Sepsis prediction system, prediction model construction method and system, and kit

Abstract

The present invention aims to develop and validate a prediction model for sepsis within 7 days after LT in liver transplant recipients. Data of 786 patients who underwent LT from January 2015 to January 2020 were retrospectively extracted from the big data platform of the Third Affiliated Hospital of Sun Yat-sen University. Seven ML models were developed to predict post-LT sepsis. Our results showed that a larger amount of red blood cell transfusion, ascites, blood loss and gastric drainage, less crystalloid infusion and urine output, longer anesthesia time, and higher preoperative TBIL level were the top 8 important indicators affecting postoperative sepsis. Among the 7 ML models developed in the study, the random forest classifier (RF) model showed the best overall performance in predicting post-LT sepsis.

Description

Technical Field

[0001] The present invention belongs to the field of medical data mining, and in particular relates to an early prediction system for postoperative sepsis in liver transplant patients, a prediction model construction method and a kit. Background Art

[0002] In recent years, machine learning technology has been widely used in the field of intelligent medicine, and has important practical significance in clinical decision-making, clinical diagnosis, and precision medicine. Machine learning-based models have high accuracy in predicting medical outcomes and identifying high-risk patients by utilizing the massive parameters already present in electronic medical records. However, there is currently no early prediction model for post-liver transplant sepsis based on machine learning. Liver transplantation (LT) is currently recognized as the only effective treatment for end-stage liver disease. Although the survival rate and long-term prognosis after LT have been significantly improved in recent years due to advances in surgical techniques, anesthesia management, immunosuppression technology, and intensive care unit (ICU) management, LT recipients still have various postoperative complications, among which postoperative sepsis is one of the most serious complications, often leading to septic shock, multiple organ dysfunction syndrome (MODS), and increased postoperative mortality. It is reported that the incidence of sepsis after LT is as high as 50-80%, and sepsis-related deaths account for 50% to 90% of all postoperative mortality. In addition, it is reported that for every hour of delay in sepsis treatment, the mortality rate increases by 7.6%. Existing technologies already have prediction methods for sepsis in ordinary patients. However, patients with poor immune function after liver transplantation have a higher incidence and mortality of sepsis after liver transplantation. The existing sepsis prediction models have poor accuracy and specificity for sepsis after liver transplantation, and cannot predict sepsis after liver transplantation at an early stage. Therefore, a reliable early prediction model for postoperative sepsis in liver transplantation patients is urgently needed to guide preventive intervention and treatment. Summary of the invention

[0003] In order to overcome the deficiencies in the prior art, the present invention provides a scientifically reliable, highly specific, good model performance, and highly sensitive model, method, and system for early prediction of postoperative sepsis in liver transplant patients. This study established a random forest prediction model based on machine learning, which incorporates preoperative and intraoperative variables to predict postoperative sepsis in liver transplant recipients. It is expected that in future clinical applications, postoperative sepsis in liver transplant recipients will be predicted, which will help early decision-making on sepsis in clinical work.

[0004] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0005] One aspect of the present invention is to provide a method for constructing an early prediction model for sepsis after liver transplantation, comprising the following steps:

[0006] S1. Data acquisition: Clinical sample data of patients undergoing allogeneic liver transplantation were collected, and the collected samples were used as an internal data set and an external validation set, respectively. The internal data set was used to train and test the model, and the external validation set was used to test the predictive performance of the model in subsequent cases in our hospital. Postoperative sepsis in patients was diagnosed according to the Sepsis 3.0 standard. The allocation ratio of the training set and the test set in the internal data set ranged from 6:4 to 7:3.

[0007] S2. Data preprocessing: exclude cases with incomplete data, map categorical variables, map "yes", "male" to 1, and map "no", "no", "female" to 0; standardize continuous variables using the standard deviation standardization method, the specific calculation formula is x'=x-mean / std, use the mean to fill missing values ​​for continuous variables, use the mode to fill missing values ​​for categorical variables, and ensure that the incidence of sepsis in the training set and the test set are consistent;

[0008] S3. Variable screening: Through the perioperative specialist database system composed of electronic medical records, variables including demographic variables, preoperative variables, i.e., comorbidities, etiologies, complications, and laboratory test results, intraoperative variables including intraoperative events, medication, fluid management, etc., the variables at least include intraoperative red blood cell transfusion volume, ascites volume, blood loss, gastric drainage volume, crystalloid fluid infusion volume and urine volume, anesthesia time, and preoperative total bilirubin level, t-test was used for continuous variables, and chi-square test was used for categorical variables to obtain the statistical test results of each variable, and the variables with p<0.05 in the statistical test results were selected, which were the variables obtained in the first variable screening. Based on the variables obtained in the first variable screening, the least absolute shrinkage and selection operator regression methods were used for modeling, and the variables with coefficients not equal to 0 in the model were selected, which were the results of the second variable screening;

[0009] S4. Development of machine learning model: The important variables obtained from the second variable screening are used for model training. The model training uses a grid search method combined with a K-fold cross-validation method to select the optimal parameter combination. First, the parameters and parameter value lists that need to be grid searched for each model are set, and cross-combinations are performed. The training set is then randomly divided into 5 equal parts, 1 of which is used as a test set and the remaining 4 parts are used as training sets. The models under specific parameter combinations are trained and tested, and the average value of the evaluation index of the corresponding model is calculated as the model score of the parameter combination. By comparing the model scores of each parameter combination, the optimal parameter combination for each model can be finally obtained; the model is selected from one or more of 7 models, including logistic regression algorithm, support vector machine algorithm, Gaussian Bayes algorithm, random forest algorithm, gradient boosting algorithm, adaptive boosting algorithm, and deep learning algorithm.

[0010] Preferably, the training model in step S4 is selected from the random forest algorithm RF (Random Forest).

[0011] S5 Model Validation: The model is validated by one or more of the following three methods: ① The test set is used for testing, and the bootstrap resampling method is used to resample the test set 1000 times with replacement to obtain 1000 test data sets. The above test data sets are used to test and evaluate the model. The evaluation of the model uses evaluation indicators such as the area under the ROC curve (AUC), accuracy, sensitivity, specificity, and F1 score. The 1000 test evaluation indicators are expressed in the form of "median (2.5% quantile, 97.5% quantile)". By testing the 7 models with the test set, the performance of various models can be obtained, and the optimal model can be selected. ② The external validation set on the time series in the hospital database, that is, the samples after the cutoff time of the internal data set research cohort, is used to test the optimal model, which further illustrates that the model is suitable for the prediction of future patients in this hospital. ③ The results of the traditional clinical sepsis prediction method SOFA score are compared with the evaluation results of the optimal model, which further illustrates that the model prediction results are better than traditional methods. Another aspect of the present invention is to provide a system for constructing a sepsis prediction model after liver transplantation, the system comprising one or more computer processors and a computer-readable medium, wherein the computer-readable medium stores a plurality of instructions, and the plurality of instructions instruct the one or more computer processors to execute the above-mentioned sepsis prediction model construction method.

[0012] Another aspect of the present invention is to provide a kit for predicting sepsis after liver transplantation, the kit comprising detection reagents and detection instruments for detecting the amount of red blood cell transfusion, ascites volume, blood loss, gastric drainage volume, crystalloid fluid input, urine volume, anesthesia time and total bilirubin level of patients before liver transplantation surgery.

[0013] Another aspect of the present invention is to provide a detection reagent and instrument for detecting the amount of red blood cell transfusion, ascites volume, blood loss, gastric drainage volume, crystalloid fluid input and urine volume, anesthesia time and total bilirubin level of a patient before liver transplantation surgery in the preparation of a product or kit for establishing an early prediction model for post-liver transplantation sepsis or a method for predicting post-liver transplantation sepsis. The detection reagent and instrument are routinely used in clinical practice.

[0014] Another aspect of the present invention provides an early prediction system for sepsis after liver transplantation, the system comprising an input device, a processor and a computer-readable medium; the input device is used to obtain the measured values ​​of relevant detection indicators of the tested liver transplant patient; the processor is connected to the input device, and the processor is used to process the data obtained by the input device and output the predicted value of sepsis risk; the computer-readable medium stores a plurality of instructions, and the plurality of instructions instruct the input device and the processor to execute a method for predicting sepsis after liver transplantation; the method comprises the following steps: a) obtaining the measured values ​​of eight indicators of red blood cell transfusion volume, ascites volume, blood loss and gastric drainage volume, crystalloid and urine volume, anesthesia time and preoperative TBIL level during liver transplantation surgery of the liver transplant patient; b) standardizing the measured values ​​of the eight indicators in step a) and loading them into the computer-readable medium; The trained random forest model inputs the standardized result parameters of the 8 indicators into the trained random forest model, the random forest model includes 100 different decision trees, each decision tree predicts the input sample respectively, obtains 100 prediction results, and finally adopts the voting method to obtain the model prediction result; wherein the steps of the random forest algorithm are as follows: the result parameters of the 8 indicators that have been standardized are input into the trained 100 decision tree basic model, and the corresponding demand attributes of each decision tree are substituted respectively to obtain the prediction category and probability results of each decision tree model respectively, and the prediction rule of the random forest model is to vote on the prediction classification results of the 100 decision trees to obtain the final prediction classification result, and the prediction probability results of the 100 decision trees are averaged to obtain the risk prediction value of sepsis.

[0015] The present invention finally obtained 8 important indicators that can be used for early stage prediction of the risk assessment of postoperative sepsis in liver transplant patients, namely, intraoperative red blood cell transfusion volume, ascites volume, blood loss and gastric drainage volume, crystalloid fluid infusion volume and urine volume, anesthesia time, and preoperative total bilirubin level, enabling anesthesiologists to perform early intervention during surgery. Among the 7 machine learning models developed in the study, the random forest classifier (RF) model had the best overall performance in predicting postoperative sepsis in liver transplantation, with an AUC of 0.731, an accuracy of 71.6%, a sensitivity of 62.1%, a specificity of 76.1% in the internal validation set, and an AUC of 0.755 in the external validation set. The present invention found that the RF model has higher performance than the postoperative SOFA score. The present invention uses machine learning (ML) technology to develop and verify a prediction model for sepsis within 7 days after LT in liver transplant recipients. The present invention fills the gap in the prior art in the lack of effective prediction methods and prediction models for early postoperative sepsis in liver transplant patients. In addition, the present invention uses multiple model evaluation methods to evaluate the models developed by the present invention, and the random forest classifier (RF) model finally screened out has more stable performance, more accurate prediction results, and can reliably predict the prediction model, method and system for postoperative sepsis in liver transplant recipients. It has better performance than the existing prediction sepsis scoring system. The prediction model, method and system incorporate preoperative and intraoperative variables to predict postoperative sepsis in liver transplant recipients, which is expected to predict postoperative sepsis in liver transplant recipients in future clinical applications, and help early decision-making on sepsis in clinical work. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 designing a roadmap for research;

[0017] Figure 2 This is the importance ranking chart of the 8 important variables included;

[0018] Figure 3 It is the AUC curve of 7 prediction models in the internal validation set;

[0019] Figure 4 This is the AUC curve of the RF model and SOFA model in the internal validation set;

[0020] Figure 5 The AUC curve of the RF model in the internal validation set and the external validation set;

[0021] Figure 6 Diagram of the online calculation interface for predicting the risk of sepsis after liver transplantation. DETAILED DESCRIPTION

[0022] In order to more clearly illustrate the embodiments of the present invention, specific implementations of the present invention will be described below with reference to the accompanying drawings.

[0023] Embodiment 1: The research idea of ​​the present invention is as follows Figure 1 As shown, LR represents logistic regression; SVM represents support vector machine; RF represents random forest; GBM represents gradient boosting; ADA represents adaptive enhancement; GNB represents Gaussian naive Bayes; MLP represents multi-layer perceptron. The method for constructing an early prediction model for post-liver transplantation sepsis in this application includes the following steps:

[0024] (1) Data acquisition: The dataset samples were derived from liver transplantation records on the big data platform of the Third Affiliated Hospital of Sun Yat-sen University (Guangzhou, China), including patients who received allogeneic liver transplantation during the 6-year period from January 2015 to April 2021. All liver transplant recipients were registered in the Chinese Organ Transplant Response System (www.cot.org.cn). The inclusion criteria used were: (1) age ≥ 18 years; (2) allogeneic liver transplantation. Patients with the following conditions were excluded from this study: (1) combined liver and kidney transplantation; (2) other operations performed at the same time; (3) incomplete medical records; (4) combined with preoperative sepsis. The samples collected from January 2015 to January 2020 were used as the internal dataset for training and testing the model, and the samples collected from February 2020 to April 2021 were used as the external validation set to test the prediction performance of the model in subsequent cases in our hospital. The internal dataset contained 677 samples and 60 variables. The external validation set contained 118 samples. Postoperative sepsis was diagnosed according to the Sepsis 3.0 criteria. Specifically, the daily SOFA score of each patient was collected through the electronic medical record system and manually reviewed, and the SOFA score on the day of surgery was set as the SOFA baseline score. Organ dysfunction was defined as an acute change in the SOFA score of ≥2 points caused by infection. Machine learning model for predicting sepsis within 7 days after liver transplantation.

[0025] (2) Data preprocessing: Map categorical variables, such as mapping "yes", "male", etc. to 1, and mapping "no", "no", "female", etc. to 0. Use the mean to fill missing values ​​for continuous variables, and use the mode to fill categorical variables. Use the standard deviation standardization method to standardize the data for continuous variables, and the specific calculation formula is x'=(x-mean) / std. The internal data set is randomly divided into a training set (70%) and a test set (30%) in a ratio of 7:3, and the proportion of samples with and without sepsis in the training set and the test set is kept consistent. Finally, 473 samples were obtained in the training set and 204 samples were obtained in the test set. Preferably, in step 2), the allocation ratio of the training set and the test set ranges from 6:4 to 7:3.

[0026] (3) Variable screening: Based on the summary of previous literature and the actual situation of the Third Affiliated Hospital of Sun Yat-sen University, a total of 60 variables were collected according to the perioperative specialist database system of the electronic medical records of the Third Affiliated Hospital of Sun Yat-sen University (see Table 1), including intraoperative red blood cell transfusion volume, ascites volume, blood loss, gastric drainage volume, crystalloid fluid input and urine volume, anesthesia time and preoperative total bilirubin level. Since multicollinear variables and confounding variables will affect the model fitting performance, the least absolute shrinkage and selection operator (LASSO) regression method is used to select variables with non-zero coefficients to prevent overfitting of the model and improve its clinical applicability. At the same time, the bootstrap method is combined with the LASSO method to sample 1000 different test sets, and the instability and sensitivity of the LASSO regression to sampling variability are processed. The t test is used for continuous variables, and the chi-square test is used for categorical variables to obtain the statistical test results of each variable. The variables with p<0.05 in the statistical test results are selected, which are the variables obtained by the first variable screening. Based on the variables obtained from the first variable screening, the least absolute shrinkage and selection operator (LASSO) regression method is used to build a model, and the variables with coefficients not equal to 0 in the model are selected, which are the results of the second variable screening. Figure 2 As shown in Figure 2, the eight indicators of red blood cell transfusion volume, ascites volume, blood loss, gastric drainage volume, crystalloid fluid infusion volume and urine volume, anesthesia time and preoperative total bilirubin level during liver transplantation surgery are the eight most important variables affecting the prediction of sepsis after LT (see Figure 2 ). Finally, 8 important indicators for constructing an early prediction model for sepsis after liver transplantation were screened, namely, intraoperative red blood cell transfusion volume, ascites volume, blood loss, gastric drainage volume, crystalloid fluid infusion volume and urine volume, anesthesia time and preoperative total bilirubin level. The final selection of these 8 important indicators is the innovation of the present invention, which ensures the accuracy and reliability of the prediction results.

[0027] After deleting variables with too many missing items, extreme deviations, ordinal variables, postoperative variables, and designated variables, 60 variables remained. The remaining variables were subjected to t-test and chi-square test, among which 34 variables had a P value less than 0.05. The test results are shown in Table 1:

[0028] Table 1

[0029]

[0030]

[0031]

[0032] (4) Development of machine learning models: The eight important variables obtained from the second variable screening are used for model training, where important variables refer to variables with high rankings. The training model is selected from one or more of seven classic machine learning model algorithms, including logistic regression algorithm LR (LogisticRegression), support vector machine algorithm SVM (Support Vector Machine), Gaussian Naive Bayesian algorithm GNB (Gaussian Naive Bayesian), random forest algorithm RF (Random Forest), gradient boosting algorithm GBM (Gradient Boosting Machine), adaptive boosting algorithm ADA (AdaBoost), and deep learning algorithm MLP (Multi-Layer Perceptron). Model training uses the grid search method combined with the K-fold cross-validation method to select the optimal parameter combination. First, the parameters and parameter value lists that need to be grid searched for each model are set and cross-combined. Then the training set is randomly divided into 5 equal parts, 1 of which is used as the test set and the remaining 4 are used as the training set. The models under specific parameter combinations are trained and tested, and the average value of the evaluation index of the corresponding model is calculated as the model score of the parameter combination. By comparing the model scores of each parameter combination, the optimal parameter combination for each model can be obtained.

[0033] The present invention compares seven modeling prediction methods: logistic regression (LR), support vector machine (SVM), random forest classifier (RF), gradient boosting (GBM), adaptive boosting classifier (ADA), Gaussian naive Bayes (GNB) and multilayer perceptron (MLP). All of the above models are established by Scikit-learn software package (https: / / github.com / scikit-learn / scikit-learn). Preferably, the training model in step 4) is selected from the random forest algorithm RF (Random Forest), which has advantages.

[0034] The basic principle of the random forest algorithm is to use the Bootstrap resampling method to obtain sample sets with different sample sizes of s (s does not exceed the total sample size) for building the basic model, thereby increasing the difference between models and improving the ability of extrapolation prediction. For attribute selection: First, randomly select a subset containing k (k is usually log2d, where d is the size of the attribute set) attributes from the attribute set of the base decision tree node; second, select the best attribute from this subset for division; finally, use the majority result method, average method, voting method, etc. to comprehensively consider the prediction results of multiple decision trees to determine the final prediction result of the random forest algorithm.

[0035] In this application, when training the classification model of random forest, first, S samples are randomly selected from the original training set, and a sample set consisting of k attributes is randomly selected to train and establish a decision tree basic model; secondly, the Bootstrap method is used to perform N samplings to establish N decision tree basic models, and finally, for each sample, the trained N decision tree basic models are substituted to obtain N predicted classification results, and each sample is voted according to the classification results; finally, the classification is determined, where I(·) is an indicative function. The formula is as follows:

[0036]

[0037] When h i When (x) = Y, I(h i (x) = Y) = 1, otherwise I (h i (x)=Y)=0. Given a set of classification models m1(x), m2(x), …, mN(x), the training data of each classification is sampled from the original data (X, Y), where N is greater than or equal to 1 and is an integer, X represents the numerical vector of the independent variable, and Y represents the numerical vector of the dependent variable. Therefore, the residual function f(X, Y) is used to find the specific situation where the correct classification is greater than the wrong classification. The formula is as follows:

[0038]

[0039] It can be seen that f(X,Y) is closely related to the classification prediction results. The larger the f(X,Y), the more accurate the prediction results. Therefore, the extrapolation error of the model is:

[0040] PE * =P X，Y [f(X,Y)<0]

[0041] As the number of decision tree classifications increases, the generalization error increases, and all decision trees converge to the formula:

[0042]

[0043] In the above formula, n is the number of decision trees in the forest, and θ is the decision tree segmentation threshold parameter. As the decision tree becomes larger, the generalization error PE tends to the upper bound, that is, the random forest algorithm has good convergence and the ability to prevent overfitting.

[0044] (5) Performance evaluation of the model and its use in predicting sepsis after liver transplantation: The model was validated by one or more of the following three methods: ① The model was tested using a test set and the bootstrap resampling method was used to resample the test set 1000 times with replacement to obtain 1000 test data sets. The model was tested and evaluated using the above test data sets. The model evaluation used evaluation indicators such as the area under the ROC curve (AUC), accuracy, sensitivity, specificity, and F1 score. The 1000 test evaluation indicators were expressed in the form of "median (2.5% quantile, 97.5 quantile)". The performance of various models was obtained by testing the 7 models with the test set, and the optimal model was selected. ② The optimal model was tested using an external validation set on the time series in the hospital database, that is, samples after the cutoff time of the internal data set research cohort, to further illustrate that the model is suitable for predicting future patients in our hospital. ③The results of the traditional clinical sepsis prediction method SOFA score were compared with the evaluation results of the optimal model, further demonstrating that the model prediction results are superior to traditional methods.

[0045] Among the seven predictive modeling methods studied, the results were Figure 3 As shown, the RF model has the largest AUC (0.731, CI 0.649 to 0.802), the highest F1 score (0.581, CI 0.476 to 0.676), and relatively balanced sensitivity (0.621, CI 0.493 to 0.725) and specificity (0.761, CI 0.69 to 0.832). Therefore, the present invention finally selected the RF model for further analysis and application.

[0046] The characteristics of the external validation set included a high proportion of males (90.68%) and a mean age of 47.3 years (Table S3). On the one hand, the anesthesia time and urine output of the external validation set were significantly lower than those of the development set (all P < 0.05). On the other hand, the crystalloid fluid input and ascites volume were higher in the external validation set. In the external validation set, the incidence of sepsis was 28.81%, and the RF model achieved an AUC comparable to that of the internal validation set (0.755, CI 0.652 to 0.839) ( Figure 3 ).

[0047] Since the SOFA score has been reported to predict postoperative sepsis in ICU patients, the present invention validated and compared the performance of the SOFA score and the RF-based model of the present invention. Figure 3It turns out that in the internal validation set of the present invention, the AUC of the SOFA score (0.637, CI 0.551 to 0.692) is lower than the AUC of the RF model in the validation set (0.745, CI 0.645 to 0.824).

[0048] Since the 8 important indicators included in the model of the present invention can be obtained in clinical practice, so as to conveniently calculate the risk of post-LT sepsis, the present invention has also developed an online risk calculator, the algorithm of which is fixed. The online risk calculator allows anesthesiologists and colleagues around the world to use this model. By inputting the result parameters of the 8 indicators of red blood cell transfusion volume, ascites volume, blood loss and gastric drainage volume, crystalloid and urine volume, anesthesia time and pre-operative TBIL level during liver transplant surgery into the online risk calculator, the predicted value of the risk of sepsis after LT in liver transplant patients can be obtained. The result output format is 1 / 0 (0-1). Figure 6 As shown in the figure, "1" indicates a positive result for post-LT sepsis, "0" indicates a negative result, and the value in brackets is the probability of post-LT sepsis. For example, the predicted output for patient No. 10 is "0(0.76)", which means that the probability of post-LT sepsis in this patient is only 24%. An online risk calculator for calculating the risk of LT sepsis can be accessed at http: / / wb.aidcloud.cn / zssy / sepsis_web.html.

[0049] Embodiment 2: The present invention provides a system for constructing an early prediction model of sepsis after liver transplantation, the system comprising one or more computer processors and a computer-readable medium, wherein the computer-readable medium stores a plurality of instructions, and the plurality of instructions instruct the one or more computer processors to execute the above-mentioned method for constructing a sepsis prediction model.

[0050] Embodiment 3: The present invention provides an early prediction system for sepsis after liver transplantation, the system comprising an input device, a processor and a computer-readable medium; the input device is used to obtain the measured values ​​of relevant detection indicators of the tested liver transplant patient; the processor is connected to the input device, and the processor is used to process the data obtained by the input device and output the predicted value of sepsis risk; the computer-readable medium stores a plurality of instructions, and the plurality of instructions instruct the input device and the processor to execute a method for predicting sepsis after liver transplantation; the method comprises the following steps: a) obtaining the measured values ​​of eight indicators of red blood cell transfusion volume, ascites volume, blood loss and gastric drainage volume, crystalloid and urine volume, anesthesia time and preoperative TBIL level during liver transplantation surgery of the liver transplant patient; b) standardizing the measured values ​​of the eight indicators in step a) and loading them into the computer-readable medium; The trained random forest model inputs the result parameters of the standardized 8 indicators into the trained random forest model, and the random forest model contains 100 different decision trees, each decision tree predicts the input sample respectively, obtains 100 prediction results, and finally adopts the voting method to obtain the model prediction result; wherein the steps of the random forest algorithm are as follows: the result parameters of the 8 indicators that have been standardized are input into the trained 100 decision tree basic model, and the corresponding demand attributes of each decision tree are substituted respectively, and the prediction category and probability results of each decision tree model are obtained respectively, and the prediction rule of the random forest model is to vote on the prediction classification results of the 100 decision trees to obtain the final prediction classification result, and the prediction probability results of the 100 decision trees are averaged to obtain the risk prediction value of sepsis. The prediction results of patient No. 10 obtained by using this system and the online risk calculator developed based on the method for predicting sepsis after liver transplantation are consistent.

[0051] Embodiment 4: The present invention provides a kit for predicting sepsis after liver transplantation, the kit comprising detection reagents and detection instruments for detecting the amount of red blood cell transfusion, ascites volume, blood loss, gastric drainage volume, crystalloid fluid input, urine volume, anesthesia time and total bilirubin level of liver transplant patients before liver transplantation surgery.

[0052] The statistical analysis of the present invention was implemented in the anaconda basic environment (https: / / www.anaconda.com) of python 3.7, including scikit-learn 0.22, numpy 1.17.0 and pandas 1.2.3. Independent sample t-test was used to compare normally distributed data. For non-normally distributed data, Mann–Whitney U test was used in univariate analysis. Categorical variables were analyzed by chi-square test or Fisher test, and long-term survival was analyzed by Kaplan-Meier method. Only categorical variables had missing values, and the missing proportion was less than 10%. Categorical variables were filled with mode. Continuous variables in both training and validation sets were standardized according to the mean and standard deviation of the training set, while categorical variables were treated as dummy variables. All models were developed in the same 70% training set and validated in the 30% validation set.

[0053] The preoperative, intraoperative and postoperative time of the present invention is based on the time of the first liver transplantation operation, which are respectively the last time before the operation, intraoperatively and within 1-3 days after the operation.

[0054] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, according to the ideas provided by the present invention, there will be changes in the specific implementation methods, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. A prediction system for post-liver transplant sepsis, the system comprising an input device and a processor; and a computer-readable medium, the computer-readable medium storing a plurality of instructions, the input device being used to obtain the measured values ​​of the relevant detection indexes of the tested liver transplant patient; the processor being connected to the input device, the processor being used to process the data obtained by the input device and output the predicted value of the risk of sepsis; the instructions instructing the input device and the processor to execute a method for predicting post-liver transplant sepsis; the method comprising the following steps: a) obtaining the measured values ​​of eight indexes, namely, the amount of red blood cell transfusion, ascites volume, blood loss and gastric drainage volume, crystalloid volume and urine volume, anesthesia time and TBIL before surgery in the liver transplant patient; b) standardizing the measured values ​​of the eight indexes in step a) and loading the trained The trained random forest model inputs the standardized result parameters of the 8 indicators into the trained random forest model. The random forest model contains 100 different decision trees, each decision tree predicts the input sample respectively, obtains 100 prediction results, and finally adopts the voting method to obtain the model prediction result; wherein the steps of the random forest algorithm are as follows: the result parameters of the 8 indicators that have been standardized are input into the trained 100 decision tree basic model, and the corresponding demand attributes of each decision tree are substituted respectively to obtain the prediction category and probability results of each decision tree model respectively. The prediction rule of the random forest model is to vote on the prediction classification results of the 100 decision trees to obtain the final prediction classification result, and the prediction probability results of the 100 decision trees are averaged to obtain the risk prediction value of sepsis.

2. A method for constructing a prediction model for sepsis after liver transplantation, characterized in that: The prediction model is an early prediction model, and the method comprises the following steps: S1. Data acquisition: Clinical sample data of patients undergoing allogeneic liver transplantation were collected, and the collected samples were used as an internal data set and an external validation set, respectively. The internal data set was used to train and test the model, and the external validation set was used to test the predictive performance of the model in subsequent cases in our hospital. Postoperative sepsis in patients was diagnosed according to the Sepsis 3.0 standard. The allocation ratio of the training set and the test set in the internal data set ranged from 6:4 to 7:

3. S2. Data preprocessing: Cases with incomplete data were excluded, and categorical variables were mapped, with "yes", "male" mapped to 1, and "no", "no", and "female" mapped to 0; continuous variables were standardized using the standard deviation standardization method, and the specific calculation formula was x , = x-mean / std, use the mean to fill missing values ​​for continuous variables, use the mode to fill missing values ​​for categorical variables, and ensure that the incidence of sepsis in the training set and the test set is consistent; S3. Variable screening: Through the perioperative specialist database system composed of electronic medical records, variables including demographic variables, preoperative variables, i.e., comorbidities, etiologies, complications, and laboratory test results, intraoperative variables including intraoperative events, medication, and fluid management, the variables at least including intraoperative red blood cell transfusion volume, ascites volume, blood loss, gastric drainage volume, crystalloid fluid infusion volume and urine volume, anesthesia time, and preoperative total bilirubin level were collected. The t-test was used for continuous variables, and the chi-square test was used for categorical variables to obtain the statistical test results of each variable. The variables with p<0.05 in the statistical test results were selected as the variables obtained in the first variable screening. Based on the variables obtained in the first variable screening, the least absolute shrinkage and selection operator regression methods were used for modeling, and the variables with coefficients not equal to 0 in the model were selected as the results of the second variable screening. S4. Development of machine learning model: The important variables obtained from the second variable screening are used for model training. The model training uses a grid search method combined with a K-fold cross-validation method to select the optimal parameter combination. First, the parameters and parameter value lists that need to be grid searched for each model are set, and cross-combinations are performed. The training set is then randomly divided into 5 equal parts, 1 of which is used as a test set and the remaining 4 parts are used as training sets. The models under specific parameter combinations are trained and tested, and the average value of the evaluation index of the corresponding model is calculated as the model score of the parameter combination. By comparing the model scores of each parameter combination, the optimal parameter combination for each model can be finally obtained; the model is selected from a logistic regression algorithm, a support vector machine algorithm, a Gaussian Bayes algorithm, a random forest algorithm, a gradient boosting algorithm, an adaptive boosting algorithm, and a deep learning algorithm.

3. The method for constructing a sepsis prediction model according to claim 2, characterized in that: It also includes a performance evaluation step for the model, and the performance evaluation of the model uses the area under the ROC curve, accuracy, sensitivity, specificity, and F1 score evaluation indicators.

4. A sepsis prediction model construction system, the system comprising one or more computer processors and a computer-readable medium, wherein the computer-readable medium stores a plurality of instructions, wherein the plurality of instructions instruct the one or more computer processors to execute the sepsis prediction model construction method according to any one of claims 2 to 3.

5. A kit for predicting sepsis after liver transplantation, characterized in that: The kit comprises detection reagents and detection instruments for detecting the amount of red blood cell transfusion, ascites volume, blood loss, gastric drainage volume, crystalloid fluid input, urine volume, anesthesia time and the total bilirubin level of the patient before the liver transplantation operation during the liver transplantation operation.

6. Use of reagents and instruments for detecting the amount of red blood cell transfusion, ascites volume, blood loss, gastric drainage volume, crystalloid fluid infusion, urine volume, anesthesia time and total bilirubin level before liver transplantation in the preparation of an early prediction model for post-liver transplantation sepsis or a kit for predicting the risk of post-liver transplantation sepsis.

Images