Method and system for predicting pathogens of bloodstream infection

Abstract

The present invention relates to a method and system for predicting pathogenic bacteria of bloodstream infection, comprising the following steps: obtaining input features and preprocessing, wherein the input features are physical examination data; selecting the optimal input features from the preprocessed input features to form an input feature set, and outputting prediction results through a prediction model; wherein the prediction model is a method for predicting pathogenic bacteria of bloodstream infection by using multiple first classifiers in parallel as base classifiers of the prediction model, using a cost-sensitive decision tree as the base classifier of each first classifier, and the misclassification weight of a single sample is the product of the sample weight of the first classifier and the weight of the class misclassified in the decision tree. The prediction model is constructed by connecting the physical examination data of bloodstream infection patients with positive and negative bacteria in pathogenic bacteria, so as to predict pathogenic bacteria of bloodstream infection in samples without increasing additional costs.

Description

Technical Field

[0001] The present invention relates to the technical field of data classification, and in particular to a method for predicting pathogenic bacteria of bloodstream infection. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Bloodstream infection is the main cause of sepsis and septic shock. Choosing appropriate antimicrobial agents can cope with the harm caused by bloodstream infection. As microbial resistance to antimicrobial agents is increasing, the current method is to perform blood culture (BC) on the collected samples under laboratory conditions to select appropriate antimicrobial agents by identifying positive pathogenic microorganisms. However, this method is time-consuming and the accuracy of positive identification is not high. Although PCR (polymerase chain reaction)-based detection takes relatively less time, it is limited by the number of PCR probes and the number of samples, and is difficult to be used on a large scale in the detection of Gram-positive bacteria in bloodstream infection. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a method and system for predicting bloodstream infection pathogens. By mining the relationship between the physical examination data indicators of bloodstream infection patients and the pathogens, it aims to achieve accurate screening of positive bacteria, negative bacteria and fungi among the bloodstream infection pathogens in sample data by machine learning without adding additional costs.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] The first aspect of the present invention provides a method for predicting pathogens of bloodstream infection, comprising the following steps:

[0007] Obtain input features and preprocess them, where the input features are physical examination data;

[0008] Filter out the best input features from the preprocessed input features to form an input feature set, and output the prediction results through the prediction model;

[0009] The prediction model uses multiple parallel first classifiers as base classifiers of the prediction model, a cost-sensitive decision tree as the base classifier of each first classifier, and the misclassification weight of a single sample is the product of the sample weight of the first classifier and the weight of the misclassified class in the decision tree.

[0010] The first classifier is the AdaBoost classifier.

[0011] The input features are obtained and preprocessed. The input features are physical examination data, including judging whether the model input indicators are complete based on the preprocessed features and whether these indicators are obtained within a set time period. If they are satisfied, the preprocessed features are input into the prediction model to obtain the prediction results.

[0012] Preprocessing includes deleting features with missing data ratios exceeding a set value and samples with missing values; outliers are considered as abnormal values ​​and deleted through box plots or scatter plots of the features.

[0013] The input features of the prediction model are all the optimal features selected.

[0014] The training process of the prediction model includes:

[0015] Obtain physical examination data of patients with bloodstream infection and perform preprocessing to remove features that do not differ between Gram-positive bacteria, Gram-negative bacteria, and fungi in bloodstream infection;

[0016] Check and delete the continuous features whose difference coefficients between the selected feature groups are greater than the set value;

[0017] Test and delete gender characteristics whose coefficient of difference between the two groups is greater than the set value;

[0018] Check and delete features with inter-feature correlation coefficient |γ|>0.8;

[0019] The remaining feature data after the test are scaled to a standard normal distribution and divided into a training set and a test set;

[0020] Input the training set and test set to complete the training of the prediction model.

[0021] A second aspect of the present invention provides a system for implementing the above method, comprising:

[0022] A preprocessing module is configured to: obtain input features and preprocess them, where the input features are physical examination data;

[0023] The prediction output module is configured to: select the optimal input features from the preprocessed input features to form an input feature set, and output the prediction results through the prediction model;

[0024] Among them, the prediction model is to use multiple parallel first classifiers as the base classifiers of the prediction model, and a cost-sensitive decision tree as the base classifier of each first classifier. The misclassification weight of a single sample is the product of the sample weight of the first classifier and the weight of the misclassified class in the decision tree, and the first classifier is an AdaBoost classifier.

[0025] A third aspect of the present invention provides a computer-readable storage medium.

[0026] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the method for predicting pathogens of bloodstream infection as described above.

[0027] A fourth aspect of the present invention provides a computer device.

[0028] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for predicting pathogens of bloodstream infection as described above are implemented.

[0029] Compared with the prior art, one or more of the above technical solutions have the following beneficial effects:

[0030] 1. A prediction model is constructed through the connection between the physical examination data of patients with bloodstream infection and the negative bacteria, positive bacteria and fungi among the pathogenic bacteria. Without adding additional costs, the positive bacteria, negative bacteria and fungi of the pathogenic bacteria of bloodstream infection in the sample data can be predicted, thereby realizing the screening of pathogenic bacteria in the sample.

[0031] 2. Using the integrated model as the prediction model, multiple parallel AdaBoost classifiers are used as the base classifiers of the prediction model to form the integration of the first layer of models. After sampling, the data in the sample is first input into multiple parallel AdaBoost classifiers, and each AdaBoost classifier itself is an integrated model, and its base classifier is a cost-sensitive decision tree, which forms the second layer of integration. After all the preprocessed features are input into the integrated model, the results are more complete and more accurate, and the positive bacteria, negative bacteria and fungi in the sample data can be predicted, thereby realizing the screening of pathogenic bacteria in the sample in a machine learning way.

[0032] 3. When the prediction model processes each data set after feature elimination, it is processed by each AdaBoost classifier respectively. The classifier will adjust the sample weight according to the classification result until a better result is obtained. The base classifier of each AdaBoost classifier is a cost-sensitive decision tree, which will generate various weights according to the number of input samples of each category. The cost of misclassification will be calculated based on the sample weight in the base classifier, and then the sample weight will be adjusted according to the classification result to obtain a more accurate result.

[0033] 4. In order to improve the performance of the model, grid search is used to optimize the hyperparameters of the model and select the input features, thereby reducing the feature dimension, improving the generalization ability of the model, and achieving better prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] Figure 1 It is a flow chart of a method for predicting positive bacteria, negative bacteria and fungi that are pathogenic bacteria in bloodstream infection provided by one or more embodiments of the present invention;

[0036] Figure 2 is a feature correlation heat map in a prediction method provided by one or more embodiments of the present invention;

[0037] Figure 3 is a flowchart of model training in a prediction method provided by one or more embodiments of the present invention;

[0038] Figure 4 It is a flowchart of implementing prediction using a trained model provided by one or more embodiments of the present invention. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0041] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0042] As described in the background technology, currently, blood culture (BC) is performed on the collected samples under laboratory conditions to select appropriate antimicrobial agents by identifying positive pathogenic microorganisms. The samples come from patients with bloodstream infections, and the parameters such as blood cell counts, acute phase proteins and electrolytes are data that are easily obtained in an experimental environment. These data contain characteristics related to sepsis and septic shock. By using machine learning, the positive bacteria, negative bacteria and fungi that cause sepsis and septic shock in the samples can be predicted and classified, thereby accelerating the detection of pathogenic microorganisms.

[0043] Prognosis refers to the prediction of the possible course and outcome of a disease. Samples from patients with bloodstream infections can also be used to predict the possible course and outcome of the disease using machine learning.

[0044] Therefore, the following embodiments provide a method and system for predicting Gram-positive bacteria, Gram-negative bacteria and fungi that cause bloodstream infection in a sample. By mining the relationship between the physical examination data indicators of bloodstream infection patients and the pathogenic bacteria, a prediction model is selected, aiming to achieve the prediction of positive bacteria, negative bacteria and fungi in the sample by machine learning without adding additional costs.

[0045] Embodiment 1:

[0046] like Figure 1-4 As shown, the prediction method for three types of bloodstream infection includes the following steps:

[0047] Step 1: receiving input features, where the input features are the patient's physical examination data;

[0048] Step 2: Perform data cleaning, significance and correlation detection, and data standardization preprocessing on the input features;

[0049] Step 3: Filter out the optimal input features of each layer from the preprocessed input features and form an input feature set, and output the prediction results through the prediction model;

[0050] Specific:

[0051] The input features of the prediction model are all the optimal features selected, and the optimization process of the prediction model in this embodiment also performs hyperparameter optimization.

[0052] Before optimizing the classifier, sample data were collected, including blood culture characteristics, liver function characteristics, etc. The research subjects were patients with bloodstream infection. The extracted feature data were processed to construct a data queue. The performance of the classifier was evaluated using 10-fold cross-validation.

[0053] The prediction model is as follows:

[0054] 1. Obtain indicator information of the research object

[0055] The research subjects were all patients with bloodstream infection; a total of 2118 samples were used in this example, all of which were derived from the MIMIC database.

[0056] The index information includes BaseExcess, FreeCalcium, Lactate, pCO2, pH, pO2, %HemoglobinA1c, 25-OHVitaminD, AlanineAminotransferase, Albumin, AlkalinePhosphatase, Alpha-Fetoprotein, Ammonia, AsparateAminotransferase, Bicarbonate, BilirubinDirect, BilirubinIndirect, BilirubinTotal, C-ReactiveProtei C-reactive protein, C3 (β1C-globulin), C4 (β1E-globulin), Calciu m Total (Total Calcium), Chloride (Chloride), Cholesterol Total (Total Cholesterol), Creatine Kinase (CK Creatine Kinase), Creatine Kinase MB Isoenzyme (Creatine Kinase MB Isoenzyme), Creatinine (Creatinine), D-Dimer (D-Dimer), Ferritin (Ferritin), Glucose (Glucose), HIV Antibody (HIV Antibody), Immunoglobulin A (Immunoglobulin A), Immunoglobulin G (Immunoglobulin G), Immunoglobulin M (Immunoglobulin M), Iron, Lactate Dehydrogen ase(LD Lactate Dehydrogenase), Lithium(Lithium), Magnesium(Magnesium), NTproBNP(N-terminal pro-brain natriuretic peptide), Phosphate(Phosphate), Potassium(Potassium), Sodium(Sodium), Thyroxine(T4)Free(Demethoxine), Triglycerides(Triglycerides), Triiodothyronine(T3 Triiodothyronine), TroponinI(Troponin I), TroponinT(Troponin T), UreaNitrogen(Urea Nitrogen), VitaminB12(Vitamin B12), AbsoluteCD3Count(Absolute CD3 Count), Absolut eCD4Count(Absolute CD4 Count),AbsoluteCD8Count (absolute CD8 count), AbsoluteLymphocyteCount (absolute lymphocyte count), Antithrombin (antithrombin), Basophils (basophils), CD10 (neprilysin), CD103, CD117, CD11c (integrin αX), CD13 (aminopeptidase N), CD138, CD14 (receptor for LPS and LBP complex), CD15 (fucosyltransferase), CD16 / 56, CD16 (low affinity Fc receptor), CD16 / 56 Absolute Count (absolute count of CD16 / 56), CD16 / 56% (percentage of CD16 / 56), CD19 (B lymphocyte surface antigen B4), CD19% (percentage of B lymphocyte surface antigen B4), CD19 Absolute Count (absolute count of B lymphocyte surface antigen B4), CD2 (sheep erythrocyte receptor), CD20 (B lymphocyte surface antigen B1), CD20% (CD20 %, CD20AbsoluteCount(CD20 absolute count), CD22(T cell surface antigen Leu-14), CD23(low affinity immunoglobulin epsilon Fc receptor), CD25(TAC antigen), CD3%(CD3 percentage), CD3AbsoluteCount(CD3 absolute count), CD3CellsPercent(CD3 cell percentage), CD33(sialic acid binding immunoglobulin-like lectin 3), CD34, CD38(ADP-ribose Cyclase 1), CD4 Cells Percent (CD4 cell percentage), CD4 / CD8 Ratio (CD4 and CD8 bilv), CD41 (integrin α-IIb), CD45 (leukocyte common antigen), CD5, CD5% (CD5 percentage), CD5 Absolute Count (CD5 absolute count), CD55 (complement decay accelerating factor), CD56, CD57 (human natural killer cell 1), CD59 (MAC-inhibitory protein), CD64 (IgG Fc receptor I), CD7, CD71 (Transferrin receptor protein 1), Eosinophils (eosinophils), HLA-DR (human leukocyte DR antigen), INR (PT), Lymphocytes (lymphocytes), MCH (mean corpuscular hemoglobin), MCHC (mean corpuscular hemoglobin concentration), MCV (mean corpuscular volume), MonocyteCount (monocyte count), Neutrophils (neutrophils), PlateletCount (platelet count), PT (prothrombin time), PTT (partial thromboplastin time),Red Blood Cells (red blood cells), Reticulocyte Count Absolute (absolute reticulocyte count), WBC Count (white blood cell count), White Blood Cells (white blood cells). CD is a white blood cell classification antigen.

[0057] 2. Process the acquired data and determine the input parameters of the model.

[0058] Feature processing mainly includes data cleaning and statistical analysis. Test the continuous features with differences between the selected feature groups, test the differences in gender features between the two groups, and test the correlation between the features. Scale the remaining feature data after the test to a standard normal distribution, and use a random forest classifier to obtain the classification results.

[0059] The correlation between features was checked and determined, including the removal of all features with correlation coefficient |γ|>0.8.

[0060] Specific:

[0061] The acquired data are all continuous features except gender, with 0 and 1 representing female and male respectively.

[0062] Data cleaning mainly includes missing value processing and outlier processing; features with missing ratios exceeding 40% and samples with missing values ​​are deleted; the distribution of feature values ​​is viewed by drawing a box plot or scatter plot of the feature, and outliers are regarded as outliers and deleted. In this embodiment, there are 80 feature samples with missing ratios exceeding 40%, so the remaining 27 features are selected.

[0063] Statistical methods were used to eliminate features with no significant differences among the three categories and features with high correlation. The feature correlations were as follows: Figure 2 As shown, the spatial dimension of the input features is reduced, reducing the risk of model overfitting.

[0064] In this embodiment, the statistical test information of 27 features is shown in Table 1:

[0065] Table 1: Statistical test information of 27 features

[0066]

[0067]

[0068] Through statistical difference analysis, the relationship between physical examination data and positive and negative bacteria can be obtained, that is, the p value in Table 1. The Kruskal-Wallis H test method was used to test the continuous characteristics with significant differences between the selected groups. The chi-square test method was used to test whether there was a significant difference in gender characteristics between the two groups, with a significance level of p = 0.05. The Pearson correlation test was used to determine the correlation between the characteristics, and all the characteristics with a Pearson correlation coefficient |γ|>0.8 between the characteristics were deleted. After statistical testing, the number of characteristics was reduced to 13.

[0069] Since the model is sensitive to feature dimensions, the continuous features with significant differences are selected for standardization, and the data distribution is scaled to a standard normal distribution to eliminate the impact of dimensions between different features. The conversion formula for feature standardization is:

[0070] Among them, X new , X, μ, and σ represent the transformed eigenvalue, original eigenvalue, characteristic population mean, and characteristic population standard deviation, respectively.

[0071] OneHot encoding is used for discrete features to avoid sorting unordered features.

[0072] 3. Use the classifier to get the prediction results.

[0073] The cost-sensitive random forest, EasyEnsemble, RUSBoost, and AdaBoost (AdaptiveBoosting) classifiers were selected for comparison with the prediction model in this embodiment. The optimization of the classifier was performed by simultaneously optimizing the model hyperparameters using a grid search method, and the optimal parameters were selected according to the classification effect of the model.

[0074] Random forest is an ensemble learning method based on decision trees that can solve nonlinear classification problems. Cost sensitivity means that the cost of misclassification of different classes is different, so as to correct the model's prediction tendency for most sample classes. The parameters that need to be optimized in random forest include the number of decision trees, the maximum number of features, the maximum depth of the decision tree, and the minimum number of separable samples at the node. If the number of decision trees is too small, it is easy to underfit, and if it is too large, the model cannot be significantly improved. If the maximum number of features is too small, the performance will be poor, and if it is too large, the calculation amount will be too large. Improper settings of the depth of the decision tree and the minimum number of separable samples at the node will lead to underfitting or overfitting. In this embodiment, the number of decision trees is set to [50,400], with an accuracy of 50; the maximum number of features is in the range of 1 to the square root of the number of sample features, with an accuracy of 1; the maximum depth of the decision tree is in the range of [5,20], with an accuracy of 2; the percentage range of the minimum number of separable samples at the node to the sample size is [0.01,0.02], with an accuracy of 0.002.

[0075] Easyensemble is an ensemble learning method for unbalanced data. It first performs undersampling with replacement on classes with large data volumes to form many data sets, so that the sample sizes of various types of data in the data sets are balanced. Each data set is used to train the AdaBoost model, and finally the output results of multiple models are voted to determine the final prediction results. The parameters that need to be optimized include the number of models and the learning rate. Too many models or too small a learning rate will result in too long training time, and too few models or too large a learning rate will result in reduced accuracy. In this embodiment, the number of models is set in the range of [10, 100], with a precision of 10; the range of the learning rate is 0.1 to 1, with a precision of 0.1.

[0076] RUSBoost is an ensemble learning method for unbalanced data. It first randomly undersamples the data, randomly extracts a certain amount of majority class samples and minority class samples from the data set to form a balanced distribution of training data sets, and each data set is used to train the AdaBoost.M2 model. Finally, the output results of multiple models are voted to determine the final prediction results. The parameters that need to be optimized include the number of models and the learning rate. Too many models or too small a learning rate will result in too long training time, and too few models or too large a learning rate will result in reduced accuracy. In this embodiment, the number of models is set to [10, 100], with a precision of 10; the range of the learning rate is 0.1 to 1, and the precision is 0.1.

[0077] AdaBoost is an integrated learning model that enhances the base classifier by continuously modifying the sample weights. This embodiment uses a decision tree as the base classifier. The hyperparameters that need to be adjusted are the learning rate and the number of base models. If the learning rate is too large, the training speed will be fast, but the error will become larger. If the learning rate is too small, the error will be small but the training speed will be very slow. This embodiment sets the learning rate range to [0.2, 0.7] and the number of base models to [10, 100].

[0078] The prediction model of this embodiment is an integrated model, specifically:

[0079] The prediction model uses multiple parallel AdaBoost classifiers as the base classifiers of the prediction model. The base classifier of each AdaBoost classifier is a cost-sensitive decision tree. The misclassification weight of a single sample is the product of the sample weight of the AdaBoost classifier and the weight of the misclassified class in the decision tree.

[0080] Among them, the prediction model has multiple base classifiers in parallel, such as Figure 3As shown in the figure, multiple parallel AdaBoost classifiers form an integration of the first layer model. After sampling, the data in the sample is first input into multiple parallel AdaBoost classifiers; and each AdaBoost classifier itself is an integrated model, and its base classifier is a cost-sensitive decision tree, which forms the second layer of integration. The AdaBoost classifier is as follows: Figure 4 shown.

[0081] The details are as follows:

[0082] First, the input imbalanced category data samples are sampled with replacement. The sampling process does not distinguish between different types of samples, and several data sets are generated. Then, the proportions of each type of the sampled data sets are verified one by one. If the imbalance problem of a data set after sampling worsens (the number of samples of the majority class / the number of samples of the minority class increases), the data set is removed and resampled, and finally a specified number (the number is adjustable) of data sets is generated.

[0083] Then, each newly generated data set is randomly feature-removed. This process only removes a small number of features, and the proportion of features retained is above 70% to avoid losing too much information. The number of features retained can be adjusted according to the final accuracy. The above two sampling links increase the diversity of data, which is conducive to improving the performance of the integrated model.

[0084] Each data set after feature elimination is input into each AdaBoost classifier respectively, and the initial sample weight is 1. The classifier will adjust the sample weight according to the classification result until a better result is obtained.

[0085] The base classifier of each AdaBoost classifier is a cost-sensitive decision tree. The base classifier generates weights for each category according to the number of input samples in each category. For each sample, its sample weight in the base classifier = its sample weight in the AdaBoost classifier * the class weight corresponding to the sample. The cost of misclassification will be calculated based on the sample weight in the base classifier, and then the sample weight will be adjusted according to the classification result.

[0086] After the training of each AdaBoost classifier is completed, the classification probabilities of these classification results are averaged, that is, soft voting, to obtain the final prediction result.

[0087] The sample sampling ratio of the entire model is 1, the number of features retained is 11, and the number of data sets is 50. The model requires hyperparameter optimization. The above parameters are the results of optimization using the grid search method based on the validation set.

[0088] Each Adaboost classifier needs to be optimized for hyperparameters. The grid search method is used for hyperparameter optimization. During the optimization process, the classifier is evaluated for model performance using 10 10-fold cross validation. The optimized hyperparameters of each classifier may be different.

[0089] A cost-sensitive decision tree is used as the base classifier. To prevent overfitting, the maximum depth of the decision tree is set to 2. Since the maximum depth is set very small, no restrictions are imposed on other pruning hyperparameters. Cost sensitivity means that the misclassification costs of different classes are different, in order to correct the model's prediction tendency for the majority of sample classes. The misclassification costs of each class are calculated based on the proportion of the number of samples in each class. The weight of a single sample is equal to the sample weight of the AdaBoost classifier multiplied by the misclassification weight of the class.

[0090] Decision tree is a classic supervised classification algorithm for machine learning. It is a nonlinear classifier that represents a mapping relationship between object attributes and object values. It has the advantages of high execution efficiency and strong interpretability during prediction, and is less affected by data imbalance. It is often used as the base classifier of unbalanced ensemble models.

[0091] The grid search method is used for hyperparameter optimization. The grid search method is an exhaustive search method for specifying parameter values. The optimal learning algorithm is obtained by optimizing the parameters of the estimation function through cross-validation.

[0092] The process of predicting the positive and negative bacteria in the sample is a classification process for unbalanced data, so the model used is suitable for the case of unbalanced data.

[0093] Ten-fold cross validation is a commonly used method in machine learning to avoid model overfitting. This method divides the training set into ten parts with equal sample size, and selects nine of them for model training in turn, and uses the remaining one as a test. This is repeated ten times to ensure that all samples are tested once and only once; the results of the ten tests are used as the evaluation of the model.

[0094] In order to determine the optimal hyperparameters, an indicator needs to be defined to define the quality of the model effect. In medical diagnosis, the most common evaluation indicators are sensitivity and specificity. In order to comprehensively measure the size of the two, this embodiment uses the AUC area as an indicator. The definitions of sensitivity, specificity and score are as shown in the following formula:

[0095]

[0096]

[0097]

[0098] Among them, TN, TP, FN, and FP represent true negative, true positive, false negative, and false positive, respectively, which are obtained from the confusion matrix. The confusion matrix is ​​shown in Table 3:

[0099] Table 3: Confusion matrix for model classification

[0100]

[0101] The definition of accuracy is as follows:

[0102]

[0103] ROC analysis is also a commonly used evaluation index. It uses the false positive rate (FPR) as the horizontal axis and the true positive rate (TPR) as the vertical axis. By continuously lowering the classification threshold, a curve is drawn. In order to objectively evaluate the effect of ROC, the area under the ROC curve (AUC) is used to represent it. The definition of the true positive rate and the false positive rate is shown in the following formula:

[0104]

[0105]

[0106] F1-score is a measurement indicator for classification problems. It is the harmonic mean of precision and recall, with a maximum of 1 and a minimum of 0.

[0107]

[0108] The G-score is another system performance evaluation criterion that unifies precision and recall. The G-score is defined as the geometric mean of recall and precision.

[0109] G = √precision·recall

[0110] Among them, precision refers to the proportion of positive samples in the positive examples determined by the classifier.

[0111]

[0112] Recall: refers to the proportion of positive examples predicted to be positive to the total positive examples

[0113]

[0114] For the three-classification problem, the above indicators are also applicable. For a certain category, calculate the binary classification indicators of this category and non-this category, and then average the indicators of the three categories to get the final indicator.

[0115] This example uses the classifier to classify Gram-positive and Gram-negative bacteria through the above steps to predict the positive and negative bacteria in the sample; at the same time, in order to avoid the influence of the sample order on the classifier, the data set is randomly shuffled ten times, and the individual with the best average classification performance on the ten data sets is selected as the optimal solution of the model.

[0116] By comparing the indicators of the four algorithms with those of the integrated model, it is found that the performance of the integrated model is better than other classifiers.

[0117] 5. Use the classifier with parameter optimization and feature selection to obtain the classification result. The process of parameter optimization is the same as the process of hyperparameter optimization using the grid search method mentioned above.

[0118] This embodiment realizes the screening of positive bacteria, negative bacteria and fungal species without adding additional costs by mining effective information in physical examination data.

[0119] The prediction model adopts an integrated model, and the input features are all the features after preprocessing, which are more perfect and have higher accuracy, so that the positive, negative and fungi of bleeding stream infection pathogens can be predicted.

[0120] Embodiment 2

[0121] This embodiment provides a bloodstream infection Gram-positive and negative bacteria and fungi prediction system, which includes:

[0122] (1) An input feature receiving module, which is used to receive input features, where the input features are individual information and laboratory test indicators.

[0123] Personal information includes age and gender.

[0124] The laboratory examination indicators are the indicator information of the first embodiment.

[0125] (2) Preprocessing module, which is used to perform data cleaning, significance and correlation detection, and data standardization preprocessing on input features.

[0126] Specifically, the processing of extracted features mainly includes the processing of missing values ​​and outliers, and the use of statistical methods to select features with significant differences between Gram-positive, Gram-negative and fungal samples; the specific method is to use the Kruskal-Wallis H test method and the chi-square test to compare the differences in continuous features and categorical features between the two groups, respectively, and use the Pearson correlation test to detect the correlation between continuous features; the significance level p = 0.05, the Pearson correlation coefficient |γ| = 0.8; OneHot encoding is performed on categorical features to avoid sorting disordered features; continuous features are standardized and the feature data are converted into a standard normal distribution.

[0127] Feature processing mainly includes data cleaning and statistical analysis.

[0128] The acquired data are all continuous features except gender, with 0 and 1 representing female and male respectively.

[0129] Data cleaning mainly includes missing value processing and outlier processing; features with missing ratios exceeding 40% and samples with missing values ​​are deleted; the distribution of feature values ​​is viewed by drawing box plots or scatter plots of the features, and outliers are regarded as outliers and deleted.

[0130] Statistical methods are used to eliminate features that have no significant differences among the three categories and are highly correlated, thereby reducing the spatial dimension of the input features and reducing the risk of model overfitting.

[0131] (3) a prediction output module, which is used to select the optimal input features from the preprocessed input features and form an input feature set, and output the prediction results through the prediction model;

[0132] The prediction model uses multiple parallel AdaBoost classifiers as the base classifiers of the prediction model. The base classifier of each AdaBoost classifier is a cost-sensitive decision tree. The misclassification weight of a single sample is the product of the sample weight of the AdaBoost classifier and the weight of the misclassified class in the decision tree.

[0133] Embodiment 3

[0134] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the method for predicting pathogens of bloodstream infection proposed in the above-mentioned embodiment 1 are implemented.

[0135] Embodiment 4

[0136] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for predicting pathogens of bloodstream infection proposed in the first embodiment are implemented.

[0137] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0138] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0139] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0140] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0141] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0142] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting pathogenic bacteria of bloodstream infection, characterized by: The following steps are involved: Obtain and preprocess input features, the input features are physical examination data, including individual information and laboratory examination indicators; Individual information includes age and gender; Laboratory test indicators include: BaseExcess, FreeCalcium, Lactate, pCO2, pH, pO2, %Hemoglobin, A1c, 25-OHVitaminD, AlanineAminotransferase, ALT, Albumin, AlkalinePhosphatase, Alpha-Fetoprotein, Ammonia, AsparateAminotransferase, AST, Bicarbonate, BilirubinDirect, BilirubinInd irect, BilirubinTotal, C-ReactiveProtein, C3 β1C-globulin, C4 β1E-globulin, CalciumTotal, Chloride, CholesterolTotal, CreatineKinase CK, CreatineKinaseMBIsoenzyme, Creatinine, D-Dimer, Ferritin, Glucose, HIVAntibody, ImmunoglobulinA, ImmunoglobulinG, ImmunoglobulinM, Iron, LactateDehydrogenaseLD, Lithium, Magnesium, NTproBNP, Phosphate, Potassium, Sodium, Thyroxine(T4)Free, Triglycerides, Triiodothyronine T3, TroponinI, TroponinT, UreaNitrogen, VitaminB12, AbsolueCD3Count CD3 absolute count, AbsoluteCD4Count CD4 absolute count, AbsoluteCD8Count CD8 absolute count, AbsoluteLymphocyteCount absolute lymphocyte count, Antithrombin antithrombin, Basophils basophils,CD10 Neprilysin, CD103, CD117, CD11c Integrin αX, CD13 Aminopeptidase N, CD138, CD14 Receptor for LPS and LBP complex, CD15 Fucosyltransferase, CD16 / 56, CD16 low affinity Fc receptor, CD16 / 56AbsoluteCount CD16 / 56 absolute count, CD16 / 56% CD16 / 56 percentage, CD19 B lymphocyte surface antigen B4, CD19% B lymphocyte surface antigen B4 percentage, CD19AbsoluteCount B lymphocyte surface antigen B4 absolute count, CD2 Sheep erythrocyte receptor, CD20 B lymphocyte surface antigen B1, CD20% CD20 percentage, CD20AbsoluteCount CD20 absolute count, CD22 T cell surface antigen Leu-14, CD23 low affinity immunoglobulin epsilon Fc receptor, CD25 TAC antigen, CD3% CD3 percentage, CD3AbsoluteCount CD3 absolute count, CD3CellsPercent CD3 cell percentage, CD33 Sialic acid-binding immunoglobulin-like lectin 3, CD34, CD38 ADP ribosyl cyclase 1, CD4CellsPercent CD4 cell percentage, CD4 / CD8Ratio CD4 and CD8bilv, CD41 Integrin α-IIb, CD45 Leukocyte common antigen, CD5, CD5% CD5 percentage, CD5AbsoluteCount CD5 absolute count, CD55 Complement decay accelerating factor, CD56, CD57 Human natural killer cell 1, CD59 MAC-inhibitory protein, CD64 IgGFc receptor I, CD7, CD71 Transferrin receptor protein 1, Eosinophils Eosinophils, HLA-DR Human leukocyte DR antigen, INR (PT) , Lymphocytes lymphocytes, MCH mean corpuscular hemoglobin, MCHC mean corpuscular hemoglobin concentration, MCV mean corpuscular volume, MonocyteCount monocyte count, Neutrophils neutrophils (white) blood cells, PlateletCount platelet count, PT plasma prothrombin time, PTT partial thromboplastin time, Red Blood Cells red blood cells, ReticulocyteCountAbsolute net cell absolute count, WBCCount white blood cell count, White Blood Cells white blood cells; CD is the white blood cell classification antigen;, The optimal input features are selected from the preprocessed input features to form an input feature set, and the prediction model outputs the prediction results, which are the output results of Gram-positive bacteria, Gram-negative bacteria, and fungi; The prediction model is a model that uses multiple first classifiers in parallel as base classifiers of the prediction model, uses a cost-sensitive decision tree as the base classifier of each first classifier, and the misclassification weight of a single sample is the product of the sample weight of the first classifier and the weight of the class misclassified in the decision tree; Multiple parallel AdaBoost classifiers form an integration of the first layer of models. After sampling, the data in the sample is first input into multiple parallel AdaBoost classifiers; and each AdaBoost classifier itself is an integrated model, and its base classifier is a cost-sensitive decision tree, which forms the second layer of integration; Each Adaboost classifier needs to be optimized for hyperparameters. The grid search method is used for hyperparameter optimization. During the optimization process, the classifier is evaluated for model performance using ten-fold cross validation. A cost-sensitive decision tree is used as the base classifier. To prevent overfitting, the maximum depth of the decision tree is set to 2.

2. The method for predicting bloodstream infection pathogens according to claim 1, characterized in that: The preprocessing process includes: Delete features with missing data ratios exceeding the set value and samples with missing values; Through the box plot or scatter plot of the features, outliers are considered as abnormal values ​​and deleted.

3. The method for predicting bloodstream infection pathogens according to claim 2, characterized in that: The preprocessing process also includes: Based on the preprocessed features, it is determined whether the model input indicators are complete and whether these indicators are obtained within the set time period. If so, the preprocessed features are input into the prediction model to obtain the prediction results.

4. The method for predicting bloodstream infection pathogens according to claim 1, characterized in that: The training process of the prediction model includes: Obtain physical examination data of patients with bloodstream infection and remove features that do not differ between bloodstream infection Gram-positive bacteria, Gram-negative bacteria, and fungi; Check and delete the continuous features whose difference coefficients between the selected feature groups are greater than the set value.

5. The method for predicting bloodstream infection pathogens according to claim 4, characterized in that: The training process of the prediction model also includes: Test and delete gender characteristics whose coefficient of difference between the two groups is greater than the set value; Check and remove the correlation coefficient between features Features; The remaining feature data after the test are scaled to a standard normal distribution and divided into a training set and a test set; Input the training set and test set to complete the training of the prediction model.

6. A bloodstream infection pathogen prediction system, characterized by: include: The preprocessing module is configured to: obtain input features and preprocess them, where the input features are physical examination data, including individual information and laboratory examination indicators; Individual information includes age and gender; Laboratory test indicators include: BaseExcess, FreeCalcium, Lactate, pCO2, pH, pO2, %Hemoglobin, A1c, 25-OHVitaminD, AlanineAminotransferase, ALT, Albumin, AlkalinePhosphatase, Alpha-Fetoprotein, Ammonia, AsparateAminotransferase, AST, Bicarbonate, BilirubinDirect, BilirubinInd irect, BilirubinTotal, C-ReactiveProtein, C3 β1C-globulin, C4 β1E-globulin, CalciumTotal, Chloride, CholesterolTotal, CreatineKinase CK, CreatineKinaseMBIsoenzyme, Creatinine, D-Dimer, Ferritin, Glucose, HIVAntibody, ImmunoglobulinA, ImmunoglobulinG, ImmunoglobulinM, Iron, LactateDehydrogenaseLD, Lithium, Magnesium, NTproBNP, Phosphate, Potassium, Sodium, Thyroxine(T4)Free, Triglycerides, Triiodothyronine T3, TroponinI, TroponinT, UreaNitrogen, VitaminB12, AbsolueCD3Count CD3 absolute count, AbsoluteCD4Count CD4 absolute count, AbsoluteCD8Count CD8 absolute count, AbsoluteLymphocyteCount absolute lymphocyte count, Antithrombin antithrombin, Basophils basophils,CD10 Neprilysin, CD103, CD117, CD11c Integrin αX, CD13 Aminopeptidase N, CD138, CD14 Receptor for LPS and LBP complex, CD15 Fucosyltransferase, CD16 / 56, CD16 low affinity Fc receptor, CD16 / 56AbsoluteCount CD16 / 56 absolute count, CD16 / 56% CD16 / 56 percentage, CD19 B lymphocyte surface antigen B4, CD19% B lymphocyte surface antigen B4 percentage, CD19AbsoluteCount B lymphocyte surface antigen B4 absolute count, CD2 Sheep erythrocyte receptor, CD20 B lymphocyte surface antigen B1, CD20% CD20 percentage, CD20AbsoluteCount CD20 absolute count, CD22 T cell surface antigen Leu-14, CD23 low affinity immunoglobulin epsilon Fc receptor, CD25 TAC antigen, CD3% CD3 percentage, CD3AbsoluteCount CD3 absolute count, CD3CellsPercent CD3 cell percentage, CD33 Sialic acid-binding immunoglobulin-like lectin 3, CD34, CD38 ADP ribosyl cyclase 1, CD4CellsPercent CD4 cell percentage, CD4 / CD8Ratio CD4 and CD8bilv, CD41 Integrin α-IIb, CD45 Leukocyte common antigen, CD5, CD5% CD5 percentage, CD5AbsoluteCount CD5 absolute count, CD55 Complement decay accelerating factor, CD56, CD57 Human natural killer cell 1, CD59 MAC-inhibitory protein, CD64 IgGFc receptor I, CD7, CD71 Transferrin receptor protein 1, Eosinophils Eosinophils, HLA-DR Human leukocyte DR antigen, INR (PT) , Lymphocytes lymphocytes, MCH mean corpuscular hemoglobin, MCHC mean corpuscular hemoglobin concentration, MCV mean corpuscular volume, MonocyteCount monocyte count, Neutrophils neutrophils (white) blood cells, PlateletCount platelet count, PT plasma prothrombin time, PTT partial thromboplastin time, Red Blood Cells red blood cells, ReticulocyteCountAbsolute net cell absolute count, WBCCount white blood cell count, White Blood Cells white blood cells; CD is the white blood cell classification antigen;, The prediction output module is configured to: select the optimal input features from the preprocessed input features to form an input feature set, and output the prediction results through the prediction model, and the prediction results are the output results of Gram-positive bacteria, Gram-negative bacteria, and fungi; Preprocessing includes deleting features with missing data ratios exceeding a set value and samples with missing values; using the box plot or scatter plot of the features, outliers are considered as abnormal values ​​and deleted; The prediction model is a model that uses multiple first classifiers in parallel as base classifiers of the prediction model, uses a cost-sensitive decision tree as the base classifier of each first classifier, and the misclassification weight of a single sample is the product of the sample weight of the first classifier and the weight of the class misclassified in the decision tree; the first classifier is an AdaBoost classifier; Multiple parallel AdaBoost classifiers form an integration of the first layer of models. After sampling, the data in the sample is first input into multiple parallel AdaBoost classifiers; and each AdaBoost classifier itself is an integrated model, and its base classifier is a cost-sensitive decision tree, which forms the second layer of integration; Each Adaboost classifier needs to be optimized for hyperparameters. The grid search method is used for hyperparameter optimization. During the optimization process, the classifier is evaluated for model performance using ten-fold cross validation. A cost-sensitive decision tree is used as the base classifier. To prevent overfitting, the maximum depth of the decision tree is set to 2.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for predicting pathogens of bloodstream infection as described in any one of claims 1 to 5 are implemented.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for predicting pathogens of bloodstream infection according to any one of claims 1 to 5 are implemented.

Images