A 5-marker based system for detecting the type of infectious agent

Abstract

The application provides a kind of detection system of infectious pathogen type based on 5 markers, the 5 markers include: GSDMD, p-MLKL, IL-6, PCT, CRP, the system includes the following modules: (1) detection model: the GSDMD of detection sample, p-MLKL, IL-6, PCT, CRP content;(2) analysis module: based on random forest algorithm, neural network algorithm is analyzed to data, constructs diagnosis bacterial infection model;(3) diagnosis module: based on machine learning algorithm model is judged to bacterial infection to be diagnosed sample.The system provided by the application reaches 0.909, the diagnostic accuracy rate, the sensitivity, specificity, positive predictive value and negative predictive value of diagnosis are as follows: 0.95, 0.85, 0.90, 0.92.

Description

Technical Field

[0001] This invention discloses a pathogen type detection system, belonging to the fields of microbiology and artificial intelligence technology. Background Technology

[0002] Early and accurate diagnosis of bloodstream infection (BSI) is crucial for improving prognosis. Currently, routinely used biomarkers in clinical testing include procalcitonin (PCT), acute-phase reactive protein (CRP), and interleukin-6 (IL-6). While PCT is currently recognized as the best biomarker for bacterial infection with the highest diagnostic accuracy, its clinical application has limitations: ① Its overall specificity for identifying pathogen types is only about 70% (meta-analysis data); ② It often shows a delayed increase within 24 hours of infection, making it more suitable for assessing the severity of sepsis than for early diagnosis of bloodstream infection; ③ Notably, non-infectious factors (such as burns, intestinal ischemia, and postoperative trauma) can also lead to abnormally elevated PCT levels. Therefore, the sensitivity (generally <85%) and specificity (usually <90%) of existing biomarkers have not yet met the diagnostic standards for bloodstream infection recommended by ESCMID (European Society for Clinical Microbiology and Infectious Diseases), necessitating the discovery and validation of novel biomarkers.

[0003] The purpose of this invention is to provide a more sensitive and specific pathogen type detection system. Summary of the Invention

[0004] To achieve the above objectives, the present invention first provides a pathogen infection type detection system based on five markers, namely: GSDMD, p-MLKL, IL-6, PCT, and CRP. The system includes the following modules:

[0005] (1) Training dataset input module, the module is used to input sample data and construct training dataset. Each sample data in the training dataset contains five input features: the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP, as well as a binary diagnostic label for bacterial infection and non-bacterial infection of the sample.

[0006] (2) Training module, which is used to train the training dataset constructed by the training dataset input module through machine learning to establish a binary classification prediction model for bacterial infection and non-bacterial infection based on five input features of detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP.

[0007] (3) Prediction module: Based on the prediction model obtained from the training module, the module makes a prediction of bacterial infection or non-bacterial infection based on the five input features of the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP of the input sample to be detected.

[0008] In a preferred embodiment, the machine learning method in the training module employs a random forest classification model. This model consists of 100 decision trees as base learners. Each decision tree generates a subset of training data by randomly sampling from the training dataset using bootstrap sampling. At each node split, two features are randomly selected from all features for optimal splitting. The minimum number of split samples per node is set to 2. Each decision tree is trained independently. Furthermore, the weights of each category are automatically adjusted to balance their influence during model training. Binary classification diagnostic labels are predicted using majority voting.

[0009] In the prediction module, the binary diagnostic label of the sample to be detected is predicted using the majority voting method consistent with the training module, based on the five input features of the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP of the sample to be detected.

[0010] In a more preferred embodiment, in the training module, each decision tree uses the Gini index as the splitting criterion when splitting a node, and the Gini index of node t is defined as:

[0011] ,

[0012] Where, p k Let t represent the proportion of samples of class k in node t, where K is the total number of classes. In each split, the feature and split point that minimizes the weighted average Gini index are selected.

[0013] In another more preferred embodiment, the weight class adjustment in the training module is set as follows:

[0014] ,

[0015] Where, n samples Let n be the total number of samples, K be the number of categories, and n be the total number of samples. c Let be the number of samples in class c.

[0016] In yet another, more preferred embodiment, in the training module, the predicted category is determined by majority voting:

[0017] ,

[0018] in, To predict the category, C is the set of all possible categories, and I(.) is the indicator function, when y i The value is 1 if it equals c, otherwise it is 0.

[0019] In another preferred embodiment, the machine learning method in the training module employs a multi-layer feedforward neural network. The training module includes an input layer with five nodes containing detection values ​​for GSDMD, p-MLKL, IL-6, PCT, and CRP, one to two hidden layers, and one output layer. The output layer has one node, and the activation function is the sigmoid function, achieving binary classification probability output. The multi-layer feedforward neural network uses stochastic gradient descent (SGD) for backpropagation, with a learning rate set to 0.001. L2 regularization (weight decay) is used to prevent overfitting. A categorical cross-entropy loss function is calculated. When the loss function is observed to stabilize, the model converges, training stops, and an independent test dataset is established.

[0020] In the prediction module, five input features—the detected values ​​of GSDMD, p-MLKL, IL-6, PCT, and CRP—of the sample to be detected are input into the neural network. After calculation by each layer, the output is a probability value (ranging from 0 to 1). A judgment threshold is set (e.g., 0.5). If the output probability is greater than or equal to the threshold, it is judged as a positive class (1, bacterial infection); otherwise, it is judged as a negative class (0, non-bacterial infection).

[0021] In another preferred embodiment, each hidden layer contains 10-32 neurons, and the activation function is the sigmoid function.

[0022] ,

[0023] The output of the neuron is:

[0024] ,

[0025] in, This is the output of the i-th neuron in the previous layer. As weight, σ is the bias, and σ is the activation function.

[0026] In another, more preferred embodiment, the categorical cross-entropy loss function is:

[0027] ,

[0028] Among them, y trueTrue diagnostic label: 0 or 1, y pred Predict probabilities for the model.

[0029] In another preferred embodiment, the prediction module performs feature normalization processing on the sample to be predicted, consistent with the training phase. The sample is input into the neural network, and after calculation by each layer, the output is a probability value in the range of 0 to 1. A judgment threshold is set. If the output probability is greater than or equal to the threshold, it is judged as a positive class (1, bacterial infection); otherwise, it is judged as a negative class (0, non-bacterial infection).

[0030] More preferably, the determination threshold is 0.5.

[0031] This invention addresses the clinical challenge of early diagnosis of bloodstream infections by providing a combination of five biomarkers, including two novel biomarkers: gasdermin D (GSDMD) and phosphorylated mixed kinase domain-like protein (p-MLKL), and three conventional biomarkers: PCT, CRP, and IL-6. GSDMD and p-MLKL are key proteins activated in the pyroptosis and necroptosis pathways of host cells following microbial infection, respectively. GSDMD is the core executor of pyroptosis. Bacterial infection activates caspase-4 / 5 / 11, cleaving GSDMD and releasing its N-terminal domain. The N-terminal fragment oligomerizes and forms pores in the cell membrane, leading to ion imbalance, cell swelling and rupture, and the release of pro-inflammatory factors such as IL-1β and IL-18, ultimately mediating cell membrane perforation and directly initiating pyroptosis. In the necroptosis pathway, bacteria stimulate death receptor (such as TNFR1) signaling, activating MLKL via the RIPK1-RIPK3 complex to generate p-MLKL. This p-MLKL oligomers undergo conformational changes and oligomerization, translocating to the cell membrane, disrupting membrane integrity, leading to permeable necrosis and the release of inflammatory factors. Pyroptosis and necroptosis play important roles in regulating the host cell's innate immune response to infection, but excessive activation can lead to inflammatory diseases such as sepsis and autoimmune disorders.

[0032] The technical solution of this invention integrates five markers through a machine learning algorithm to establish a high-precision diagnostic model with an accuracy of >95%, which is significantly better than the current detection scheme. Attached Figure Description

[0033] Figure 1 A schematic diagram of an infectious pathogen type detection system based on five markers;

[0034] Figure 2Comparison of ROC curves for 5 individual biomarkers and AUC values ​​for each biomarker; ROC, receiver operating characteristic curve; AUC, area under the curve;

[0035] Figure 3 Comparison of ROC curves and AUC values ​​for each of the three biomarker combinations. The random forest model constructed from the combination of five biomarkers showed the highest AUC value in diagnosing bloodstream bacterial infection. (A) Model features constructed from the combination of PCT, CRP, and IL-6; (B) Model features constructed from the combination of GSDMD and p-MLK; (C) Model features constructed from the combination of five biomarkers: 1. Model error tree diagram; 2. Ranking of the importance of each variable; 3. ROC and AUC of the model. ROC: Subject characteristic operating curve; AUC: Area under the curve. Detailed Implementation

[0036] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result of the description. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of protection defined by the claims of the present invention.

[0037] Figure 1 The flowchart of the pathogen infection type detection system based on random forest algorithm or neural network algorithm and 5 markers of the present invention is given. The specific implementation scheme is described below.

[0038] Example 1. Sample collection and testing

[0039] 1. Collect 100 cases of bloodstream bacterial infection (including Escherichia coli, Klebsiella pneumoniae, Pseudomonas aeruginosa, Enterococcus faecalis, etc.), 150 cases of non-bacterial infection, including 100 cases of simple viral infection (influenza A virus, herpes simplex virus, Epstein-Barr virus, etc.), and 50 cases without infection.

[0040] 2. Collect whole blood / plasma from febrile patients and detect the levels of GSDMD (gasdermin D), p-MLKL (phosphorylated mixed-series kinase domain-like protein), IL-6 (interleukin-6), PCT (procalcitonin), and CRP (acute phase reactive protein) using conventional ELISA methods. The specific methods are as follows:

[0041] (1) Use human GSDMD, p-MLKL, PCT, CRP, IL-6 ELISA kit (COBIO Biotechnology Co., Ltd., Shanghai, China);

[0042] (2) Remove the kit from 4°C and allow it to equilibrate to room temperature for 30 minutes (avoid direct exposure to light);

[0043] (3) Dilute the concentrated washing solution with double-distilled water to the working concentration (e.g., 20× concentrate needs to be diluted at a ratio of 1:19).

[0044] (4) Plasma: After collection, centrifuge (2000 × g, 20 minutes) to remove the precipitate, aliquot and store at -80℃; for whole blood, take 300 µL and extract total protein using a whole blood total protein extraction kit (Solepro Biotechnology Co., Ltd., Beijing, China) and store at -80℃.

[0045] (5) Preparation of standard: Dilute the standard with standard diluent in a gradient to different concentrations (e.g., 0, 50, 100, 200, 500, 1000 pg / mL) to establish a standard curve;

[0046] (6) Antibody incubation: Place the 96-well plate pre-coated with anti-GSDMD antibody at room temperature and gently tap to remove the protective solution;

[0047] (7) Sample addition: Standard wells: Add 100 μL of different concentrations of standard; Sample wells: Add 100 μL of the sample to be tested (pre-diluted); Blank wells: Add 100 μL of sample diluent as a blank control; Negative / positive control wells: Add according to the instructions.

[0048] (8) Cover with film and incubate at 37°C for 90 minutes;

[0049] (9) Wash, discard the liquid, add 300 μL of washing solution to each well, let stand for 30 seconds and then discard, repeat 3 times (it is recommended to use a multi-channel pipette or an automatic plate washer).

[0050] (10) Add detection antibody (biotin-labeled), add 100 μL of biotinylated anti-human GSDMD / p-MLKL / PCT / CRP / IL-6 antibody working solution to each well;

[0051] (11) Cover with film, incubate at 37°C for 60 minutes, and wash 3 times again;

[0052] (12) Add streptavidin-HRP, add 100 μL of enzyme conjugate (HRP-labeled streptavidin) to each well.

[0053] (13) Cover with film to avoid light, incubate at 37°C for 30 minutes, and wash 5 times (to thoroughly reduce non-specific binding).

[0054] (14) Develop color (TMB): Add 90 μL of TMB substrate to each well and incubate at room temperature in the dark for 15-30 minutes (it will turn blue; the time needs to be optimized to avoid it being too dark).

[0055] (15) To terminate the reaction, add 50 μL of stop solution (2M H2SO4) to each well; the color will change from blue to yellow.

[0056] (16) Read the plate and read the absorbance (OD value) at 450 nm using an ELISA reader within 30 minutes. Refer to 630 nm to correct for differences between wells.

[0057] (17) Data Analysis:

[0058] The standard curve was plotted with the standard concentration as the x-axis (logarithmic scale) and the corresponding OD value as the y-axis, and a four-parameter logic (4PL) curve fitting was used.

[0059] To calculate the sample concentration, substitute the sample OD value into the standard curve equation and multiply by the dilution factor.

[0060] Example 2. Construction of the Random Forest Model

[0061] Embodiment 2 of this invention employs a Random Forest classification model as the core prediction algorithm. The Random Forest model consists of several decision tree base learners, with the following specific structure:

[0062] (1) Random Forest Model Structure:

[0063] The random forest model contains N decision trees, denoted as {T1,T2,...,T...}. N Each decision tree generates a training subset by randomly sampling from the original training dataset using bootstrap sampling. At each node split, a subset of features is randomly selected from all features for optimal splitting. Each decision tree is trained independently, and the final prediction result is output through an ensemble voting method.

[0064] (2) Setting of main technical parameters

[0065] 1) The number of trees is set to 100, meaning the model contains 100 decision trees;

[0066] 2) Maximum tree depth: Set to 10;

[0067] 3) Number of splitting features: Set to the square root of the number of features, that is, each time a node splits, features are randomly selected from all features. There are 5 features (M is the total number of features). In this embodiment, M=5, so about 2 features are randomly selected for each split.

[0068] 4) Minimum number of split samples for a node: set to 2, meaning that each internal node must contain at least 2 samples before it is allowed to continue splitting;

[0069] 5) Category weights: Set to "balanced" to automatically adjust the weights of each category.

[0070] (3) Model training

[0071] 1) Input features and processing

[0072] A training dataset was constructed, with each sample containing five input features (GSDMD, p-MLKL, IL-6, PCT, CRP) and a binary diagnostic label (bacterial infection and non-bacterial infection). First, the data was cleaned. Outliers were detected using the interquartile range (IQR) method; values ​​exceeding Q3 + 1.5 × IQR were considered outliers and removed. Then, the training samples were input into each decision tree, and the classification results for each tree were obtained.

[0073] 2) Decision tree splitting criteria

[0074] Each decision tree uses either the Gini Index or Information Gain as the splitting criterion when splitting a node. Taking the Gini Index as an example, the Gini Index of node t is defined as:

[0075] ,

[0076] Where, p k Let K be the proportion of samples of class k in node t, where K is the total number of classes (2 in this example). Each split selects the feature and split point with the smallest weighted average Gini index.

[0077] 3) Weight category adjustment

[0078] ,

[0079] Where, n samples K is the total number of samples, K is the number of categories (2 in this example), and n is the number of categories. c This represents the number of samples in class c (i.e., the number of cases in each class, such as n1=100, meaning 100 cases in the bacterial infection group, and n0=200, meaning 200 cases in the non-bacterial infection group). This weight is used to balance the influence of each class during model training.

[0080] 4) Final prediction category determination

[0081] The final predicted category is determined by majority voting, as follows:

[0082] ,

[0083] Where C is the set of all possible categories (in this embodiment, it includes both bacterial infection and non-bacterial infection), and I(.) is an indicator function, when y i =c, take 1 (indicating bacterial infection); otherwise, take 0 (indicating non-bacterial infection).

[0084] (4) Model evaluation

[0085] Establish an independent test dataset to verify the model's performance on the test set. Use specificity, sensitivity, positive predictive value (PPV), negative predictive value (NPV), and area under the ROC curve (AUC) to comprehensively evaluate the model's predictive performance.

[0086] (5) Model prediction

[0087] Perform the same data processing on the samples to be predicted as during the training phase. Input the data into the model and output the final prediction result using the same majority voting method as during the training phase. When y i =c, take 1 (indicating bacterial infection); otherwise, take 0 (indicating non-bacterial infection).

[0088] Example 3. Construction of a multilayer feedforward neural network model

[0089] Embodiment 3 of the present invention further provides a prediction model based on a neural network (NN). The neural network model adopts a multi-layer perceptron (MLP) structure, including an input layer, several hidden layers, and an output layer. It is described below:

[0090] (1) Network structure and main parameter settings:

[0091] 1) Input Layer: The number of nodes is equal to the number of input features. In this embodiment, there are 5 nodes, namely GSDMD, p-MLKL, PCT, CRP, and IL-6.

[0092] 2) Hidden Layers: One to two hidden layers are used, each containing 10-32 neurons. The activation function is the sigmoid function.

[0093] ,

[0094] Where z is the input of the function, y is the output of the function, and e is the base of the natural logarithm.

[0095] The output of a neuron, taking the j-th neuron in the l-th layer as an example:

[0096] ,

[0097] in, This is the output of the i-th neuron in the previous layer. The weights are randomly initialized and then updated through backpropagation. σ is the bias (the weight parameters are randomly initialized and then updated through backpropagation), and σ is the Sigmoid activation function.

[0098] 3) Output Layer: Set up 1 node. The activation function of the output layer is the Sigmoid function, which realizes the binary classification probability output.

[0099] (2) Model training:

[0100] 1) Input Data Processing: Construct a training dataset. Each sample in the training set contains 5 input features (GSDMD, p-MLKL, IL-6, PCT, CRP) and a binary diagnostic label (bacterial infection and non-bacterial infection). The input features are normalized, and the diagnostic label is encoded using 0-1.

[0101] 2) Loss Function: The cross-entropy loss function is used.

[0102] ,

[0103] Among them, y true For the true diagnostic label (0 or 1, 1 indicates bacterial infection, 0 indicates non-bacterial infection), y pred The predicted probability output by the model (the probability value is a floating-point number between 0 and 1).

[0104] 3) Optimization Algorithm (Optimizer): Backpropagation is performed using the Stochastic Gradient Descent (SGD) method. The SGD weight update formula is as follows:

[0105] ,

[0106] Where θ represents the model parameters, η represents the learning rate, and L... i (θ) is the loss for the i-th sample. This is the gradient operator. In this example, the learning rate is set to 0.001.

[0107] 4) Regularization method: L2 regularization (weight decay) is used to prevent overfitting. The loss function after L2 regularization is as follows:

[0108] ,

[0109] Where L is the original loss function, θ is the model parameters, and λ is the regularization coefficient. In this example, the regularization coefficient λ can be set to 0.0001.

[0110] (3) Model convergence: When the loss function is observed to be stable, the model converges and training stops.

[0111] (4) Model evaluation

[0112] Establish an independent test dataset to verify the model's performance on the test set. Use specificity, sensitivity, positive predictive value (PPV), negative predictive value (NPV), and area under the ROC curve (AUC) to comprehensively evaluate the model's predictive performance.

[0113] (5) Model prediction

[0114] The samples to be predicted undergo feature normalization processing consistent with the training phase. The samples are input into the neural network, and after calculation by each layer, the output is a probability value (ranging from 0 to 1). A decision threshold is set (e.g., 0.5). If the output probability is greater than or equal to the threshold, it is classified as positive (1, bacterial infection); otherwise, it is classified as negative (0, non-bacterial infection).

[0115] Example 4. Evaluation of the diagnostic capability of various biomarker combinations in a random forest model.

[0116] This invention is characterized by a combination of five biomarkers in plasma / whole blood for the diagnosis of bloodstream bacterial infections, whereas conventional biomarkers for the diagnosis of bloodstream infections only include PCT, CRP, and IL-6.

[0117] 1. The key to this invention: the combined use of five biomarkers for diagnosing bloodstream bacterial infections, including GSDMD (gasdermin D), p-MLKL (phosphorylated mixed series kinase domain-like protein), IL-6 (interleukin-6), PCT (procalcitonin), and CRP (acute phase reactant protein).

[0118] 2. Collect plasma from febrile patients and use conventional ELISA to detect the levels of GSDMD (gasdermin D), p-MLKL (phosphorylated mixed series kinase domain-like protein), IL-6 (interleukin-6), PCT (procalcitonin), and CRP (acute phase reactant protein) in plasma / whole blood.

[0119] 3. Model Building: Using plasma / whole blood GSDMD, p-MLKL, IL-6, PCT, and CRP values ​​from patients with bacterial infections and non-bacterial infection controls (patients with viral infections and those without infections) as datasets, the datasets were divided into a training set (210 cases) and a test set (90 cases). The training set comprised 70% of the total dataset and was used to construct a diagnostic model using random forest and neural network algorithms. The test set comprised 30% of the total dataset and was used to evaluate and validate the established model.

[0120] 4. Diagnosis of new cases: Plasma was collected from febrile patients, and the levels of GSDMD, p-MLKL, IL-6, PCT, and CRP were detected by ELISA.

[0121] Random Forest Model Diagnostic Method: The five variable values ​​of the sample to be predicted are input into the model, and the final prediction result is output using the majority voting method consistent with the training phase. Among these... To predict the category, C is the set of all possible categories (2 in this example), when y i =c, take 1 (indicating bacterial infection); otherwise, take 0 (indicating non-bacterial infection).

[0122] Neural network model diagnostic method: The sample is input into the neural network, and after calculation by each layer, the output is a probability value (ranging from 0 to 1). If the output probability is greater than or equal to the threshold (0.5), it is judged as a positive class (1, bacterial infection); otherwise, it is judged as a negative class (0, non-bacterial infection).

[0123] 5. The inventors constructed a random forest model with five variables, using positive blood cultures and the detection of bacterial nucleic acid sequences in blood pathogen screening as the criteria. The diagnostic results for new cases (test set) are shown in Table 1. The diagnostic accuracy of the combination of five biomarkers reached 0.909 (95% confidence interval: 0.757, 0.981), with sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of 0.95, 0.85, 0.90, and 0.92, respectively. In contrast, the diagnostic accuracy of the combination of three traditional biomarkers was only 0.758 (95% confidence interval: 0.577, 0.889), with sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of 0.75, 0.77, 0.83, and 0.67, respectively.

[0124] Table 1. Systematic evaluation of the diagnostic capability of various biomarker combinations in the random forest model

[0125] .

[0126] 95% CI, 95% confidence interval; PPV, Positive Predictive Value; NPV, Negative Predictive Value.

[0127] Figure 2 and Figure 3 The results of various detection methods on the subject characteristic operating curve (ROC) and area under the curve (ACU) are also shown. Figure 2 Comparison of ROC curves for five individual markers and AUC values ​​for each marker. Figure 3ROC curves for three combinations of biomarkers and comparisons of AUC values ​​for each biomarker are presented. Specifically: (A) Model features constructed using the combination of PCT, CRP, and IL-6; (B) Model features constructed using the combination of GSDMD and p-MLK; (C) Model features constructed using the combination of five biomarkers. 1. Model error tree diagram; 2. Ranking of the importance of each variable; 3. ROC and ACU of the model. Figure 2 and Figure 3 As can be seen, this invention uses a combination of five biomarkers to diagnose bloodstream bacterial infections, and its diagnostic AUC value is significantly higher than that of the conventional PCT (0.788). Figure 2 Increased to 0.962 ( Figure 3 (C-3), with an accuracy of >95%, significantly improving the diagnostic accuracy of bloodstream bacterial infections. Figure 3 Among them, the random forest model constructed by combining 5 biomarkers had the highest AUC value for diagnosing bloodstream bacterial infection (0.962), which was significantly higher than that of the PCT, CRP and IL-6 group (0.765) and the GSDMD and p-MLK group (0.9).

[0128] The following examples illustrate that five methods for detecting combined biomarkers are superior to traditional methods for detecting single biomarkers or traditional methods for detecting combined biomarkers:

[0129] Case 1. A patient with myelodysplastic syndrome, hospitalized for bone marrow transplantation, suddenly developed a high fever. Plasma PCT, CRP, and IL-6 levels were 0.4 ng / mL (mildly elevated), 1.8 mg / L (normal), and 10 pg / mL (mildly elevated), respectively. Using these three variables in the model, the infection was classified as non-bacterial. However, plasma GSDMD and p-MLKL levels were 1593.907 ng / mL and 630.666 ng / mL, respectively. Using all five biomarkers in the model simultaneously, the infection was classified as bacterial. Subsequent high-throughput sequencing of whole blood pathogens revealed Pseudomonas aeruginosa infection, with a bacterial nucleic acid sequence count of 610.

[0130] Case 2. A patient with multiple myeloma was admitted to the hospital with a fever of 38.8 degrees Celsius. Plasma PCT, CRP, and IL-6 levels were measured at 0.36 ng / mL (slightly elevated), 3.1 mg / L (slightly elevated), and 14.4 pg / mL (elevated), respectively. Using these three variables in the model, a bacterial infection was identified. However, plasma GSDMD and p-MLKL levels were measured at 68.217 ng / mL and 16.407 ng / mL, respectively. Using all five biomarkers in the model simultaneously, a non-bacterial infection was identified. Subsequent high-throughput sequencing of whole blood confirmed the absence of pathogen nucleic acid sequences in the patient's blood.

[0131] Case 3. A patient with acute lymphoblastic leukemia presented with persistent low-grade fever (38.0-38.5 degrees Celsius) during hospitalization. Plasma PCT, CRP, and IL-6 levels were 0.09 ng / mL (slightly elevated), 1.3 mg / L (normal), and 4.6 pg / mL (normal), respectively. Using these three variables in the model, the infection was initially classified as non-bacterial. However, plasma GSDMD and p-MLKL levels were 3188.466 ng / mL and 761.948 ng / mL, respectively. Using all five biomarkers in the model simultaneously, a bacterial infection was identified. Subsequent high-throughput sequencing of whole blood revealed Klebsiella pneumoniae infection, with a bacterial nucleic acid sequence count of 870.

Claims

1. A pathogen infection type detection system based on five markers for non-diagnostic purposes, characterized in that, The five markers include: GSDMD, p-MLKL, IL-6, PCT, and CRP. The system includes the following modules: (1) Training dataset input module, the module is used to input sample data and construct training dataset. Each sample data in the training dataset contains five input features: the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP, as well as a binary diagnostic label for bacterial infection and non-bacterial infection of the sample. (2) Training module, which is used to train the training dataset constructed by the training dataset input module through machine learning to establish a binary classification prediction model for bacterial infection and non-bacterial infection based on five input features of detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP. (3) Prediction module: Based on the prediction model obtained from the training module, the module predicts a binary diagnostic label of bacterial infection or non-bacterial infection based on the five input features of the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP of the input sample to be detected; The machine learning method used in the training module employs a random forest classification model. This model consists of 100 decision trees as base learners. Each decision tree generates a subset of training data by randomly sampling from the training dataset using a bootstrap sampling method. At each node split, two features are randomly selected from all features for optimal splitting. The minimum number of split samples per node is set to 2. Each decision tree is trained independently. Furthermore, the weights of each category are automatically adjusted to balance their influence during model training. Binary classification diagnostic labels are predicted using a majority voting method. In the prediction module, the binary diagnostic label of the sample to be detected is predicted using the majority voting method consistent with the training module, based on the five input features of the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP of the sample to be detected.

2. The pathogen infection type detection system based on five markers according to claim 1, characterized in that, In the training module, the Gini index is used as the splitting criterion when each decision tree splits a node. The Gini index of node t is defined as: ， in, Let t represent the proportion of samples of class k in node t, where K is the total number of classes. In each split, the feature and split point that minimizes the weighted average Gini index are selected.

3. The pathogen infection type detection system based on five markers according to claim 1, characterized in that, In the training module, the weight class adjustment is set as follows: ， in, Let K be the total number of samples, K be the number of categories, and c be the number of samples in the c-th category.

4. The pathogen infection type detection system based on five markers according to claim 1, characterized in that, In the training module, the predicted category is determined by majority voting: ， in, To predict the category, C is the set of all possible categories, and I(.) is an indicator function that takes the value 1 if the value is true and 0 otherwise.

5. A pathogen infection type detection system based on five markers for non-diagnostic purposes, characterized in that, The five markers include: GSDMD, p-MLKL, IL-6, PCT, and CRP. The system includes the following modules: (1) Training dataset input module, the module is used to input sample data and construct training dataset. Each sample data in the training dataset contains five input features: the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP, as well as a binary diagnostic label for bacterial infection and non-bacterial infection of the sample. (2) Training module, which is used to train the training dataset constructed by the training dataset input module through machine learning to establish a binary classification prediction model for bacterial infection and non-bacterial infection based on five input features of detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP. (3) Prediction module: Based on the prediction model obtained from the training module, the module predicts a binary diagnostic label of bacterial infection or non-bacterial infection based on the five input features of the detection values ​​of GSDMD, p-MLKL, IL-6, PCT and CRP of the input sample to be detected; The machine learning method used in the training module employs a multi-layer feedforward neural network. This training module includes an input layer with five nodes containing detection values ​​for GSDMD, p-MLKL, IL-6, PCT, and CRP, one to two hidden layers, and one output layer. The output layer has one node, and the sigmoid function is used as the activation function to achieve binary classification probability output. The multi-layer feedforward neural network uses stochastic gradient descent for backpropagation, with a learning rate of 0.001, and L2 regularization is used to prevent overfitting. The cross-entropy loss function is calculated, and when the loss function is observed to stabilize, the model converges, training stops, and an independent test dataset is established. In the prediction module, five input features—the detected values ​​of GSDMD, p-MLKL, IL-6, PCT, and CRP of the sample to be detected—are input into the neural network. After calculation by each layer, the output is a probability of 0 to 1. A judgment threshold is set. If the output probability is greater than or equal to the threshold, it is judged as positive class: 1, that is, bacterial infection; otherwise, it is judged as negative class: 0, that is, non-bacterial infection.

6. The pathogen infection type detection system based on five markers according to claim 5, characterized in that, Each hidden layer contains 10-32 neurons, and the activation function is the sigmoid function. The output of the neuron is: in, This is the output of the i-th neuron in the previous layer. As weight, For bias, This is the activation function.

7. The pathogen infection type detection system based on five markers according to claim 5, characterized in that, The cross-entropy loss function is: in, True diagnostic label: 0 or 1 Predict probabilities for the model.

8. The pathogen infection type detection system based on five markers according to claim 5, characterized in that, The prediction module performs feature normalization processing on the samples to be predicted, consistent with the training phase. The samples are input into the neural network, and after calculation by each layer, the output is a probability value in the range of 0 to 1. A judgment threshold is set. If the output probability is greater than or equal to the threshold, it is judged as positive class: 1, that is, bacterial infection; otherwise, it is judged as negative class: 0, that is, non-bacterial infection.

9. The pathogen infection type detection system based on five markers according to claim 8, characterized in that, The determination threshold is 0.5.

Images