Damage time prediction model and construction method and application thereof

Abstract

The invention relates to the field of biomedicine, and discloses an injury time prediction model and a construction method and application thereof. The construction of the model comprises the following steps: carrying out data preprocessing on a data set of abundance values of metabolites of excrement and serum; constructing at least two of a support vector regression model, a random forest regression model and a gradient boosting regression model; and performing hyper-parameter tuning and model training. The prediction model can accurately predict the injury time by fusing excrement and serum metabolite data. A support vector regression (SVR) model of an excrement group has extremely high prediction accuracy, the determination coefficient (R2) of the SVR model reaches 0.9630, 96.30% of data variation can be explained, and the SVR model is remarkably higher than a traditional method. The random forest model of the serogroup also provides valuable supplementary information for prediction. The successful construction of the model provides a new thought and method for the inference of the injury time in forensic medicine.

Description

Technical Field

[0001] The invention belongs to the field of biomedicine, and specifically relates to an injury time prediction model and a construction method and application thereof. Background Art

[0002] In recent years, forensic research has gradually deepened the metabolic changes after tissue injury, which has important application value in injury identification and injury time inference. Studies have shown that by analyzing the dynamic changes of metabolites after injury, the physiological and biochemical reaction processes of the body can be revealed, providing an objective and scientific basis for forensic medicine. The development of metabolomics technology, especially the progress of high-throughput and large-scale metabolite detection, makes it possible to study metabolic patterns after injury, thus providing new tools and methods for forensic medicine. However, existing technologies still face many challenges and shortcomings in application.

[0003] In the mouse model, feces and serum are commonly used biological samples that can reflect the intestinal microbial environment and systemic metabolic status, respectively. Metabolites in feces are mainly derived from the metabolic activities of intestinal microorganisms, and thus can reflect changes in the intestinal environment and microbial communities; while serum metabolites can provide clues to the body's systemic metabolic state. Therefore, metabolic analysis of these biological samples has great application potential in forensic medicine, especially in the evaluation of responses after tissue damage. However, metabolic changes at different time points after injury are extremely complex, and data characteristics are highly individual and time-series, making it quite difficult to accurately establish a prediction model suitable for forensic needs.

[0004] With the widespread application of machine learning methods in the biomedical field, especially in pattern recognition and prediction tasks of high-dimensional and complex data, the field of forensic medicine has attempted to introduce methods such as support vector machines (SVM), random forests, and gradient boosting. These methods have the advantages of processing small samples and high-dimensional data, and can predict the time of injury based on multi-omics data. However, traditional machine learning models still have significant shortcomings in the practical application of forensic medicine: on the one hand, the metabolic changes of different biological samples at different time points after injury are quite diverse, and existing algorithms are difficult to effectively identify the key features of these changes; on the other hand, different algorithms have different performances in accuracy and stability, and the prediction effect of a single model is often not ideal.

[0005] In addition, feature selection is an essential step in the model building process. Existing technologies mostly use methods such as analysis of variance and partial least squares discriminant analysis (PLS-DA) to screen key features, but when faced with complex metabolic data, these methods may have problems such as unstable feature selection and poor interpretability, which limits the prediction accuracy and scope of application of the model. Summary of the invention

[0006] The present invention provides a method for constructing a damage time prediction model to overcome the problems of the above-mentioned prior art in processing high-dimensional complex data, single model limitations, unstable feature selection, and low calculation efficiency;

[0007] Another object of the present invention is to provide a damage time prediction model;

[0008] Another object of the present invention is to provide an application of a damage time prediction model;

[0009] Another object of the present invention is to provide a method for predicting damage time.

[0010] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0011] A method for constructing a damage time prediction model comprises the following steps:

[0012] S1. Preprocessing the data sets of feces and serum metabolite abundance values;

[0013] S2. Construct at least two models among support vector regression model, random forest regression model and gradient boosting regression model;

[0014] S3, hyperparameter tuning and final model training;

[0015] S4. The prediction results of each model are integrated and output.

[0016] Furthermore, the feces and serum are collected at different time points after injury.

[0017] Preferably, feces and serum are collected 1 day, 3 days, 7 days, and 11 days after injury.

[0018] Furthermore, the metabolites include lipid metabolites, amino acids and their derivatives, organic acid metabolites, nucleic acids and their metabolites, phenolic and polyphenolic compounds, and sugars.

[0019] Preferably, among the fecal metabolites, lipid metabolites include coenzyme Q, lysophosphatidylcholine (glycerophospholipids), bile acid, amino acids and their derivatives include aminocaprylic acid, glutamine, tyrosine, glutamic acid, organic acid metabolites include 2,3-dihydroxyisovaleric acid, gluconate, and also include alkaloids.

[0020] Preferably, among the serum metabolites, lipid metabolites include 3-hydroxyhexadecanoic acid, amino acids and their derivatives include DL-isoleucine, nucleic acids and their metabolites include uracil, phenolic and polyphenolic compounds include ferulic acid, 4-vinylphenol sulfate, isoflavones, sugars include glucose, and also include 1-naphthylamine.

[0021] In the fecal group model, lipid metabolites (such as glycerophospholipids) were identified as key features. These metabolites play an important role in cell membrane stability, signal transduction, and inflammatory response. After injury, the body needs to quickly repair damaged cell membranes, and changes in lipid metabolism reflect this physiological need. In addition, changes in amino acids and their derivatives suggest that the body has an increased demand for protein synthesis and immune regulation after injury. Although the serum group model has lower prediction accuracy than the fecal group, it still provides valuable information. Changes in metabolites in serum more reflect systemic metabolic responses, including processes such as energy metabolism, oxidative stress, and inflammatory response. The dynamic changes in serum metabolites provide a supplement for a comprehensive understanding of the physiological response after injury.

[0022] Preferably, the dataset contains sample names, feature columns, label columns, and statistical significance rows.

[0023] Furthermore, data preprocessing is performed through feature weighting and neighborhood component analysis for feature selection.

[0024] Preferably, data preprocessing includes the following steps: reading and integrating, standardizing, and feature weighting the data set of metabolite abundance values, and after obtaining the weighted feature matrix, performing feature selection using neighborhood component analysis.

[0025] Further preferably, the reading and integration includes the following steps: reading data, extracting feature names, obtaining significance information, extracting feature data, data type conversion, feature alignment, and feature weighting.

[0026] Further preferably, the standardization comprises the following steps: calculating the mean and standard deviation, processing constant value features, substituting into a standardization formula, and checking NaN values.

[0027] Furthermore, neighborhood component analysis is performed on the weighted feature data and labels; the weight of each feature is obtained; features with weights greater than 1% of the maximum weight are selected as thresholds; and features are filtered according to the thresholds.

[0028] Preferably, the fsrnca function of MATLAB is used to perform neighborhood component analysis on the weighted feature data and labels.

[0029] Preferably, the support vector regression model sets the following hyperparameters:

[0030] Kernel function: linear, rbf, polynomial;

[0031] Penalty parameter C: 0.01, 0.1, 1, 10;

[0032] ε-insensitive loss function parameters: 0.01, 0.1, 0.5, 1.

[0033] The random forest regression model sets the following hyperparameters:

[0034] Number of decision trees: 50, 100, 200;

[0035] Minimum number of leaf node samples: 1, 5, 10.

[0036] The gradient boosting regression model sets the following hyperparameters:

[0037] Number of learning cycles: 50, 100, 200;

[0038] Learning rate: 0.01, 0.1, 0.20.

[0039] Further preferably, when training the feces group model, a support vector regression model is constructed, and the following hyperparameters are set: kernel function: linear; penalty parameter C: 1; ε-insensitive loss function parameter: 0.1.

[0040] Further preferably, when training the serum group model, a random forest regression model is constructed, and the following hyperparameters are set: number of decision trees: 50; minimum number of leaf node samples: 1.

[0041] Preferably, cross-validation and model evaluation are performed on the constructed model.

[0042] Furthermore, 10-fold cross validation is used to ensure the stability of model evaluation; the evaluation indicators include mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), determination coefficient (R 2 ).

[0043] Furthermore, the hyperparameter tuning includes the following steps: generating hyperparameter combinations; parallel computing; model training and evaluation; and selecting the best model and parameters.

[0044] Preferably, parallel computing utilizes MATLAB's parfor loop to accelerate model training and evaluation.

[0045] Preferably, model training and evaluation include: for each parameter combination, training the model on each fold of cross validation and calculating the evaluation index; recording the average MSE of each parameter combination.

[0046] Preferably, the final model is trained on the full dataset using the best possible combination of model type and hyperparameters.

[0047] Furthermore, model evaluation and visualization were performed after the final model training, including manual cross-validation predictions and calculation of evaluation metrics;

[0048] The prediction results of each model are integrated and output: according to the R of each model on the validation set 2value, calculate their respective weights; perform weighted average on the prediction results of each model to get the final prediction result.

[0049] Preferably, visualization can be achieved by drawing a scatter plot and a reference line.

[0050] A damage time prediction model is constructed by the damage time prediction model construction method.

[0051] An application of an injury time prediction model is used for inferring injury time in forensic medicine; the injury time prediction model is constructed by the injury time prediction model construction method.

[0052] A method for predicting injury time comprises collecting feces and serum samples at different time points to obtain a data set of metabolite abundance values; and using the injury time prediction model to predict injury time.

[0053] The method is aimed at non-disease diagnosis.

[0054] Preferably, processing the sample data set includes the following steps: reading data; aligning features of the sample with a feature set of a damage time prediction model; and standardizing the sample data using the mean and standard deviation of the training data.

[0055] Preferably, the feces group model and the serum group model are used to predict the samples respectively to obtain a prediction result vector; according to the R 2 value, calculate their respective weights; perform weighted average on the prediction results of the model to obtain the final prediction result.

[0056] Preferably, the predicted time point of each new sample is listed; a bar graph is drawn to visualize the predicted time point of the sample.

[0057] The present invention utilizes the synergistic effect of multi-omics data and improves the accuracy of injury time prediction by integrating fecal and serum omics data. This emphasizes the importance of multi-omics data integration in biomedical research. Fecal metabolites mainly reflect the intestinal microbiota and local metabolic state, while serum metabolites provide systemic metabolic information. The combination of the two can more comprehensively reveal the complex metabolic changes of the body after injury and provide richer data for the discovery of biomarkers and mechanism research.

[0058] The optimized feature selection and model training process of the present invention weights and selects features through neighborhood component analysis (NCA), optimizes the feature selection process, reduces the complexity of the model, and improves the training efficiency. At the same time, the present invention adopts parallel computing and detailed hyperparameter tuning to ensure the optimal performance of the model on complex data sets.

[0059] The results of the model of the present invention corroborate with previous biological knowledge, demonstrating the reliability of the model. The key metabolites identified by the model, such as lipid metabolites, amino acids and their derivatives, organic acid metabolites, etc., are highly consistent with the metabolic pathways involved in the known injury repair process. The core role of lipid metabolism in cell membrane stability and repair, as well as the importance of amino acids in inflammatory response and antioxidant defense, are accurately captured by the model. In addition, the model reveals the dynamic changes of metabolites at different time points, which is consistent with the staged process of tissue repair after injury. Even metabolites that are initially regarded as non-significant are selected as important features by the model, suggesting the existence of synergistic effects in complex metabolic networks. These findings together demonstrate the reliability and effectiveness of the model in predicting and explaining biological processes.

[0060] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0061] 1. Improve prediction accuracy: The present invention constructs a high-precision injury time prediction model by integrating fecal and serum metabolite data. The support vector regression (SVR) model of the fecal group has extremely high prediction accuracy, and its determination coefficient (R 2 ) reached 0.9630, which indicates that the model can explain 96.30% of the data variation, which is significantly higher than the traditional method. The prediction accuracy of the random forest model for the serum group R 2 =0.7189, which provides valuable supplementary information for prediction.

[0062] 2. Identification of key biomarkers: The model of the present invention successfully identified key metabolites closely related to the injury and recovery process, such as lipid metabolites and amino acids and their derivatives, which play a key role in cell membrane stability, signal transduction and inflammatory response. This discovery helps to deepen the understanding of the biological mechanism after injury and may provide targets for the development of new diagnostic and therapeutic methods.

[0063] 3. Expanding forensic applications: The model of the present invention can not only be used for scientific research, but also can be applied to injury time inference in forensic medicine, providing an objective and scientific tool that helps to improve the accuracy and reliability of forensic identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The scatter plot is a comparison of the true value and predicted value of the feces group model;

[0065] Figure 2 The figure shows a scatter plot comparing the true value and predicted value of the serum group model. DETAILED DESCRIPTION

[0066] The present invention is further described below in conjunction with the accompanying drawings and specific examples, but the examples do not limit the present invention in any form. Unless otherwise specified, the reagents, methods and equipment used in the present invention are conventional reagents, methods and equipment in the art.

[0067] Unless otherwise specified, the reagents and materials used in the following examples are commercially available.

[0068] Example 1

[0069] 1. Sample collection and processing:

[0070] An injury model was established in mice, and samples were collected at the following time points: D0 (control group), D1 (1 day after injury), D3 (3 days after injury), D7 (7 days after injury), and D11 (11 days after injury).

[0071] The number of samples in the feces experimental group was 10, and 1971 metabolites were measured, including coenzyme Q (CoenzymeQ4), alkaloids (Isocorydine), bile acid (3b,7a,12a-Trihydroxy-5a-Chol anoic acid), aminocaprylic acid D (l-2-Aminooctanoic acid), 2,3-Dihydroxyisovaleric Acid (2,3-dihydroxyisovaleric acid), glutamine (Gln), tyrosine (Tyr), glutamic acid (Glu), gluconate (Gluconate), and lysophosphatidylcholine LPC (LPC O-18:2);

[0072] The number of samples in the serum experimental group was 6, and 1044 metabolites were measured, including uracil, genistein, DL-isoleucine, 4-vinylphenolsulfate, 1-naphthylamine, ferulic acid, b-glucose, and 3-hydroxyhexadecanoic acid.

[0073] All samples were preprocessed to obtain the abundance values ​​of metabolites.

[0074] 2. Model Construction

[0075] 2.1 Data Preprocessing

[0076] 2.1.1 Dataset organization

[0077] The fecal and serum data of 1 day, 3 days, 7 days, and 11 days were saved in Excel files respectively. Each data set contained the following contents: sample name (first column), feature data (second to second to last columns), label column (last column, indicating time point), statistical significance row (last row, marking the significance of each feature "yes" or "no")

[0078] 2.1.2 Data reading and integration

[0079] For each dataset, the following steps were performed:

[0080] 1. Read data: Use MATLAB's readcell function to read the Excel file.

[0081] 2. Extract feature names: Get the feature names except the first and last columns.

[0082] 3. Obtaining significance information: Read the significance mark in the last row and convert it into feature weight ("yes" corresponds to weight 1, "no" corresponds to weight 0.1). In the ANOVA test, the p value and q value are obtained, and the VIP is obtained through the PLS-DA model. When p<0.05, q<0.1, VIP>1, the feature is set as a significant feature, which is represented as "YES" in the significance matrix, otherwise it is represented as "NO", so as to classify the significance of the feature.

[0083] 4. Extract feature data: exclude the title row, first column, last column and significant rows to obtain pure feature data.

[0084] 5. Data type conversion: Convert cell data into numerical matrices and handle missing values ​​and non-numerical data.

[0085] 6. Feature alignment: Align the features of different data sets to build a global feature set.

[0086] 7. Feature weighting: Assign corresponding weights to each feature based on the significance information.

[0087] Through the above steps, the integrated feature data matrix X, label vector Y and feature weight matrix W are obtained.

[0088] 2.1.3 Data Standardization

[0089] In order to eliminate the influence of different feature dimensions, the data was standardized:

[0090] Calculate mean and standard deviation: Calculate the mean μ and standard deviation σ for each feature.

[0091] Handling constant value features: For features with a standard deviation of zero (constant value features), set the standard deviation to 1 to avoid division by zero errors.

[0092] Standardization formula:

[0093] Check for NaN values: Make sure there are no NaN values ​​in the normalized data.

[0094] 2.1.4 Feature Weighting

[0095] Multiply the standardized feature data by the feature weight matrix to obtain the weighted feature matrix:

[0096] X wcightcd =X standardizcd ×W.

[0097] 2.2 Feature Selection

[0098] In order to reduce the complexity of the model and improve the prediction performance, "Neighborhood Component Analysis (NCA)" is used for feature selection. The following steps are included:

[0099] 1. Use MATLAB's fsrnca function to perform NCA on the weighted feature data and labels.

[0100] 2. Get the weight of each feature.

[0101] 3. Set the threshold: Select features whose weight is greater than 1% of the maximum weight.

[0102] 4. Filter features: Select features based on the threshold and update the feature matrix and feature name list.

[0103] After NCA feature selection, the number of features is greatly reduced and the training efficiency of the model is improved.

[0104] 2.3 Model construction and hyperparameter tuning

[0105] 2.3.1 Model selection

[0106] In order to predict the time of mouse injury, three commonly used regression models were selected for comparison:

[0107] 1. Support Vector Regression (SVR),

[0108] 2. Random forest regression,

[0109] 3. Gradient Boosting.

[0110] 2.3.2 Hyperparameter Settings

[0111] For each model, the following different hyperparameter combinations were set:

[0112] 1. SVR hyperparameters:

[0113] Kernel function: linear, rbf, polynomial

[0114] Penalty parameter C(BoxConstraint): 0.01, 0.1, 1, 10

[0115] ε-insensitive loss function parameters (Epsilon): 0.01, 0.1, 0.5, 1

[0116] 2. Random Forest Hyperparameters:

[0117] Number of decision trees (NumTrees): 50, 100, 200

[0118] Minimum number of leaf node samples (MinLeafSize): 1, 5, 10

[0119] 3. Gradient Boosting Hyperparameters:

[0120] Number of learning cycles (NumLearningCycles): 50, 100, 200

[0121] Learning rate: 0.01, 0.1, 0.2

[0122] 2.3.3 Cross-validation and model evaluation

[0123] Cross-validation strategy: 10-fold cross-validation is used to ensure the stability of model evaluation. The following evaluation indicators are included:

[0124] Mean square error (MSE),

[0125] Root mean square error (RMSE),

[0126] Mean absolute error (MAE),

[0127] Coefficient of determination (R 2 ).

[0128] 2.3.4 Hyperparameter Tuning Process

[0129] 1. Parameter combination generation: Generate all possible hyperparameter combinations in advance to form a parameter space.

[0130] 2. Parallel computing: Use MATLAB's parfor loop to accelerate model training and evaluation.

[0131] 3. Model training and evaluation: For each parameter combination, train the model on each fold of cross validation and calculate the evaluation index. Record the average MSE for each parameter combination.

[0132] 4. Select the best model and parameters: Determine the best model type and hyperparameter combination based on the smallest average MSE.

[0133] 2.3.5 Final model training

[0134] Train the final model on the full dataset using the best combination of model type and hyperparameters.

[0135] 2.4 Model Evaluation and Visualization

[0136] Manual cross-validation prediction: In order to obtain more stable prediction results, we retrain the model for each fold and predict the validation set based on cross-validation.

[0137] Calculate evaluation indicators: Calculate MSE, RMSE, MAE and R based on the predicted values ​​and true labels 2 .

[0138] Result visualization: Draw a scatter plot of the true value and the predicted value, and add a reference line of the ideal prediction to intuitively display the prediction performance of the model.

[0139] 2.5 New Sample Prediction and Result Fusion

[0140] 2.5.1 New Sample Collection

[0141] A total of 5 new fecal and serum samples were collected on the 3rd day (F3-1, F3-2), 7th day (F7-1, F7-2) and 11th day (F11-1) of injury, and the abundance value data of metabolites were measured after pretreatment of the new samples. The new sample data were obtained according to the "2.1 Data Preprocessing" method.

[0142] 2.5.2 New Data Processing

[0143] 1. Read new sample data: Read the feature data of the new sample from Copy_of_StackingData.xlsx.

[0144] 2. Feature alignment: Align the features of the new sample with the feature set of the model, and fill the missing features with zeros.

[0145] 3. Standardization: Use the mean and standard deviation of the training data to standardize the new samples.

[0146] 2.5.3 Model prediction

[0147] The stool group model and serum group model are used to predict new samples respectively, and two prediction result vectors are obtained.

[0148] 2.5.4 Result Fusion

[0149] Weight determination: According to the R of the two models on the validation set 2 values ​​and calculate their respective weights.

[0150] Weighted average: The prediction results of the two models are weighted averaged to obtain the final prediction result.

[0151] 2.5.5 Results display

[0152] Output prediction results: List the predicted time points for each new sample.

[0153] Draw a bar chart: visualize the prediction time points of new samples and intuitively show the prediction effect of the model.

[0154] Results Analysis

[0155] 1. Model evaluation (I) Model 1: fecal group model evaluation

[0156] 1. Cross-validation results

[0157] When training the feces group model, 10-fold cross validation was used to evaluate the performance of the model. The best model was support vector regression (SVR), and its optimal hyperparameters were as follows:

[0158] Kernel Function linear Penalty parameter (BoxConstraint) 1 ε-Insensitive loss function parameter (Epsilon) 0.1

[0159] The evaluation indicators of the model in cross validation are:

[0160] Mean Square Error (MSE) 0.5451 Root mean square error (RMSE) 0.7383 Mean Absolute Error (MAE) 0.5640 <![CDATA[Coefficient of determination (R 2 )]]> 0.9630

[0161] High R 2 The value (close to 1) indicates that the model has a good fit to the data and is able to explain 96.30% of the variation in the label variable. This means that the stool group model has high accuracy and reliability in predicting time points.

[0162] 2. Visualization of prediction results

[0163] In order to intuitively demonstrate the prediction performance of the model, a scatter plot of the actual time point and the predicted time point is drawn ( Figure 1 ).like Figure 1 As shown, most of the data points are close to the ideal prediction line (45-degree diagonal line), which further verifies the accuracy of the model.

[0164] (II) Model 2: Serum group model evaluation

[0165] 1. Cross-validation results

[0166] When training the serum group model, 10-fold cross validation was used. The best model was a random forest regression with the following optimal hyperparameters:

[0167] Number of decision trees (NumTrees) 50 Minimum number of leaf node samples (MinLeafSize) 1

[0168] The evaluation indicators of the model in cross validation are:

[0169] Mean Square Error (MSE) 4.1467 Root mean square error (RMSE) 2.0363 Mean Absolute Error (MAE) 1.4800 <![CDATA[Coefficient of determination (R 2 )]]> 0.7189

[0170] Compared with the feces group model, the serum group model has a higher R 2 The value is low, indicating that it fits the data slightly less well, explaining 71.89% of the variation in the label variable. This may be due to the fact that there are fewer features in the serogroup data that are strongly correlated with time points, or that the data is noisy.

[0171] 2. Visualization of prediction results

[0172] Draw a scatter plot of the actual time points and predicted time points of the serum group model ( Figure 2 ). It can be observed that the distribution of data points is relatively scattered, and some points deviate far from the ideal prediction line, which is consistent with the low R 2 The values ​​are consistent.

[0173] (III) Feature Importance Analysis

[0174] Feature selection and feature importance assessment are key steps in the model building process. Neighborhood component analysis (NCA) was used to assess the contribution of features to the model prediction performance. The following are the top ten features in terms of feature importance scores in the feces group model:

[0175] 1.NEG_410.231_4.8247:1.1585

[0176] 2.NEG_231.139_5.8908:1.1144

[0177] 3.POS_496.338_5.9447:1.1134

[0178] 4.POS_276.119_0.7867:1.1018

[0179] 5.NEG_467.242_4.7236:1.1009

[0180] 6.POS_72.0813_0.9831:0.9952

[0181] 7.NEG_303.232_7.3796:0.9886

[0182] 8.NEG_714.501_7.7304:0.9825

[0183] 9.POS_230.247_5.3723:0.9235

[0184] 10.NEG_923.495_5.4594: 0.8926

[0185] Feature explanation: The prefixes "POS" and "NEG" in the feature name indicate the features detected in positive and negative ion modes, respectively. The subsequent numbers represent the mass-to-charge ratio (m / z) and retention time (RT), for example, "NEG_410.231_4.8247" indicates a feature with a mass-to-charge ratio of 410.231 and a retention time of 4.8247 minutes in negative ion mode.

[0186] The importance score of a feature reflects its impact on the model's predictive performance. The higher the score, the more important the feature is to the model. The following is an analysis of the top ten features:

[0187] NEG_410.231_4.8247 (score: 1.1585): This is the most important feature in the model, which may be related to a key metabolite and significantly affects the prediction of the time point;

[0188] NEG_231.139_5.8908 (score: 1.1144) and POS_496.338_5.9447 (score: 1.1134): followed closely, indicating that these two features have important contributions to the model in different ion modes;

[0189] POS_276.119_0.7867 (score: 1.1018) and NEG_467.242_4.7236 (score: 1.1009): The importance scores of these features are very close, and they may be involved in a biological process together;

[0190] POS_72.0813_0.9831 (score: 0.9952) to NEG_923.495_5.4594 (score: 0.8926): These features have slightly lower scores but still contribute significantly to the model.

[0191] 2. Biological analysis using the results obtained from the model

[0192] (I) Classification and function of metabolites:

[0193] 1. Lipid metabolites (such as "NEG_410.231_4.8247" and "NEG_467.242_4.7236"): These metabolites mainly belong to glycerophospholipids. They may be related to the stability and repair of cell membranes and play an important role in the repair process after injury.

[0194] 2. Amino acids and derivatives (such as "POS_276.119_0.7867"): During the inflammatory response and immune regulation after injury, amino acid metabolism may be accelerated to provide energy and raw materials for cell repair.

[0195] 3. Organic acid metabolites (e.g., “NEG_231.139_5.8908”): These compounds may be related to energy metabolism and antioxidant responses, reflecting metabolic stress after injury.

[0196] 2. Biological significance of non-significant features

[0197] Through NCA feature selection, some metabolites marked as non-significant (lower weight) are still selected, indicating that although these metabolites seem to have little effect at the beginning, they may play an important role in more complex metabolic networks. This phenomenon explains:

[0198] 1. Indirect contributions and synergistic effects: In biological systems, some metabolites may play important roles in specific biological processes through indirect pathways or by cooperating with other molecules. Although the changes of these metabolites are not significant individually, they may play an important role in the regulation of the entire metabolic network. The discovery of such "hidden" features can guide new research directions to further explore the potential functions of these metabolites in cell signaling, membrane remodeling or immune response.

[0199] 2. Complex metabolic network reactions: The selection of non-significant metabolites may also suggest that during the damage repair process, multiple metabolic pathways work in parallel and regulate each other, forming a complex feedback mechanism. For example, lipid metabolites (such as "NEG_410.231_4.8247") may play an indirect but critical role in the cell membrane repair process by regulating membrane fluidity or signal molecule transport.

[0200] 3. Time-dependent metabolite response

[0201] Samples were collected at different time points (D1, D3, D7, D11), involving different stages of acute stress, inflammatory response and tissue repair. From the feature selection of metabolites, the change pattern of metabolites at different time points can be found, revealing the time-dependent metabolite response after injury:

[0202] Acute stage (D1, D3): In the acute stage after injury, changes in energy metabolites and lipid metabolites may be to cope with the high metabolic demand after injury and to maintain the integrity of cell membranes and tissues. For example, LPE 13:0 (present in "NEG_410.231_4.8247") is related to the reconstruction of cell membranes, which may be particularly important in the early stages.

[0203] Recovery phase (D7, D11): In the late repair phase, amino acid metabolism and lipid oxidation metabolism may dominate, supporting cell proliferation, tissue regeneration and anti-inflammatory processes. For example, "NEG_467.242_4.7236" may be related to the glycerophospholipid pathway, which is an important component in the cell membrane repair process.

[0204] Such metabolite changes over a time series provide a more detailed perspective for understanding the biological processes after injury. The dynamic response of metabolites reflects the different needs of cells and tissues at different stages. These changes can be used as early disease markers or indicators for monitoring therapeutic effects in the future.

[0205] 4. The core role of lipid metabolism

[0206] In the results, multiple lipid metabolites were selected as important features, indicating the central role of lipid metabolism in damage response and repair. For example, multiple glycerophospholipid metabolites (such as LysoPC and LPG) are closely related to membrane remodeling and signal transduction:

[0207] LysoPC (lysophosphatidylcholine): These metabolites are usually released in tissue damage and inflammatory responses. They are not only degradation products of membrane phospholipids, but also may act as intercellular signaling molecules to regulate the chemotaxis of immune cells. Therefore, the high weight of LysoPC metabolites may suggest their potential new mechanism for regulating inflammatory responses after injury.

[0208] Glycerophospholipids and cell signaling: Changes in glycerophospholipid metabolism reflect the repair of cell membranes and the regulation of signal transduction. The frequent appearance of this type of metabolites suggests that perturbations in lipid metabolism may be the key to rebalancing cellular metabolism after injury.

[0209] (V) The relationship between amino acids and antioxidant defense

[0210] Features such as POS_276.119_0.7867 may be related to amino acid metabolism (such as glutamine and glutamate). Such metabolites play an important role in responding to oxidative stress and promoting tissue repair:

[0211] Antioxidant defense: Glutamate is a precursor of the antioxidant glutathione and participates in the cell's defense mechanism against oxidative stress. Oxidative stress after injury is a hallmark of cell damage, and activation of amino acid metabolism may help mitigate oxidative damage.

[0212] The role of amino acids in repair: Amino acids not only serve as the basic units of protein synthesis, but also play a role in signal transduction and immune regulation. Changes in amino acid metabolites may provide new targets for the treatment of tissue damage in the future.

[0213] (VI) New Sample Prediction

[0214] The data of the five new samples are stored in a file named Copy_of_StackingData.xlsx and the code is run for prediction, and the following results are obtained:

[0215] Sample name Model 1 predicts time points Model 2 predicts time points Final prediction time point F3-1 6.38 4.00 5.37 F3-2 3.66 2.80 3.29 F7-1 4.90 8.28 6.35 F7-2 5.17 7.88 6.33 F11-1 11.76 8.44 10.34

[0216] By combining the prediction results of Model 1 and Model 2, a more accurate time point prediction was obtained. It is worth noting that the random forest model of the serogroup (Model 2) provided important complementary information on some samples. For example, for samples F7-1 and F7-2, the predicted time points of Model 2 were closer to the actual situation, which indicates that the serogroup data have a stronger correlation at these time points.

[0217] Therefore, the comprehensive use of the data from the stool group and serum group through the weighted average method can improve the predictive performance of the model and obtain more reliable prediction results.

[0218] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for constructing a damage time prediction model, characterized in that: The following steps are involved: S1. Preprocessing the data sets of feces and serum metabolite abundance values; S2. Construct at least two models among support vector regression model, random forest regression model and gradient boosting regression model; S3, hyperparameter tuning and model training; S4. The prediction results of each model are integrated and output.

2. The method for constructing the damage time prediction model according to claim 1, characterized in that: The feces and serum were collected at different time points after injury.

3. The method for constructing the damage time prediction model according to claim 1, characterized in that: The metabolites include lipid metabolites, amino acids and their derivatives, organic acid metabolites, nucleic acids and their metabolites, phenolic and polyphenolic compounds, and sugars.

4. The method for constructing the damage time prediction model according to claim 1, characterized in that: During data preprocessing, feature selection is performed through feature weighting and neighborhood component analysis.

5. The method for constructing the damage time prediction model according to claim 4, characterized in that: The domain component analysis includes the following steps: performing neighborhood component analysis on weighted feature data and labels; obtaining the weight of each feature; selecting features whose weights are greater than 1% of the maximum weight as thresholds; and screening features according to the thresholds.

6. The method for constructing the damage time prediction model according to claim 1, characterized in that: The hyperparameter tuning includes the following steps: generating hyperparameter combinations; parallel computing; model training and evaluation; and selecting the best model and parameters.

7. The method for constructing the damage time prediction model according to claim 1, characterized in that: Perform model evaluation and visualization after model training, including manual cross-validation predictions and calculation of evaluation metrics; The prediction results of each model are integrated and output: according to the R of each model on the validation set 2 values, calculate their respective weights; The prediction results of each model are weighted averaged to obtain the final prediction result.

8. A damage time prediction model, characterized in that: The damage time prediction model is constructed by the damage time prediction model construction method described in any one of claims 1 to 7.

9. An application of the damage time prediction model according to claim 8, characterized in that: Used for inferring the time of injury in forensic medicine.

10. A method for predicting damage time, characterized in that: Feces and serum samples are collected at different time points, and a data set of metabolite abundance values ​​is obtained by measurement; and the injury time prediction model described in claim 8 is used to predict the injury time.

Images