Cardiovascular age prediction method and device based on deep learning and metabolic profile analysis

Abstract

The present invention discloses a cardiovascular age prediction method and device based on deep learning and metabolic spectrum analysis, which relates to the field of biological data processing, including: obtaining blood samples of healthy people and people with cardiovascular diseases and performing data processing and feature selection to determine relevant metabolites; constructing a cardiovascular age prediction model based on deep learning and training it to obtain a trained cardiovascular age prediction model; obtaining blood samples of people to be tested and processing it to obtain a metabolomics feature matrix composed of metabolic features corresponding to relevant metabolites, inputting the metabolomics feature matrix into the trained cardiovascular age prediction model, passing through a metabolic feature representation module to obtain a metabolic feature representation, inputting the metabolic feature representation into a feature representation processing module, predicting the cardiovascular age of the person to be tested, and calculating the cardiovascular age acceleration. The present invention solves the problems of high cost, low accuracy, lack of extensive verification, etc. of existing cardiovascular age prediction.

Description

Technical Field

[0001] The present invention relates to the field of biological data processing, and in particular to a cardiovascular age prediction method and device based on deep learning and metabolic spectrum analysis. Background Art

[0002] As the global population ages, the burden of cardiovascular disease has increased significantly. This trend has prompted researchers to explore the complex and close biological connection between aging and cardiovascular health risks in more depth. Aging is usually defined as a process in which functional ability and resilience decline over time. It is not only the basis for the development of most chronic diseases, but is also considered a major risk factor for a variety of fatal diseases such as heart disease, neurodegenerative diseases and cancer. However, the decline trajectory not only varies significantly between individuals, but also shows a high degree of heterogeneity between different organ systems. Especially in the cardiovascular system, this heterogeneity reflects the significant differences in the aging rate of individuals under physiological and pathological conditions. Therefore, it is particularly important to find indicators that can replace chronological age and more accurately reflect the physiological state of the cardiovascular system in order to predict cardiovascular health risks.

[0003] Biological age has been widely regarded as a more accurate indicator for disease risk assessment and mortality prediction than actual age. Biological age (BA) measures the physiological state of an individual through a series of biomarkers. The results may deviate significantly from the individual's actual age, and an increase in biological age is often closely associated with a higher risk of disease. In recent years, researchers have proposed biological aging markers covering multiple fields such as epigenetics, proteomics, and glycomics, such as the "epigenetic clock" measured by the DNA methylation state of CpG sites. This method can convert DNA methylation data into a predicted age that is highly correlated with actual age. In addition, leukocyte telomere length, as a marker of aging that shortens due to an increase in the number of cell divisions, as well as N-glycan profiles and Ink4a / Arf locus expression that reflect changes in glycosylation patterns, together reveal differences in the rate of aging between individuals and reflect the complexity of the aging process.

[0004] At the same time, with the explosive growth of biomedical data, the availability of large-scale omics data has also brought new challenges to the analysis and interpretation of molecular aging. Traditional statistical methods have some drawbacks in processing and interpreting these complex data. Machine learning (ML) and deep learning (DL), as the core technologies of modern artificial intelligence, are becoming increasingly key tools in biomedical research due to their powerful capabilities in large-scale data analysis. Machine learning algorithms can handle complex nonlinear relationships by identifying patterns in data, and are suitable for analyzing high-dimensional data in biology, especially in the application of omics data, helping to screen and identify important features related to disease or health status, and establish predictive models. Deep learning is a subfield of machine learning, which has achieved remarkable results in image recognition, natural language processing and other fields, and its application potential in the biomedical field is gradually emerging. Researchers are increasingly using machine learning and deep learning techniques to refine omics data into composite aging biomarkers to help explain the complex biology of aging and guide clinical decision-making. Metabolomics, as an important part of systems biology, measures the concentration of the final products of cellular metabolism with relatively small molecular weights, and is a promising technology for comprehensive assessment of biological aging. Compared with other omics methods, metabolomics can provide a more complete picture of biological processes and stronger phenotypic representation, showing significant advantages in predicting biological age. Through non-targeted metabolomics analysis, researchers can capture the information of all metabolites in the sample to the greatest extent. Especially in cardiovascular health research, through chromatography-mass spectrometry (LC-MS), metabolomics can efficiently detect and analyze metabolites and provide key information about the state of the cardiovascular system. However, there is currently no research on predicting and constructing the age of the cardiovascular system through metabolomics.

[0005] There are the following defects in the current prior art:

[0006] 1. Lack of metabolomics-based cardiovascular age prediction research: There is currently no research using metabolomics to construct and predict the biological age of the cardiovascular system, indicating that research in this field is still in its infancy and the methodology is insufficient.

[0007] 2. Low prediction accuracy: Traditional actual age cannot accurately reflect the biological aging process of an individual's cardiovascular system, resulting in low accuracy in predicting brain health risks. The age prediction accuracy of existing omics technologies (such as transcriptomics clocks) is highly variable in different cohorts and significantly lower than that of epigenetic clocks. This inaccuracy may be partly attributed to the increase in technical noise during data integration, especially when microarray and sequencing data come from multiple platforms.

[0008] 3. The model is based on a small cohort and lacks extensive validation: Currently, most biological age prediction models are based on small sample sizes and lack validation in large-scale independent cohorts. Therefore, the accuracy and repeatability of the model in large populations is unclear.

[0009] 4. High cost limits application: Many omics methods (such as proteomics and methylomics) are expensive, which limits their application in large-scale biobanks and population health. In particular, some high-throughput omics technologies, such as epigenomics and transcriptomics, although they can provide rich information, their high cost and technical barriers hinder their widespread application in the field of brain health.

[0010] 5. The prediction model is too simple: Most current biological age prediction models (including statistical and simple machine learning methods) perform poorly when dealing with high-dimensional, complex omics data. Simple linear regression models cannot capture nonlinear relationships and interaction effects in the data, are sensitive to collinearity, and lead to model instability, increased prediction errors, and lack the ability to handle complex data.

[0011] The above defects indicate that although cardiovascular biological age prediction has great potential in personalized health management and disease intervention, the limitations of existing technologies and methods hinder its development in clinical and large-scale applications. Summary of the invention

[0012] The purpose of this application is to propose a cardiovascular age prediction method and device based on deep learning and metabolic spectrum analysis to address the above-mentioned technical problems.

[0013] In a first aspect, the present invention provides a cardiovascular age prediction method based on deep learning and metabolic profile analysis, comprising the following steps:

[0014] Obtaining blood samples from healthy people and people with cardiovascular diseases and performing metabolite extraction and chromatography-mass spectrometry analysis to obtain mass spectrometry data of each metabolite in several metabolomes corresponding to the healthy people and the people with cardiovascular diseases, performing data processing on the mass spectrometry data of each metabolite in each metabolome to obtain several metabolic features, wherein the metabolic features are formed by splicing the chromatographic retention time and the mass-to-charge ratio in the mass spectrometry data of the metabolites, performing feature selection on several metabolic features corresponding to the healthy people and the people with cardiovascular diseases to determine related metabolites, and constructing first training data using the metabolic features corresponding to the related metabolites of the healthy people;

[0015] A cardiovascular age prediction model based on deep learning is constructed and trained using the first training data to obtain a trained cardiovascular age prediction model, wherein the trained cardiovascular age prediction model includes a metabolic feature representation module and a feature representation processing module, wherein the metabolic feature representation module includes an embedding layer, a multi-head self-attention layer, a feedforward fully connected layer, and a layer normalization layer, wherein the embedding layer, the multi-head self-attention layer, and the feedforward fully connected layer are sequentially connected, wherein the output features of the multi-head self-attention layer and the output features of the feedforward fully connected layer are first residually connected and then input into the layer normalization layer, and the feature representation processing module uses a multilayer perceptron;

[0016] A blood sample of the person to be tested is obtained and processed to obtain a metabolomics feature matrix composed of metabolic characteristics corresponding to relevant metabolites. The metabolomics feature matrix is ​​input into a trained cardiovascular age prediction model, and passed through a metabolic feature representation module to obtain a metabolic feature representation. The metabolic feature representation is input into a feature representation processing module to predict the cardiovascular age of the person to be tested, and the cardiovascular age acceleration is calculated.

[0017] Preferably, the method used for feature selection includes a differential analysis feature screening method and a Boruta method feature screening method, and the metabolite set obtained by the differential analysis feature screening method and the metabolite set obtained by the Boruta method feature screening method are intersected to obtain related metabolites, and the metabolic characteristics corresponding to the related metabolites are metabolic characteristics related to the age of cardiovascular disease.

[0018] Preferably, the differential analysis feature screening method specifically includes:

[0019] A two-sample t-test was performed on the metabolic characteristics of the healthy population and the metabolic characteristics of the population with cardiovascular diseases to obtain a p-value, and the p-value was corrected using the BH method to obtain a corrected p-value. The metabolites corresponding to the metabolic characteristics that satisfy the corrected p-value less than 0.05 and the difference multiple greater than 2 or less than 1 / 2 constitute the metabolic set obtained by the difference analysis feature screening method. The difference multiple is the multiple ratio of the median of the metabolic characteristics of the population with cardiovascular diseases to the median of the metabolic characteristics of the healthy population for each metabolite;

[0020] The Boruta method feature screening methods include:

[0021] Constructing a random forest model, the random forest model includes multiple decision trees, taking the metabolic characteristics of healthy people and the metabolic characteristics of people with cardiovascular diseases as original characteristics, randomly shuffling the values ​​in the original characteristics to generate shadow characteristics, constructing second training data based on the original characteristics and the shadow characteristics, using the second training data to train the random forest model, inputting the metabolic characteristics in the second training data into the random forest model, and obtaining a classification result, the classification result is a healthy population or a population with cardiovascular diseases, for each original characteristic and each shadow characteristic, calculating and accumulating the reduction of the Gini index caused by each original characteristic and each shadow characteristic on each node in each decision tree during the training process, and obtaining the total reduction of the Gini index of each decision tree, averaging the total reduction of the Gini index of all decision trees, calculating the average reduction of the Gini index of each original characteristic and the average reduction of the Gini index of each shadow characteristic, and determining the maximum average reduction of the Gini index among all shadow characteristics according to the average reduction of the Gini index of each shadow characteristic;

[0022] In response to when the average reduction of the Gini index of one of the original features is greater than the maximum average reduction of the Gini index of all the shadow features, marking the one of the original features as an important feature;

[0023] In response to when the average reduction of the Gini index of one of the original features is less than or equal to the maximum average reduction of the Gini index of all the shadow features, marking one of the original features as an unimportant feature;

[0024] The metabolites marked as corresponding to important features are used as the metabolite set obtained by the Boruta feature screening method.

[0025] Preferably, the metabolomics feature matrix is ​​input into the metabolic feature representation module, and each metabolic feature in the metabolomics feature matrix with a dimension of n×m is first mapped into a 256-dimensional vector through an embedding layer to obtain an embedded feature representation matrix with a dimension of n×256, wherein n is the total number of samples and m is the number of metabolic features corresponding to the relevant metabolites; the multi-head self-attention layer contains 6 self-attention heads, and the n×256 embedded feature representation matrix is ​​mapped into 6 groups of different query matrices, key matrices and value matrices, and each group of query matrices, key matrices and value matrices passes through each self-attention head. Output each output feature matrix, concatenate the six output feature matrices to obtain a concatenated matrix, input the concatenated matrix into the linear transformation layer, and obtain the output features of the multi-head self-attention layer; the feedforward fully connected layer includes a first fully connected layer and a second fully connected layer connected in sequence, the first fully connected layer uses a ReLU activation function, the output features of the multi-head self-attention layer are input into the feedforward fully connected layer, and pass through the first fully connected layer and the second fully connected layer in sequence to obtain the output features of the feedforward fully connected layer, the output features of the multi-head self-attention layer are added to the output features of the feedforward fully connected layer, and then input into the normalization layer to obtain the metabolic feature representation.

[0026] Preferably, the multilayer perceptron includes an input layer, a first hidden layer, a second hidden layer, a batch normalization layer, a Dropout layer and an output layer connected in sequence; the input layer is used to receive metabolic feature representation, the first hidden layer contains 256 neurons, and uses a ReLU activation function; the second hidden layer contains 128 neurons, and uses a ReLU activation function; the dropout ratio in the Dropout layer is 0.5, the output layer contains 1 neuron, and uses a linear activation function to output cardiovascular age.

[0027] Preferably, the total loss function used in the training process of the deep learning-based cardiovascular age prediction model is for:

[0028] ;

[0029] Where n is the total number of samples, , is the mean square error loss function of the i-th sample, is the weighted loss function of the i-th sample, α and β are the weight coefficients of the mean square error loss function and the weighted loss function respectively, and the weighted loss function is expressed as:

[0030] ;

[0031] in, represents the weight factor, , γ represents a hyperparameter, represents the actual age of the i-th sample, Represents the cardiovascular age predicted by the cardiovascular age prediction model based on deep learning for the i-th sample.

[0032] In a second aspect, the present invention provides a cardiovascular age prediction device based on deep learning and metabolic spectrum analysis, comprising:

[0033] a feature selection module configured to obtain blood samples from healthy people and people with cardiovascular diseases and perform metabolite extraction and chromatography-mass spectrometry analysis to obtain mass spectrometry data of each metabolite in several metabolomes corresponding to the healthy people and the people with cardiovascular diseases, perform data processing on the mass spectrometry data of each metabolite in each metabolome to obtain several metabolic features, wherein the metabolic features are formed by concatenating the chromatographic retention time and the mass-to-charge ratio in the mass spectrometry data of the metabolites, perform feature selection on several metabolic features corresponding to the healthy people and the people with cardiovascular diseases, determine relevant metabolites, and construct first training data using the metabolic features corresponding to the relevant metabolites of the healthy people;

[0034] a model building module, configured to build a cardiovascular age prediction model based on deep learning and use the first training data for training to obtain a trained cardiovascular age prediction model, wherein the trained cardiovascular age prediction model includes a metabolic feature representation module and a feature representation processing module, wherein the metabolic feature representation module includes an embedding layer, a multi-head self-attention layer, a feedforward fully connected layer, and a layer normalization layer, wherein the embedding layer, the multi-head self-attention layer, and the feedforward fully connected layer are connected in sequence, wherein the output features of the multi-head self-attention layer and the output features of the feedforward fully connected layer are first residually connected and then input into the layer normalization layer, and the feature representation processing module uses a multi-layer perceptron;

[0035] The prediction module is configured to obtain a blood sample from a person to be tested and process it to obtain a metabolomics feature matrix composed of metabolic features corresponding to relevant metabolites, input the metabolomics feature matrix into a trained cardiovascular age prediction model, pass it through a metabolic feature representation module, obtain a metabolic feature representation, input the metabolic feature representation into a feature representation processing module, predict the cardiovascular age of the person to be tested, and calculate the cardiovascular age acceleration.

[0036] In a third aspect, the present invention provides an electronic device comprising one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner in the first aspect.

[0037] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any implementation manner in the first aspect.

[0038] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in any implementation manner in the first aspect.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The cardiovascular age prediction method based on deep learning and metabolic profile analysis proposed in this invention combines metabolomics data with deep learning models, providing a new direction for the prediction of the biological age of the cardiovascular system. This combination enables the model to utilize the biomarkers implicit in metabolomics to more accurately reflect the cardiovascular biological state of an individual. Metabolomics captures the changes that occur during cellular metabolism, which are closely related to the aging process of the cardiovascular system. Compared with traditional clinical indicators, metabolomics data can better reflect the true physiological state and degree of aging of an individual. It can provide multi-dimensional biological information.

[0041] (2) The cardiovascular age prediction model in the cardiovascular age prediction method based on deep learning and metabolic profile analysis proposed in the present invention can efficiently capture the complex interactions and nonlinear relationships between metabolites by using the encoder structure of the improved Transformer model in the metabolic feature representation module. The multi-head self-attention mechanism in the encoder structure of the improved Transformer model enables it to process multi-dimensional features in parallel and capture richer feature relationships, surpassing the capabilities of traditional linear regression and simple machine learning models. By using the embedding layer and the multi-head self-attention mechanism, high-dimensional and sparse metabolomics data can be converted into low-dimensional and dense metabolic feature representations, reducing data noise and improving the model's feature expression capabilities and prediction accuracy.

[0042] (3) The cardiovascular age prediction method based on deep learning and metabolic spectrum analysis proposed in the present invention can accurately predict cardiovascular age, identify early signs of aging in individuals, provide a basis for personalized health management and disease prevention, and help formulate personalized intervention measures. By analyzing the metabolomics features extracted by the model, it is possible to further explore metabolites closely related to cardiovascular aging, providing important biomarkers for future clinical research and intervention strategies. The detection cost of metabolomics is relatively low, and data collection and processing are simpler, especially in large-scale biological sample libraries and population health research. It has high economic feasibility, which enables the present invention to be widely used in clinical, public health and personalized health management fields, and provide support for large-scale cardiovascular health monitoring and intervention strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 A schematic diagram of a process of a cardiovascular age prediction method based on deep learning and metabolic profile analysis according to an embodiment of the present application;

[0045] Figure 2 A schematic diagram of a cardiovascular age prediction model of a cardiovascular age prediction method based on deep learning and metabolic profile analysis according to an embodiment of the present application;

[0046] Figure 3 A relationship diagram between the cardiovascular age predicted by the cardiovascular age prediction method based on deep learning and metabolic spectrum analysis and the actual age in an embodiment of the present application;

[0047] Figure 4 A schematic diagram of the distribution of cardiovascular age acceleration obtained by the cardiovascular age prediction method based on deep learning and metabolic spectrum analysis in an embodiment of the present application;

[0048] Figure 5 A schematic diagram of a cardiovascular age prediction device based on deep learning and metabolic profile analysis according to an embodiment of the present application;

[0049] Figure 6 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Figure 1 A cardiovascular age prediction method based on deep learning and metabolic profile analysis provided in an embodiment of the present application is shown, comprising the following steps:

[0052] S1, obtaining blood samples from healthy people and people with cardiovascular diseases and performing metabolite extraction and chromatography-mass spectrometry analysis to obtain mass spectrometry data of each metabolite in several metabolomes corresponding to the healthy people and the people with cardiovascular diseases, performing data processing on the mass spectrometry data of each metabolite in each metabolome to obtain several metabolic features, wherein the metabolic features are spliced ​​by the chromatographic retention time and mass-to-charge ratio in the mass spectrometry data of the metabolites, performing feature selection on several metabolic features corresponding to the healthy people and the people with cardiovascular diseases to determine related metabolites, and constructing the first training data using the metabolic features corresponding to the related metabolites of the healthy people.

[0053] In a specific embodiment, the feature selection method includes a differential analysis feature screening method and a Boruta method feature screening method. The metabolite set obtained by the differential analysis feature screening method and the metabolite set obtained by the Boruta method feature screening method are intersected to obtain related metabolites. The metabolic characteristics corresponding to the related metabolites are metabolic characteristics related to the age of cardiovascular disease.

[0054] In a specific embodiment, the difference analysis feature screening method specifically includes:

[0055] A two-sample t-test was performed on the metabolic characteristics of the healthy population and the metabolic characteristics of the population with cardiovascular diseases to obtain a p-value, and the p-value was corrected using the BH method to obtain a corrected p-value. The metabolites corresponding to the metabolic characteristics that satisfy the corrected p-value less than 0.05 and the difference multiple greater than 2 or less than 1 / 2 constitute the metabolic set obtained by the difference analysis feature screening method. The difference multiple is the multiple ratio of the median of the metabolic characteristics of the population with cardiovascular diseases to the median of the metabolic characteristics of the healthy population for each metabolite;

[0056] The Boruta method feature screening methods include:

[0057] Constructing a random forest model, the random forest model includes multiple decision trees, taking the metabolic characteristics of healthy people and the metabolic characteristics of people with cardiovascular diseases as original characteristics, randomly shuffling the values ​​in the original characteristics to generate shadow characteristics, constructing second training data based on the original characteristics and the shadow characteristics, using the second training data to train the random forest model, inputting the metabolic characteristics in the second training data into the random forest model, and obtaining a classification result, the classification result is a healthy population or a population with cardiovascular diseases, for each original characteristic and each shadow characteristic, calculating and accumulating the reduction of the Gini index caused by each original characteristic and each shadow characteristic on each node in each decision tree during the training process, and obtaining the total reduction of the Gini index of each decision tree, averaging the total reduction of the Gini index of all decision trees, calculating the average reduction of the Gini index of each original characteristic and the average reduction of the Gini index of each shadow characteristic, and determining the maximum average reduction of the Gini index among all shadow characteristics according to the average reduction of the Gini index of each shadow characteristic;

[0058] In response to when the average reduction of the Gini index of one of the original features is greater than the maximum average reduction of the Gini index of all the shadow features, marking the one of the original features as an important feature;

[0059] In response to when the average reduction of the Gini index of one of the original features is less than or equal to the maximum average reduction of the Gini index of all the shadow features, marking one of the original features as an unimportant feature;

[0060] The metabolites marked as corresponding to important features are used as the metabolite set obtained by the Boruta feature screening method.

[0061] Specifically, the experimental analysis is carried out first, and the experimental process includes sample collection, metabolite extraction, mass spectrometry detection and data analysis. First, sample collection and metabolite extraction are carried out. Blood samples are collected from healthy people without obvious major diseases and patients diagnosed with cardiovascular diseases (cardiovascular diseases (I00-I99), congestive heart failure (I50.0, I50.1, I50.9), coronary artery disease (I20-I25.9), myocardial infarction (I21-I23) and angina pectoris (I20.0-I20.9), etc.) whose probability of disease increases with age. The blood samples are thawed and fully mixed in a low-temperature environment, and then pre-cooled methanol is added for treatment. Metabolites are extracted by ultrasonic treatment, and then impurities are precipitated at ultra-low temperatures, and the supernatant is separated by high-speed centrifugation. The supernatant is evaporated in a high-speed vacuum concentrator centrifuge for use in mass spectrometry detection. In the mass spectrometry detection stage, the dried samples are re-dissolved with a pre-cooled methanol-water solution and centrifuged again to ensure that the samples are pure and uniform. Subsequently, the samples were introduced into an ultra-high performance liquid chromatography system (UHPLC) for separation, using high-efficiency chromatographic columns and precise gradient elution programs to achieve optimal separation and detection of metabolites. The entire analysis process was carried out under strictly controlled temperature and flow rate conditions to ensure the accuracy and repeatability of the data. During the data analysis phase, the raw mass spectrometry data were processed using MSDIAL software, including peak alignment, retention time correction, and peak area extraction, to obtain the mass spectrometry data of each metabolite in several metabolomes corresponding to healthy people and people with cardiovascular disease. The structural identification of metabolites was performed by accurate mass matching and secondary spectrum matching, and public databases such as HMDB, MassBank, and GNPS were searched to determine the names of the metabolites.

[0062] Furthermore, based on the mass spectrometry data of each metabolite in several metabolomes corresponding to healthy people and people with cardiovascular diseases, the mass spectrometry data is processed. The data processing method includes data cleaning and normalization, aiming to ensure the quality and consistency of the data for subsequent analysis and interpretation. The missing values ​​in the mass spectrometry data are processed by using the KNN interpolation method. If there are too many missing values, the metabolite or sample is eliminated to avoid bias in the analysis results. The data is logarithmically transformed and each value is unit-added to allow zero value conversion, and then centered on the mean and scaled to unit variance to reduce the skewed distribution of the data. The mean of each feature is made to be 0 and the standard deviation is 1.

[0063] Next, we enter the feature selection stage. In order to find a model that provides the best cardiovascular system age prediction performance, the embodiment of this application combines three metabolomics data sets, including indicators: MS HPOS in serum (hydrophilic interaction chromatography HILIC can enhance the separation of highly polar small molecules ionized in positive mode), MS LNEG in serum (lipid-targeted reversed-phase chromatography can provide the maximum resolution of fatty acids, triglycerides and phospholipids ionized in negative mode), and MSLPOS in serum (lipid-targeted reversed-phase chromatography, ionized in positive mode). The offline data is peak aligned, retention time corrected, and peak area extracted. After data cleaning and normalization, a total of 14,549 metabolic features are obtained for feature selection of model construction. Feature selection methods include differential analysis feature screening methods and Boruta method feature screening methods.

[0064] 1) Differential analysis feature screening: Through statistical analysis, the difference in metabolite concentrations between healthy people and people with cardiovascular diseases is compared. The specific method includes performing a two-sample t-test on each metabolic feature to compare the concentration difference between the two groups of metabolites, and adjusting the p-value by the Benjamini-Hochberg (BH) method correction to control the false positive rate. Set the discrimination criteria: the corrected p-value < 0.05, and the fold change (difference multiple) > 2 or fold change (difference multiple) < 1 / 2. Metabolic features that meet this condition are considered to be metabolic features that change significantly with age, with a total of 941 metabolites. Through this detailed differential analysis process, the embodiments of the present application can identify metabolic features that change significantly with age, providing important biomarkers for subsequent cardiovascular age prediction models. This method not only improves the accuracy of metabolomics data analysis, but also provides new insights into understanding metabolic changes during cardiovascular aging.

[0065] 2) Boruta method feature screening: By grouping the metabolic features of healthy people and the metabolic features of people with cardiovascular diseases, the healthy people are marked as 0 and the people with cardiovascular diseases are marked as 1, and the second training data is constructed. In order to perform feature selection by the Boruta method, a random forest model is first constructed, and a random forest model is trained using the second training data so that the random forest model can distinguish between healthy people and people with cardiovascular diseases, and the contribution of each metabolic feature to the reduction of impurity of the random forest model is calculated during the training process. The metabolic features of healthy people and the metabolic features of people with cardiovascular diseases are used as original features. The Boruta method evaluates the importance of features by introducing shadow features (randomly shuffled original features). In the original features, each original feature (metabolic feature) is copied, and the values ​​of the original features are randomly shuffled (i.e., randomly permuted) during the copying process. This copying process creates a corresponding shadow feature for each original feature. For example, a RT_m / z feature corresponds to a new RT_m / z_shadow feature to destroy any potential association between the shadow feature and the target variable. The shuffling process usually involves randomly rearranging the data within each shadow feature column. This shuffling ensures that the shadow features have no real association with the target variable and only contain noise. Shadow features are used as a benchmark to help distinguish which original features contribute significantly to the model. Specifically, the Boruta method calculates the importance of metabolic features based on the random forest model and examines the contribution of each feature (including original features and shadow features) in the split node of the decision tree to the reduction of model impurity.

[0066] In order to perform feature selection using the Boruta method, we first need to calculate the contribution of each feature (including original features and shadow features) to the reduction of model impurity during the training process of the random forest model. Specifically, the Boruta method analyzes all nodes of all trees and calculates the reduction of the Gini index of each feature in the split node in the decision tree of the random forest model. The definition of the Gini index Gini is as follows:

[0067] ;

[0068] in, represents the number of categories, Indicates that the node belongs to the category The sample proportion of

[0069] The reduction in the Gini index (i.e., the reduction in impurity) is calculated before and after the node splitting of the feature, as shown in the following formula:

[0070] ;

[0071] in, represents the reduction in the Gini index, represents the Gini index before the node splits, Represents the Gini index after node splitting;

[0072] The Gini index reduction is accumulated in the entire decision tree and then averaged in the entire random forest model to obtain the average reduction in the Gini index of the feature. The Boruta method compares the average reduction in the Gini index of each original feature with the average reduction in the maximum Gini index of the shadow feature through multiple iterations. If the Gini index reduction of an original feature is greater than the average reduction in the maximum Gini index of the shadow feature, the original feature is marked as an important feature. If the Gini index reduction of an original feature is less than or equal to the average reduction in the maximum Gini index of the shadow feature, the original feature is marked as an unimportant feature. The original features marked as important features are summarized and sorted from large to small according to the average reduction in the Gini index, and the top 1000 are screened, resulting in a total of 1000 metabolites.

[0073] Finally, the above two analysis results were integrated to determine a group of related metabolites that were highly correlated with the biological age of the cardiovascular system, and they were applied to the construction and optimization of the cardiovascular age prediction model. Through the significance of differential analysis and the Boruta method to screen the importance, 832 related metabolites related to cardiovascular age were finally selected. This process not only helps to understand the role of metabolic changes in the aging of the cardiovascular system, but also provides potential biomarkers for personalized health management and disease intervention. The optimized feature set was further applied to the deep learning model for the prediction and prevention of cardiovascular diseases.

[0074] S2, constructing a cardiovascular age prediction model based on deep learning and using the first training data for training to obtain a trained cardiovascular age prediction model, the trained cardiovascular age prediction model includes a metabolic feature representation module and a feature representation processing module, the metabolic feature representation module includes an embedding layer, a multi-head self-attention layer, a feedforward fully connected layer and a layer normalization layer, the embedding layer, the multi-head self-attention layer and the feedforward fully connected layer are connected in sequence, the output features of the multi-head self-attention layer and the output features of the feedforward fully connected layer are first residually connected and then input into the layer normalization layer, and the feature representation processing module uses a multi-layer perceptron.

[0075] In a specific embodiment, the metabolomics feature matrix is ​​input into the metabolic feature representation module, and each metabolic feature in the metabolomics feature matrix with a dimension of n×m is first mapped into a 256-dimensional vector through an embedding layer to obtain an embedded feature representation matrix with a dimension of n×256, wherein n is the total number of samples and m is the number of metabolic features corresponding to the relevant metabolites; the multi-head self-attention layer contains 6 self-attention heads, and the n×256 embedded feature representation matrix is ​​mapped into 6 groups of different query matrices, key matrices and value matrices, and each group of query matrices, key matrices and value matrices passes through each self-attention layer. Head, output each output feature matrix, splice the 6 output feature matrices to obtain the splicing matrix, input the splicing matrix into the linear transformation layer, and obtain the output features of the multi-head self-attention layer; the feedforward fully connected layer includes the first fully connected layer and the second fully connected layer connected in sequence, the first fully connected layer uses the ReLU activation function, the output features of the multi-head self-attention layer are input into the feedforward fully connected layer, and pass through the first fully connected layer and the second fully connected layer in sequence to obtain the output features of the feedforward fully connected layer, the output features of the multi-head self-attention layer are added to the output features of the feedforward fully connected layer, and then input into the normalization layer to obtain the metabolic feature representation.

[0076] In a specific embodiment, the multilayer perceptron includes an input layer, a first hidden layer, a second hidden layer, a batch normalization layer, a Dropout layer and an output layer connected in sequence; the input layer is used to receive metabolic feature representation, the first hidden layer contains 256 neurons and uses a ReLU activation function; the second hidden layer contains 128 neurons and uses a ReLU activation function; the dropout ratio in the Dropout layer is 0.5, the output layer contains 1 neuron, and uses a linear activation function to output cardiovascular age.

[0077] In a specific embodiment, the total loss function used in the training process of the deep learning-based cardiovascular age prediction model is for:

[0078] ;

[0079] Where n is the total number of samples, , is the mean square error loss function of the i-th sample, is the weighted loss function of the i-th sample, α and β are the weight coefficients of the mean square error loss function and the weighted loss function respectively, and the weighted loss function is expressed as:

[0080] ;

[0081] in, represents the weight factor, , γ represents a hyperparameter, represents the actual age of the i-th sample, Represents the cardiovascular age predicted by the cardiovascular age prediction model based on deep learning for the i-th sample.

[0082] Specifically, a cardiovascular age prediction model based on deep learning is constructed, which includes a metabolic feature representation module and a feature representation processing module, wherein the metabolic feature representation module adopts the encoder structure of the improved Transformer model to perform feature representation learning on the metabolic features corresponding to the relevant metabolites. The Transformer model is particularly suitable for processing high-dimensional data with complex associations. The encoder structure of the improved Transformer model can efficiently capture the complex relationship between cardiovascular-related metabolites and provide high-quality input features for the feature representation processing module. The core of the encoder structure of the improved Transformer model is a structure based on the attention mechanism, which can process sequence data in parallel and capture the global relationship between features. The metabolic feature representation module includes an embedding layer, a multi-head self-attention layer, a feedforward fully connected layer, and a layer normalization layer.

[0083] Specifically, refer to Figure 2 The input of the metabolic feature representation module is the metabolomics feature matrix composed of the metabolic features corresponding to the preprocessed related metabolites, which contains 832 cardiovascular-related metabolites (e.g., lipids, amino acids, sugar metabolites, etc.). The feature matrix size is n×m, where n represents the total number of samples and m represents the number of metabolic features corresponding to the related metabolites, which is 832. During the training process, the number of samples is the total number of people in the healthy population.

[0084] Embedding layer: Feature mapping is achieved through an embedding matrix. In the embodiment of the present application, the embedding matrix is ​​a parameter matrix with a dimension of 832×256, where 832 is the number of metabolic features corresponding to the relevant metabolites, and 256 is the dimension of the set embedding space. Each metabolic feature is mapped to the corresponding row in the embedding matrix, thereby obtaining a dense vector of length 256, that is, the embedded feature representation matrix. These vectors represent the position of the metabolic features in the embedding space. Adding an embedding layer has many advantages:

[0085] 1. Feature enrichment: By mapping metabolic features from the original high-dimensional sparse space (832 dimensions) to a lower-dimensional dense space (256 dimensions), each metabolic feature can be represented and processed more efficiently, which helps the model capture the potential relationship between metabolic features instead of directly calculating in the original high-dimensional sparse space;

[0086] 2. Improve computational efficiency: The dimension of the embedding space is smaller than that of the original high-dimensional sparse space, so the computational and storage requirements are reduced, which improves the training and reasoning efficiency of the model;

[0087] 3. Enhanced model capabilities: By introducing high-dimensional embedding representations, the model can better capture the complex nonlinear relationships between metabolic features. The embedding layer enables the model to utilize richer feature information, thereby improving the accuracy of cardiovascular age prediction.

[0088] Multi-head self-attention layer: The embedded feature representation matrix (with dimension n×256) generated by the embedding layer is input into a multi-head self-attention layer with 6 self-attention heads. The working principle of the self-attention layer is as follows: First, the input embedded feature representation matrix is ​​transformed into a query matrix, a key matrix, and a value matrix through linear transformation, each of which has a dimension of n×256. In the multi-head self-attention mechanism, these linear transformations are performed independently for each attention head. Therefore, the input feature matrix is ​​mapped into 6 different query matrices, key matrices, and value matrices, each with a dimension of 256 / 6=42, that is, the dimension of each self-attention head is 42. Next, each self-attention head determines the relationship between each feature by calculating the similarity score between the query matrix and the key matrix. Specifically, for each self-attention head, the dot product of the query matrix and all key matrices is calculated, and the normalized attention weights are obtained through the Softmax function. Then, these attention weights are used to weighted average the value matrix to obtain the output feature matrix of each self-attention head. The output feature matrices of these 6 self-attention heads are then concatenated together to form a concatenated matrix of dimension n×256. Finally, the concatenated matrix is ​​mapped back to the original dimension through a linear transformation layer to generate the output features of the final multi-head self-attention layer. In this way, the multi-head self-attention layer effectively fuses the information from different self-attention heads, enabling the model to capture the complex feature relationships in the input data.

[0089] Feed-Forward Layer: The output features of the multi-head self-attention layer are passed as input to the feed-forward fully connected layer. The feed-forward fully connected layer consists of two linear transformations and an activation function (usually ReLU), which is mainly used to further process the features output by the self-attention layer to improve the nonlinear expression ability. The feed-forward fully connected layer includes the first fully connected layer and the second fully connected layer connected in sequence. First, the input features are transformed by the first linear transformation, that is, the first fully connected layer, to map each feature vector to a new vector space. The dimension is kept at 256, and the nonlinear activation function, the ReLU activation function, is applied for activation. Then, the features processed by the ReLU activation function are linearly transformed again through a second fully connected layer (the second linear transformation), and the output is consistent with the original input dimension, which is still n×256. This process enhances the complexity and nonlinearity of feature representation, allowing the model to learn deeper feature interactions.

[0090] Layer Normalization Layer: The output features of the feed-forward fully connected layer are residually connected (i.e. added) to the output features of the multi-head self-attention layer, and then layer normalization is performed. The residual connection ensures that the input features can directly affect the output, avoiding information loss caused by too deep layers. At the same time, the layer normalization layer further standardizes the features to improve the stability of the model and ensure smooth feature transmission during training. The final output metabolic feature representation is input into the feature representation processing module.

[0091] The feature representation processing module uses a multi-layer perceptron to represent the high-quality metabolic features extracted from the metabolic feature representation module for cardiovascular age prediction. The multi-layer perceptron consists of an input layer, two hidden layers, and an output layer. It gradually extracts the high-order relationship of features to achieve accurate cardiovascular age prediction.

[0092] Input layer: Receives metabolic feature representation from the metabolic feature representation module. The input dimension is n×256, where n is the number of samples and 256 is the feature dimension.

[0093] The first hidden layer: contains 256 neurons and uses the ReLU activation function to introduce nonlinearity and enhance the model's ability to express complex feature patterns.

[0094] The second hidden layer contains 128 neurons and uses the ReLU activation function to further refine the feature representation and further process the information extracted by the first hidden layer, so that the model can capture the characteristic patterns that contribute to cardiovascular age prediction in more detail.

[0095] Batch Normalization Layer and Dropout Layer: The output of the second hidden layer is normalized through the batch normalization layer to make the distribution of features more stable, accelerate the training process and improve the generalization ability of the model; at the same time, some neurons are randomly discarded through the Dropout layer to prevent the model from over-relying on specific features and reduce overfitting. The discard ratio is set to 0.5.

[0096] Output layer: The output layer contains 1 neuron, which uses a linear activation function to directly output the predicted cardiovascular age.

[0097] The embodiment of the present application designs a total loss function based on mean square error (MSE) and weighted loss on the basis of the cardiovascular age prediction model. The goal of the total loss function is to simultaneously reduce the overall prediction error and enhance the model's attention to abnormal samples. The mean square error loss function (MSE Loss) is used to measure the average difference between the predicted cardiovascular age and the actual age. The actual age in the embodiment of the present application is calculated based on the information recorded on the ID card; the weighted loss function (Weighted MSE Loss) introduces a weight factor related to the prediction error for each sample on the basis of the mean square error loss function to enhance the attention to samples with large errors. MSE Loss is a standard mean square error loss function used to measure the difference between the predicted cardiovascular age and the actual age; Weighted MSE Loss is a weighted loss function that gives higher weights to samples with large prediction errors, enhancing the model's attention to abnormal samples; α and β are weight coefficients of the two loss functions, respectively, used to adjust their influence on the total loss. Here, α and β are set to 1 and 0.1, respectively. γ in the weighted loss function is a hyperparameter used to control the degree of weight amplification. In the embodiment of the present application, it is set to 0.5. When the prediction error is large, the weight It will also increase, and the model will pay more attention to this sample. In each training batch, the loss of each sample is first calculated, including its MSE and Weighted MSE, and then these losses are aggregated by averaging to obtain the total loss function value of the current batch. The model is optimized using the Adaptive Moment Estimation (AdamW) optimization algorithm, which prevents model overfitting by dynamically adjusting the learning rate and introducing weight decay. During the training process, the first training data is divided into a training set and a validation set with a ratio of 7:3 and a batch size of 64. For each training batch, the total loss function is calculated, the gradient is calculated by the back propagation algorithm, and the model parameters are updated using the optimization algorithm. At the end of each complete training iteration, the validation set is used to evaluate the model performance and the total loss function of the validation set is calculated. At the same time, an early stopping mechanism is introduced. When the loss of the validation set no longer decreases within 5 complete training iterations, the training is terminated in advance to prevent the model from overfitting.

[0098] refer to Figure 3 , the trained cardiovascular age prediction model obtained by training in the embodiment of the present application performs well in predicting cardiovascular age. Its Pearson correlation coefficient on the validation set is 0.73, the determination coefficient is 0.76, and the total loss function value is 17.66. The Pearson correlation coefficient (0.73) and the determination coefficient (R²=0.76) between the cardiovascular age predicted by the trained cardiovascular age prediction model in the embodiment of the present application and the actual age show that there is a strong linear relationship between the cardiovascular age predicted by the trained cardiovascular age prediction model proposed in the embodiment of the present application and the actual age, and the total loss function value is low, indicating that the prediction error is small and the prediction accuracy of the model is high. This result shows that the trained cardiovascular age prediction model can not only accurately predict the cardiovascular age of an individual, but also effectively capture the metabolic characteristics related to its aging, thereby providing an important biomarker for individualized health management and disease intervention.

[0099] S3, obtaining a blood sample from the person to be tested and processing it to obtain a metabolomics feature matrix composed of metabolic features corresponding to relevant metabolites, inputting the metabolomics feature matrix into a trained cardiovascular age prediction model, passing through a metabolic feature representation module to obtain a metabolic feature representation, inputting the metabolic feature representation into a feature representation processing module, predicting the cardiovascular age of the person to be tested, and calculating the cardiovascular age acceleration.

[0100] Specifically, after the trained cardiovascular age prediction model is deployed, the metabolomics feature matrix composed of the metabolic characteristics corresponding to the relevant metabolites obtained by processing the blood samples of the person to be tested can be predicted to obtain the corresponding cardiovascular age of the person to be tested.

[0101] The embodiment of the present application adopts the concept of Cardiovascular Age Acceleration (CAA) to further characterize the difference between the cardiovascular age predicted by the trained cardiovascular age prediction model proposed in the embodiment of the present application and the actual age, that is, the part of the cardiovascular age predicted by the trained cardiovascular age prediction model proposed in the embodiment of the present application that exceeds or is lower than the actual age. The specific definitions are as follows:

[0102] Cardiovascular age acceleration = predicted cardiovascular age – actual age.

[0103] A positive CAA value indicates that the individual's cardiovascular state may be "older" than the actual age, suggesting that there may be accelerated cardiovascular aging; conversely, a negative CAA value indicates that the individual's cardiovascular state may be "younger" than the actual age, suggesting a healthier cardiovascular state. CAA can be used as an important indicator to assess an individual's health status and the aging rate of the cardiovascular system. Therefore, after obtaining the cardiovascular age of the person to be tested, the corresponding cardiovascular age acceleration of the person to be tested can be calculated in combination with the actual age of the person to be tested.

[0104] refer to Figure 4 , the distribution of cardiovascular age acceleration (CAA) is roughly normal, and most individuals have CAA values ​​between -8 and 8. This means that most people's cardiovascular age is close to their actual age, but a few people have larger CAA values, indicating that their cardiovascular age deviates significantly from their actual age. This result helps to identify those people whose cardiovascular age is significantly accelerated, and provides an important reference for personalized health management and early intervention for people with a higher risk of cardiovascular disease.

[0105] Further references Figure 5 As an implementation of the methods shown in the above figures, the present application provides an embodiment of a cardiovascular age prediction device based on deep learning and metabolic spectrum analysis. Figure 1 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0106] The embodiment of the present application provides a cardiovascular age prediction device based on deep learning and metabolic spectrum analysis, comprising:

[0107] The feature selection module 1 is configured to obtain blood samples from healthy people and people with cardiovascular diseases and perform metabolite extraction and chromatography-mass spectrometry analysis to obtain mass spectrometry data of each metabolite in several metabolomes corresponding to the healthy people and the people with cardiovascular diseases, perform data processing on the mass spectrometry data of each metabolite in each metabolome to obtain several metabolic features, wherein the metabolic features are formed by splicing the chromatographic retention time and the mass-to-charge ratio in the mass spectrometry data of the metabolites, perform feature selection on several metabolic features corresponding to the healthy people and the people with cardiovascular diseases, determine related metabolites, and construct first training data using the metabolic features corresponding to the related metabolites of the healthy people;

[0108] A model building module 2 is configured to build a cardiovascular age prediction model based on deep learning and use the first training data for training to obtain a trained cardiovascular age prediction model, wherein the trained cardiovascular age prediction model includes a metabolic feature representation module and a feature representation processing module, wherein the metabolic feature representation module includes an embedding layer, a multi-head self-attention layer, a feedforward fully connected layer, and a layer normalization layer, wherein the embedding layer, the multi-head self-attention layer, and the feedforward fully connected layer are connected in sequence, wherein the output features of the multi-head self-attention layer and the output features of the feedforward fully connected layer are first residually connected and then input into the layer normalization layer, and the feature representation processing module uses a multi-layer perceptron;

[0109] The prediction module 3 is configured to obtain a blood sample from the person to be tested and process it to obtain a metabolomics feature matrix composed of metabolic features corresponding to relevant metabolites, input the metabolomics feature matrix into the trained cardiovascular age prediction model, pass it through the metabolic feature representation module, obtain the metabolic feature representation, input the metabolic feature representation into the feature representation processing module, predict the cardiovascular age of the person to be tested, and calculate the cardiovascular age acceleration.

[0110] Figure 6 Schematic diagram of the hardware structure of the electronic device provided by the embodiment of the present invention. Figure 6 As shown, the electronic device of this embodiment includes: a processor 601 and a memory 602; wherein the memory 602 is used to store computer-executable instructions; the processor 601 is used to execute the computer-executable instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant description in the above method embodiment.

[0111] Optionally, the memory 602 may be independent or integrated with the processor 601 .

[0112] When the memory 602 is independently provided, the electronic device further includes a bus 603 for connecting the memory 602 and the processor 601 .

[0113] The embodiment of the present invention further provides a computer storage medium, in which computer execution instructions are stored. When the processor 601 executes the computer execution instructions, the above method is implemented.

[0114] The embodiment of the present invention further provides a computer program product, including a computer program. When the computer program is executed by the processor 601, the above method is implemented.

[0115] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0116] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to implement the solution of this embodiment.

[0117] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, each module may exist physically separately, or two or more modules may be integrated into one unit. The unit formed by the above modules may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0118] The above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor 601 to perform some steps of the methods of various embodiments of the present application.

[0119] It should be understood that the processor 601 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or the processor 601 may be any conventional processor 601, etc. The steps of the method disclosed in the invention may be directly embodied in the execution of the hardware processor 601, or may be executed by a combination of hardware and software modules in the processor 601.

[0120] The memory 602 may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk, or an optical disk.

[0121] The bus 603 may be an Industry Standard Architecture (ISA), a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus 603 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus 603 in the drawings of the present application is not limited to only one bus 603 or one type of bus 603.

[0122] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0123] An exemplary storage medium is coupled to the processor 601, so that the processor 601 can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor 601. The processor 601 and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor 601 and the storage medium can also exist as discrete components in an electronic device or a main control device.

[0124] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cardiovascular age prediction method based on deep learning and metabolic profile analysis, It is characterized in that The following steps are involved: Obtaining blood samples from healthy people and people with cardiovascular diseases and performing metabolite extraction and chromatography-mass spectrometry analysis to obtain mass spectrometry data of each metabolite in several metabolomes corresponding to the healthy people and the people with cardiovascular diseases, performing data processing on the mass spectrometry data of each metabolite in each metabolome to obtain several metabolic features, wherein the metabolic features are spliced ​​from the chromatographic retention time and the mass-to-charge ratio in the mass spectrometry data of the metabolites, performing feature selection on several metabolic features corresponding to the healthy people and the people with cardiovascular diseases to determine related metabolites, and constructing first training data using the metabolic features corresponding to the related metabolites of the healthy people; A cardiovascular age prediction model based on deep learning is constructed and trained using the first training data to obtain a trained cardiovascular age prediction model, wherein the trained cardiovascular age prediction model includes a metabolic feature representation module and a feature representation processing module, wherein the metabolic feature representation module includes an embedding layer, a multi-head self-attention layer, a feedforward fully connected layer, and a layer normalization layer, wherein the embedding layer, the multi-head self-attention layer, and the feedforward fully connected layer are sequentially connected, and the output features of the multi-head self-attention layer and the output features of the feedforward fully connected layer are first residually connected and then input into the layer normalization layer, and the feature representation processing module uses a multilayer perceptron; A blood sample of the person to be tested is obtained and processed to obtain a metabolomics feature matrix composed of metabolic features corresponding to relevant metabolites, the metabolomics feature matrix is ​​input into the trained cardiovascular age prediction model, and a metabolic feature representation is obtained through the metabolic feature representation module. The metabolic feature representation is input into the feature representation processing module to predict the cardiovascular age of the person to be tested, and the cardiovascular age acceleration is calculated.

2. The cardiovascular age prediction method based on deep learning and metabolic profile analysis according to claim 1, It is characterized in that The method used for feature selection includes a differential analysis feature screening method and a Boruta method feature screening method. The metabolite set obtained by the differential analysis feature screening method and the metabolite set obtained by the Boruta method feature screening method are intersected to obtain the related metabolites. The metabolic characteristics corresponding to the related metabolites are metabolic characteristics related to the age of cardiovascular disease.

3. The cardiovascular age prediction method based on deep learning and metabolic profile analysis according to claim 2, It is characterized in that The difference analysis feature screening method specifically includes: A two-sample t-test is performed on the metabolic characteristics of a healthy population and the metabolic characteristics of a population with cardiovascular diseases to obtain a p-value, and the p-value is corrected using the BH method to obtain a corrected p-value, and the metabolites corresponding to the metabolic characteristics that satisfy the corrected p-value less than 0.05 and the difference multiple greater than 2 or less than 1 / 2 constitute the metabolic set obtained by the differential analysis feature screening method, and the difference multiple is the multiple ratio of the median of the metabolic characteristics of the population with cardiovascular diseases to the median of the metabolic characteristics of the healthy population for each metabolite; The Boruta method feature screening method specifically includes: Constructing a random forest model, the random forest model includes a plurality of decision trees, taking the metabolic characteristics of a healthy population and the metabolic characteristics of a population suffering from cardiovascular diseases as original characteristics, randomly shuffling the values ​​in the original characteristics to generate shadow characteristics, constructing second training data based on the original characteristics and the shadow characteristics, using the second training data to train the random forest model, inputting the metabolic characteristics in the second training data into the random forest model, and obtaining a classification result, wherein the classification result is a healthy population or a population suffering from cardiovascular diseases, for each original characteristic and each shadow characteristic, calculating and accumulating the reduction of the Gini index caused by each original characteristic and each shadow characteristic on each node in each decision tree during the training process, obtaining the total reduction of the Gini index of each decision tree, averaging the total reduction of the Gini index of all decision trees, calculating the average reduction of the Gini index of each original characteristic and the average reduction of the Gini index of each shadow characteristic, and determining the maximum average reduction of the Gini index among all shadow characteristics according to the average reduction of the Gini index of each shadow characteristic; In response to when the average reduction of the Gini index of one of the original features is greater than the maximum average reduction of the Gini index of all the shadow features, marking the one of the original features as an important feature; In response to the average reduction in the Gini index of one of the original features being less than or equal to the maximum average reduction in the Gini index of all the shadow features, marking the one of the original features as an unimportant feature; The metabolites marked as corresponding to the important features are used as the metabolite set obtained by the feature screening method of the Boruta method.

4. The cardiovascular age prediction method based on deep learning and metabolic profile analysis according to claim 1, It is characterized in that The metabolomics feature matrix is ​​input into the metabolic feature representation module, and each metabolic feature in the metabolomics feature matrix with a dimension of n×m is first mapped into a 256-dimensional vector through an embedding layer to obtain an embedded feature representation matrix with a dimension of n×256, wherein n is the total number of samples and m is the number of metabolic features corresponding to related metabolites; the multi-head self-attention layer contains 6 self-attention heads, and the n×256 embedded feature representation matrix is ​​mapped into 6 groups of different query matrices, key matrices and value matrices, and each group of query matrices, key matrices and value matrices passes through each self-attention head to output each output feature matrix , concatenate the six output feature matrices to obtain a concatenated matrix, and input the concatenated matrix into the linear transformation layer to obtain the output features of the multi-head self-attention layer; the feedforward fully connected layer includes a first fully connected layer and a second fully connected layer connected in sequence, and the first fully connected layer uses a ReLU activation function. The output features of the multi-head self-attention layer are input into the feedforward fully connected layer, and pass through the first fully connected layer and the second fully connected layer in sequence to obtain the output features of the feedforward fully connected layer. The output features of the multi-head self-attention layer are added to the output features of the feedforward fully connected layer and then input into the layer normalization layer to obtain the metabolic feature representation.

5. The cardiovascular age prediction method based on deep learning and metabolic profile analysis according to claim 1, It is characterized in that The multilayer perceptron includes an input layer, a first hidden layer, a second hidden layer, a batch normalization layer, a Dropout layer and an output layer connected in sequence; the input layer is used to receive the metabolic feature representation, the first hidden layer includes 256 neurons and uses a ReLU activation function; the second hidden layer includes 128 neurons and uses a ReLU activation function; the dropout ratio in the Dropout layer is 0.5, the output layer includes 1 neuron, and uses a linear activation function to output the cardiovascular age.

6. The cardiovascular age prediction method based on deep learning and metabolic profile analysis according to claim 1, It is characterized in that The total loss function used in the training process of the deep learning-based cardiovascular age prediction model for: ； Where n is the total number of samples, , is the mean square error loss function of the i-th sample, is the weighted loss function of the i-th sample, α and β are the weight coefficients of the mean square error loss function and the weighted loss function respectively, and the weighted loss function is expressed as: ； in, represents the weight factor, , γ represents a hyperparameter, represents the actual age of the i-th sample, represents the cardiovascular age predicted by the cardiovascular age prediction model based on deep learning for the i-th sample.

7. A cardiovascular age prediction device based on deep learning and metabolic profile analysis, It is characterized in that include: a feature selection module configured to obtain blood samples from healthy people and people with cardiovascular diseases and perform metabolite extraction and chromatography-mass spectrometry analysis to obtain mass spectrometry data of each metabolite in several metabolomes corresponding to the healthy people and the people with cardiovascular diseases, perform data processing on the mass spectrometry data of each metabolite in each metabolome to obtain several metabolic features, wherein the metabolic features are formed by concatenating the chromatographic retention time and the mass-to-charge ratio in the mass spectrometry data of the metabolites, perform feature selection on several metabolic features corresponding to the healthy people and the people with cardiovascular diseases, determine relevant metabolites, and construct first training data using the metabolic features corresponding to the relevant metabolites of the healthy people; A model building module is configured to build a cardiovascular age prediction model based on deep learning and use the first training data for training to obtain a trained cardiovascular age prediction model, wherein the trained cardiovascular age prediction model includes a metabolic feature representation module and a feature representation processing module, wherein the metabolic feature representation module includes an embedding layer, a multi-head self-attention layer, a feedforward fully connected layer, and a layer normalization layer, wherein the embedding layer, the multi-head self-attention layer, and the feedforward fully connected layer are connected in sequence, wherein the output features of the multi-head self-attention layer and the output features of the feedforward fully connected layer are first residually connected and then input into the layer normalization layer, and the feature representation processing module uses a multilayer perceptron; The prediction module is configured to obtain a blood sample of the person to be tested and process it to obtain a metabolomics feature matrix composed of metabolic features corresponding to relevant metabolites, input the metabolomics feature matrix into the trained cardiovascular age prediction model, pass it through the metabolic feature representation module to obtain a metabolic feature representation, input the metabolic feature representation into the feature representation processing module, predict the cardiovascular age of the person to be tested, and calculate the cardiovascular age acceleration.

8. An electronic device, include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images