Metabolite target interaction prediction method and system for myocardial injury

Abstract

The application relates to the technical field of bioinformatics, and specifically discloses a metabolite target point interaction prediction method and system for myocardial injury, aiming to solve the problems that the prior art is insufficient in processing complex heterogeneous data, and is difficult to capture dynamic changes of a metabolic network and perform fine timing prediction. Multiple source heterogeneous data are integrated, a myocardial injury risk factor and metabolite target point knowledge base is constructed, and machine learning and deep learning technologies are used to realize risk factor weighting, static and timing dynamic risk assessment. By fusing a static risk level and a dynamic adjustment factor, the system can predict metabolite target point interactions related to an individual high-risk state, and generate personalized prevention and intervention strategy suggestions accordingly. The application comprises multiple functional modules such as data acquisition and preprocessing, knowledge base construction, feature engineering, risk modeling, timing optimization and target point prediction, so as to realize early, dynamic and accurate prediction of myocardial injury risk and identification of personalized intervention target points.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of bioinformatics, and specifically relates to a metabolite target interaction prediction method and system for myocardial injury. BACKGROUND

[0002] Cardiovascular disease is a major disease that seriously threatens human health worldwide, and its high morbidity and mortality have always been the focus of research and development in the medical field. Specifically, myocardial ischemia refers to insufficient blood perfusion to the heart, which leads to reduced oxygen supply to myocardial cells, energy metabolism disorder, and ultimately impaired cardiac function. On this basis, if myocardial ischemia recovers blood perfusion after a period of time, it will exacerbate the structural damage and functional disorder of myocardial cells, induce significant decline in heart function, and even induce malignant arrhythmia. This complex pathophysiological process is called myocardial ischemia-reperfusion injury. Given the serious consequences of myocardial ischemia-reperfusion injury, in-depth understanding of its mechanism, early warning, and development of effective intervention strategies are of great significance to reduce patient morbidity and mortality.

[0003] In existing research and clinical practice on myocardial ischemia-reperfusion injury, various technical approaches have been explored and applied to the diagnosis, treatment, and mechanism elucidation of the disease. For example, as a class of small molecule polypeptides with significant biological activity, cardiac peptides have been widely proven to have a clear preventive and therapeutic effect on myocardial ischemia-reperfusion injury. In-depth pharmacological studies have shown that cardiac peptides can effectively inhibit the abnormal release of myocardial enzymes after ischemia-reperfusion, thereby reducing the degree of myocardial cell damage. Its mechanism of action is that during ischemia-reperfusion injury, myocardial cells often undergo degeneration and necrosis, leading to impaired cell membrane function and large amounts of intracellular CK enzymes escaping into the serum, resulting in a significant increase in myocardial enzyme activity. Clinical application data show that after cardiac peptide intervention, the serum myocardial enzyme activity of patients is significantly reduced, which confirms its protective effect on the myocardium. In addition, arrhythmia, as a common complication of myocardial ischemia-reperfusion injury, one of its mechanisms is the dysfunction of myocardial cell membrane ion channels. In the pathological environment of ischemia-reperfusion injury, intracellular acidosis leads to an increase in hydrogen ion concentration, which promotes intracellular Ca2+ overload through H+ / Na+ exchange and Na+ / Ca2+ exchange mechanisms. These abnormal ion activities are prone to induce fatal arrhythmia. Cardiac peptides also show a positive effect in this regard, as they can effectively reduce the incidence of severe ventricular arrhythmia in rats with ischemia-reperfusion injury by inhibiting the inward calcium current of ventricular myocardial cells, thereby playing a myocardial protective role in multiple dimensions.

[0004] Therefore, how to construct a method and system capable of deeply integrating multi-source heterogeneous data, accurately predicting metabolite target interaction related to myocardial injury through advanced machine learning and deep learning technology, and realizing dynamic and accurate assessment and early identification of cardiovascular disease risk has become a key challenge and technical problem to be solved for those skilled in the art. SUMMARY

[0005] To achieve the above object, the present application provides the following technical scheme:

[0006] According to the first aspect of the present application, a metabolite target interaction prediction method for myocardial injury is claimed, comprising the following steps:

[0007] S1, obtaining individual physiological and biochemical data, clinical diagnosis and treatment records, medical image data, genomics data and metabolomics raw data, and performing cleaning, format unification, standardization, missing value filling, outlier processing and noise reduction operation on the collected heterogeneous data;

[0008] S2, constructing a myocardial injury risk factor and metabolite target knowledge base, extracting risk factors, biomarkers, metabolic pathways, protein targets and known drug-target interaction information related to myocardial injury through automatic text mining and knowledge graph construction technology, and forming a structured knowledge network;

[0009] S3, fusing the multi-source data preprocessed by S1 and the knowledge base information constructed by S2 to generate a comprehensive feature set containing clinical risk features, genomics features, metabolomics features, time series features and graph structure embedding based features extracted from the knowledge base; and using a recursive feature elimination based feature selection method to select a key feature subset;

[0010] S4, using an adaptive enhanced AdaBoost learning algorithm to calculate the weights of the key risk factors selected by S3 to construct a weighted risk factor model, and inputting the weights output by the weighted risk factor model and the key feature subset into a gradient boosting tree GBT model to construct a static risk level classification model, and the GBT model outputs the myocardial injury risk level of the patient at the current time point;

[0011] S5, using a long short-term memory LSTM network to perform time series analysis on the individual historical physiological and biochemical data and metabolomics data preprocessed by S1 to construct a time series data dynamic optimization model, and the LSTM model outputs a dynamic adjustment factor;

[0012] S6, fusing the static risk level output by the GBT model in S4 and the dynamic adjustment factor output by the LSTM model in S5, predicting the metabolite-target interaction related to the high-risk state of the individual through reasoning based on the metabolomics features and the knowledge base and / or prediction based on the machine learning model, and generating personalized prevention and intervention strategy suggestions according to the dynamically adjusted myocardial injury risk level and the predicted metabolite-target interaction.

[0013] Further, in S1, the physiological and biochemical data include numerical data of blood pressure, blood glucose, blood lipids, and heart rate.

[0014] The clinical diagnosis and treatment records include text data of medical history, diagnosis, medication, and treatment plan.

[0015] The medical image data include echocardiogram, magnetic resonance imaging (MRI), and computed tomography (CT) image data.

[0016] The genomics data include single nucleotide polymorphism (SNP) and gene expression profile sequence data; and the metabolomics raw data is obtained by liquid chromatography-mass spectrometry (LC-MS) or gas chromatography-mass spectrometry (GC-MS) technology.

[0017] Further, in S1, the text data is processed by natural language processing techniques, including word segmentation, part-of-speech tagging, and named entity recognition, and is converted into structured information.

[0018] The image data is processed by image processing algorithms to extract features.

[0019] The sequence data is genotyped and variant-identified.

[0020] The metabolomics raw data is processed by peak extraction, peak alignment, retention time correction, isotope deconvolution, and peak area normalization, and is preliminarily identified and quantified using a standard metabolite library.

[0021] Further, in S2, the automated text mining uses a named entity recognition (NER) model based on deep learning to identify entities and a relation extraction (RE) model to identify semantic relationships between entities, and constructs triplets.

[0022] The knowledge graph represents the extracted entities and their relationships in the form of nodes and edges, forming a structured knowledge network.

[0023] The knowledge base integrates known physical interactions between metabolites and myocardial-related protein targets, signal pathway regulation relationships, enzymatic reaction relationships, and gene expression regulation relationships.

[0024] Furthermore, in S3, the comprehensive feature set includes:

[0025] Clinical risk characteristics include physiological and biochemical indicators, demographic information, and structured risk factors extracted from clinical records;

[0026] Genomic characteristics, including variation information at specific gene loci and gene expression levels associated with myocardial injury;

[0027] Metabolomics features include the relative concentrations of identified metabolites, metabolite ratios, and activity fractions of specific metabolic pathways obtained through pathway enrichment analysis.

[0028] Temporal characteristics are extracted by performing differential and moving average processing on historical physiological and biochemical data and metabolomics data to identify trend and fluctuation features.

[0029] Features based on graph structure embedding extracted from the knowledge base are used to embed the knowledge graph into a graph neural network.

[0030] Furthermore, in S4, the AdaBoost algorithm iteratively trains weak classifiers, adjusts the weights of training samples in each iteration, and assigns weights to weak classifiers based on their classification performance. The final strong classifier is composed of a weighted vote of all weak classifiers.

[0031] The GBT model constructs multiple decision trees sequentially, with each tree attempting to correct the prediction error of the previous tree, and optimizes the loss function using the gradient descent algorithm during training.

[0032] Furthermore, in S5, the LSTM network internally includes an input gate, a forget gate, an output gate, and a cell state. The input gate controls when and to what extent new information is written into the cell state. The forget gate controls how much information in the cell state of the previous time step is forgotten. The output gate controls how much information in the cell state is output to the hidden state of the current time step. The cell state is responsible for storing and transmitting information in long-term sequences.

[0033] Furthermore, in S6, the fusion mechanism is to use the dynamic adjustment factor output by the LSTM model as an independent predictor variable or as a weighting coefficient to correct the static risk level output by the GBT model.

[0034] The reasoning based on metabolomics features and the knowledge base analyzes significantly abnormal metabolites in individual metabolomics data and identifies the metabolic pathways and protein targets affected by them in conjunction with the knowledge base; the prediction based on machine learning models trains graph neural networks or random forest models and uses the individual comprehensive feature set as input to predict existing metabolite-target interactions.

[0035] According to a second aspect of the present invention, the present invention claims protection for a metabolite target interaction prediction system for myocardial injury, comprising:

[0036] Data acquisition and preprocessing module: used to perform multi-source data acquisition and preprocessing functions in step S1;

[0037] Knowledge base construction module: Used to perform the knowledge base construction function in step S2;

[0038] Metabolomics data analysis module: used for spectral processing, metabolite identification and quantification of raw metabolomics data;

[0039] Feature engineering module: Used to perform feature engineering functions of S3;

[0040] Risk Factor Weighting and Selection Module: Used to perform the feature selection function in step S3 and the weighted risk factor model construction function in step S4;

[0041] Risk level classification module: Used to perform the static risk level classification model construction function in step S4;

[0042] Time series data dynamic optimization module: used to perform the time series data dynamic optimization model construction function in step S5;

[0043] Risk assessment and target prediction module: used to perform dynamic adjustment of risk levels and prediction of metabolite-target interactions in step S6; and

[0044] Intervention strategy recommendation module: This module is used to generate personalized intervention strategy recommendations for step S6.

[0045] Furthermore, the data acquisition and preprocessing module is configured with multiple data interfaces, which support DICOM, HL7, mzML, and FASTQ data formats. The data acquisition and preprocessing module internally includes a data cleaning unit, a standardization unit, a missing value imputation unit, an outlier processing unit, and a data transformation unit.

[0046] This application relates to the field of bioinformatics technology, specifically disclosing a method and system for predicting metabolite target interactions in myocardial injury. It aims to address the shortcomings of existing technologies in handling complex heterogeneous data, capturing dynamic changes in metabolic networks, and performing precise temporal predictions. The system integrates multi-source heterogeneous data to construct a knowledge base of myocardial injury risk factors and metabolite targets, and utilizes machine learning and deep learning techniques to achieve risk factor weighting and static and temporal dynamic risk assessment. By fusing static risk levels and dynamic adjustment factors, the system can predict metabolite target interactions related to an individual's high-risk state and generate personalized prevention and intervention strategy recommendations accordingly. This invention includes multiple functional modules such as data acquisition and preprocessing, knowledge base construction, feature engineering, risk modeling, temporal optimization, and target prediction, thereby achieving early, dynamic, and accurate prediction of myocardial injury risk and identification of personalized intervention targets. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the process of a method for predicting metabolite target interactions for myocardial injury, as claimed in an embodiment of the present invention.

[0048] Figure 2 This is a structural block diagram of a metabolite target interaction prediction system for myocardial injury, as claimed in an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0050] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0051] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0052] Metabolomics, as an emerging branch of systems biology, has become an important tool for quantifying and characterizing small-molecule metabolites in organisms since its rise in the 1990s. Of particular note is non-targeted metabolomics, which, with its unique advantage of simultaneously quantifying large-scale metabolite data, has been widely applied in elucidating the pathophysiological mechanisms of diseases, discovering biomarkers, and conducting disease diagnostic research. This technology primarily involves the dynamic tracking and analysis of metabolites in cells, tissues, or body fluids. Given the relatively stable existence of metabolites in body fluids, metabolomics analysis has become a powerful tool for disease diagnostic research. Currently, the discovery of novel metabolites associated with myocardial ischemia-reperfusion injury using metabolomics methods undoubtedly has significant theoretical and practical implications for understanding disease progression and developing novel prevention and treatment strategies.

[0053] However, with the deepening of medical research and the increasing demand for precision and personalization in clinical diagnosis and treatment, some inherent characteristics of the aforementioned existing technologies at the principle level have gradually revealed deep-seated limitations in addressing early warning and precise intervention for cardiovascular diseases, especially myocardial ischemia-reperfusion injury. Specifically, although the discovery and application of cardiac peptides have provided an effective approach to the treatment of myocardial injury, and their mechanism of action has been elucidated to some extent, this is more of a post-event intervention, i.e., therapeutic protection after the injury has already occurred or begun to occur. It does not provide a comprehensive, prospective mechanism to predict the risk of myocardial injury in individuals under complex physiological conditions, or to accurately identify which type of individual will benefit most from which intervention. In other words, current research on cardiac peptides focuses on their therapeutic effects on known injuries, rather than constructing a complex model that can predict the dynamic interaction between metabolites and targets before injury occurs.

[0054] Furthermore, while metabolomics, especially non-targeted metabolomics, can efficiently identify and quantify a large number of metabolites and plays an important role in disease diagnosis and the discovery of new metabolites, it still faces significant challenges in constructing refined, multi-dimensional, and time-series predictive models of metabolite-target interactions. Traditional metabolomics analysis often focuses on differences in metabolite levels at a specific time point, and its ability to capture the dynamic evolution of metabolic networks during disease development is relatively limited, making it difficult to fully reveal the complex and non-linear interactions between metabolites and biological targets. The development of cardiovascular disease is a complex process involving multiple factors and stages, involving genetics, environment, lifestyle, and the fine regulation of metabolic networks in the body. Simple correlation analysis of metabolite levels alone is insufficient to accurately capture and quantify these intricate risk factors, let alone establish a comprehensive predictive model that integrates clinical risk factors extracted from text mining, metabolomics data, and time-series information. Existing technologies lack systematic methodologies and efficient computational tools for processing high-dimensional, heterogeneous data and extracting key metabolite-target interactions with high predictive power. Traditional statistical methods or experience-based judgments often fail to fully uncover hidden patterns in data, especially when considering the weights of different risk factors, nonlinear relationships, and time dependencies. Their predictive accuracy and robustness often fall short of the needs for early identification of high-risk individuals in clinical settings. This limitation makes it difficult to accurately identify high-risk individuals in the early stages, particularly before the disease presents obvious clinical symptoms, thus missing the optimal window for prevention and intervention.

[0055] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a method for predicting metabolite target interactions in myocardial injury, comprising the following steps:

[0056] S1 acquires individual physiological and biochemical data, clinical diagnosis and treatment records, medical imaging data, genomics data, and raw metabolomics data, and performs cleaning, format unification, standardization, missing value imputation, outlier processing, and noise reduction on the collected heterogeneous data.

[0057] S2. Construct a knowledge base of myocardial injury risk factors and metabolite targets. Through automated text mining and knowledge graph construction technology, extract information on risk factors, biomarkers, metabolic pathways, protein targets, and known drug-target interactions related to myocardial injury, and form a structured knowledge network.

[0058] S3, the multi-source data preprocessed in S1 is fused with the knowledge base information constructed in S2 to generate a comprehensive feature set including clinical risk features, genomic features, metabolomics features, temporal features, and graph-based embedding features extracted from the knowledge base; and a feature selection method based on recursive feature elimination is used to screen out a subset of key features.

[0059] S4. The AdaBoost learning algorithm is used to calculate the weights of the key risk factors selected in S3 to construct a weighted risk factor model. The weights output by the weighted risk factor model and the subset of key features are then input into the Gradient Boosting Tree (GBT) model to construct a static risk level classification model. The GBT model outputs the patient's myocardial injury risk level at the current time point.

[0060] S5, using a Long Short-Term Memory (LSTM) network to perform time-series analysis on the individual historical physiological and biochemical data and metabolomics data preprocessed in S1 to construct a time-series data dynamic optimization model, the LSTM model outputs a dynamic adjustment factor;

[0061] S6. The static risk level output by the GBT model in S4 is fused with the dynamic adjustment factor output by the LSTM model in S5. Through reasoning based on metabolomics features and the knowledge base and / or prediction based on machine learning models, the interaction of metabolite targets related to an individual's high-risk status is predicted. Based on the dynamically adjusted myocardial injury risk level and the predicted metabolite target interactions, personalized prevention and intervention strategy recommendations are generated.

[0062] In this embodiment, individual physiological and biochemical data, clinical records, medical imaging data, genomics data, and raw metabolomics data are acquired. The physiological and biochemical data includes, but is not limited to, numerical data of routine indicators such as blood pressure, blood glucose, blood lipids, and heart rate. The clinical records include, but are not limited to, textual data of medical history, diagnosis, medication, and treatment plans. The medical imaging data includes, but is not limited to, echocardiography, magnetic resonance imaging (MRI), and computed tomography (CT) images. The genomics data includes, but is not limited to, single nucleotide polymorphisms (SNPs) and gene expression profile sequence data. The raw metabolomics data includes, but is not limited to, raw mass spectra, chromatograms, and their corresponding peak intensities, retention times, and mass-to-charge ratios obtained using liquid chromatography-mass spectrometry (LC-MS) or gas chromatography-mass spectrometry (GC-MS). The acquired heterogeneous data undergoes cleaning, format standardization, normalization, missing value imputation, outlier removal, and noise reduction to ensure data quality and consistency for subsequent analyses. Textual data is processed using natural language processing techniques for word segmentation, part-of-speech tagging, and named entity recognition, and then transformed into structured information. Image data undergoes feature extraction using image processing algorithms, such as texture, morphological, and grayscale features. Sequence data is used for genotyping and variant identification. Raw metabolomics data is processed through peak extraction, peak alignment, retention time correction, isotope deconvolution, and peak area normalization, and preliminary identification and quantification of metabolites are performed using a standard metabolite library.

[0063] S2 extracts information on risk factors, biomarkers, metabolic pathways, protein targets, and known drug-target interactions related to myocardial injury from publicly available medical literature, clinical guidelines, and disease databases using automated text mining and knowledge graph construction techniques. Risk factors include, but are not limited to, hypertension, diabetes, hyperlipidemia, smoking history, age, sex, body mass index (BMI), family history, and specific genetic susceptibility loci. The text mining technique employs a deep learning-based Named Entity Recognition (NER) model to identify entities and a Relation Extraction (RE) model to identify semantic relationships between entities, constructing triples (entity 1, relation, entity 2). The knowledge graph represents the extracted entities and their relationships as nodes and edges, forming a structured knowledge network. Simultaneously, the knowledge base integrates known physical interactions, signaling pathway regulation relationships, enzyme-catalyzed reaction relationships, and gene expression regulation relationships between known metabolites and myocardial-related protein targets, forming a metabolite-target interaction network. This knowledge base provides structured prior knowledge for subsequent feature engineering, risk assessment model training, and target prediction.

[0064] S3 performs feature engineering and feature selection: It fuses the multi-source data preprocessed in the first step with the knowledge base information constructed in the second step to generate a comprehensive feature set for machine learning models. The comprehensive feature set includes:

[0065] The first aspect is clinical risk characteristics, including physiological and biochemical indicators, demographic information, and structured risk factors extracted from clinical records.

[0066] The second aspect is genomic characteristics, including variation information at specific gene loci and gene expression levels associated with myocardial injury.

[0067] Thirdly, metabolomics characteristics, including the relative concentrations of identified metabolites, metabolite ratios, and activity fractions of specific metabolic pathways obtained through pathway enrichment analysis.

[0068] Fourthly, temporal characteristics are extracted by performing differential and moving average processing on historical physiological and biochemical data and metabolomics data.

[0069] Fifthly, features based on graph structure embedding extracted from the knowledge base are used to embed the knowledge graph into a graph neural network, mapping entities and relationships to a low-dimensional vector space, thereby capturing complex relationship information between entities.

[0070] For the comprehensive feature set, a feature selection method based on recursive feature elimination (RFE) is employed. The RFE method works as follows: First, a basic machine learning model is trained on all features. Second, based on the model training results, the features with the lowest importance are removed. Next, the model is retrained on the remaining features, and the step of removing the least important features is repeated. This iterative process continues until a preset feature quantity threshold is reached, or until the model's performance metrics reach their optimal level. This method filters out a subset of key features that have the highest correlation with and greatest contribution to myocardial injury risk prediction, thereby reducing model complexity and improving model generalization ability.

[0071] In S1 and S3, the variation information of the specific gene loci refers to genetic variations at loci that have been proven in scientific research or clinical practice, through genomic analysis techniques such as genome-wide association studies (GWAS) and targeted sequencing, to be significantly associated with myocardial injury, cardiovascular disease risk, or related pathophysiological processes such as lipid metabolism, inflammatory response, blood pressure regulation, and coagulation function. These loci typically exist in the form of single nucleotide polymorphisms, insertions / deletions, etc. Examples include, but are not limited to, known risk loci located in the 9p21 chromosome region (strongly associated with the risk of coronary heart disease and myocardial infarction), the APOE gene (associated with lipid metabolism and atherosclerosis), the CETP gene (associated with high-density lipoprotein cholesterol metabolism), the PCSK9 gene (associated with low-density lipoprotein cholesterol levels), and the ACE gene (associated with angiotensin-converting enzyme activity and hypertension risk). When constructing a feature set, the variation information of the specific gene loci can be represented as the genotype (e.g., AA, AG, GG), allele frequency, or a risk score based on a genetic model.

[0072] S4. Constructing a Weighted Risk Factor Model and a Static Risk Level Classification Model: This invention utilizes an adaptive reinforcement learning algorithm to calculate the weights of the key risk factors selected in the third step, thereby constructing a weighted risk factor model. The AdaBoost algorithm iteratively trains a series of weak classifiers. In each iteration, the algorithm adjusts the weights of the training samples, assigning higher weights to samples previously misclassified by the weak classifiers, thus giving them more attention in subsequent training. After each weak classifier is trained, it is assigned a weight based on its classification performance. The final strong classifier is composed of a weighted vote from all the weak classifiers. The AdaBoost model outputs the contribution weights for each key risk factor, reflecting the relative influence of different risk factors on the risk of myocardial injury.

[0073] Furthermore, this invention combines the weights output by the weighted risk factor model with the key feature subset selected by S3, and inputs this subset into the gradient boosting tree model to construct a static risk level classification model. The GBT model is an ensemble learning model that sequentially constructs multiple decision trees, each attempting to correct the prediction error of the previous tree. During training, the GBT model utilizes a gradient descent algorithm to optimize the loss function, where each new decision tree is designed to fit the negative gradient of the prediction residual of the previous tree. The final output of the GBT model is the patient's myocardial injury risk level at the current time point, categorized into low, medium, and high risk. The classification result reflects the likelihood of myocardial injury occurring in the patient under static physiological conditions.

[0074] S5 Constructing a Dynamic Optimization Model for Time-Series Data: This invention utilizes a Long Short-Term Memory (LSTM) network to perform time-series analysis on the preprocessed individual historical physiological and biochemical data and metabolomics data from the first step, in order to construct a dynamic optimization model for time-series data. The LSTM network is a special type of recurrent neural network (RNN) structure, which internally includes an input gate, a forget gate, an output gate, and cell states. The input gate controls when and to what extent new information is written into the cell state. The forget gate controls how much information in the cell state from the previous time step is forgotten. The output gate controls how much information from the cell state is output to the hidden state of the current time step. The cell states are responsible for storing and transmitting information over long time sequences.

[0075] The LSTM model receives continuous time-series physiological and biochemical indicators and metabolomics data sequences from the patient as input. The LSTM network effectively captures and learns long-term dependencies and dynamic patterns in the time-series data through its unique gating mechanism. After processing the sequence data, the LSTM model outputs a dynamic adjustment factor. This dynamic adjustment factor, based on the patient's latest time-series data, reflects the dynamic trend of their myocardial injury risk relative to static assessment.

[0076] S6 is used for dynamic adjustment of myocardial injury risk level and prediction of metabolite target interactions:

[0077] The static risk level output by the GBT model in step S4 is fused with the dynamic adjustment factor output by the LSTM model in step 5 to achieve dynamic and precise adjustment of the myocardial injury risk level. The fusion mechanism involves using the dynamic adjustment factor output by the LSTM model as an independent predictor variable or as a weighting coefficient to correct the static risk level output by the GBT model. For example, if the LSTM model predicts an upward trend in risk, even if the GBT model assesses it as medium risk, the dynamically adjusted risk level is raised to high risk; conversely, the same applies. This dynamically adjusted risk level more accurately reflects the actual risk status of myocardial injury at different time points.

[0078] Based on this dynamic risk assessment, this invention further predicts metabolite target interactions associated with an individual's high-risk state. The prediction method includes the following aspects:

[0079] Firstly, reasoning based on metabolomics characteristics and a knowledge base: For individuals identified as high-risk, significantly abnormal metabolites in their metabolomics data are analyzed. Combined with the metabolite target knowledge base constructed in the second step, the metabolic pathways affected by these abnormal metabolites and related protein targets are identified. The analysis employs methods such as overexpression analysis or gene set enrichment analysis to identify metabolic pathways highly associated with myocardial injury pathways and exhibiting abnormalities in patients.

[0080] Secondly, predictions based on machine learning models: A specialized machine learning model is trained, using a comprehensive feature set of high-risk individuals as input and known metabolite-target interactions as training labels, to predict new or perturbed metabolite-target interactions present in that individual. The model can identify which metabolites exhibit abnormal binding, activation, or inhibition of which protein targets under specific physiological and pathological conditions.

[0081] The prediction results will clearly indicate which metabolites and their corresponding protein targets have undergone significant dynamic changes or abnormal interactions under specific myocardial injury risk states.

[0082] Personalized intervention strategy recommendations are generated: Based on dynamically adjusted myocardial injury risk levels and predicted metabolite target interactions, personalized prevention and intervention strategy recommendations are generated. These recommendations include, but are not limited to, lifestyle interventions, optimization of drug treatment regimens, and subsequent monitoring protocols. The recommendations directly target the individual's identified risk factors, abnormal metabolic pathways, and perturbed targets, aiming to minimize the incidence or severity of myocardial injury.

[0083] Furthermore, in S1, the physiological and biochemical data includes numerical data of conventional indicators such as blood pressure, blood glucose, blood lipids, and heart rate;

[0084] The clinical diagnosis and treatment records include textual data such as medical history, diagnosis, medication, and treatment plan;

[0085] The medical imaging data includes echocardiography, magnetic resonance imaging (MRI), computed tomography (CT) images.

[0086] The genomic data includes single nucleotide polymorphisms (SNPs) and gene expression profile sequence data; the raw metabolomics data are acquired using liquid chromatography-mass spectrometry (LC-MS) or gas chromatography-mass spectrometry (GC-MS).

[0087] In this embodiment, this step is fundamental to the entire prediction method, aiming to collect and standardize biomedical data from different sources and formats. Specifically, the collected data types are extremely diverse, covering multiple dimensions that comprehensively reflect an individual's physiological state and disease progression. This includes individual physiological and biochemical data, such as blood pressure, blood glucose, blood lipids, and routine heart rate indicators, which are typically stored in the form of numerical matrices. Simultaneously, clinical records, such as detailed medical histories, diagnoses, medications, surgical records, and treatment plans, exist in unstructured text format. Medical imaging data, such as echocardiograms, cardiac MRI image sequences, and coronary computed tomography (CT) images, are image-based and carry important information about cardiovascular structure and function. Genomic data, including single nucleotide polymorphism (SNP) site information, gene copy number variations, and expression profiles of specific genes, are typically presented in sequence or expression matrix formats. Crucially, the raw metabolomics data, obtained through high-throughput liquid chromatography-mass spectrometry (LC-MS) or gas chromatography-mass spectrometry (GC-MS), includes complex spectroscopic data such as raw mass spectra, chromatograms, corresponding peak intensities, retention times, and mass-to-charge ratios. Rigorous preprocessing of these heterogeneous data is essential. This process aims to eliminate noise, correct biases, fill missing values, and standardize data formats to ensure the quality and consistency of subsequent analyses. For text-based data, Natural Language Processing (NLP) techniques are employed, transforming it into structured information through dictionary-based and statistical model-based word segmentation, part-of-speech tagging, and named entity recognition (NER), extracting key entities such as disease names, drug names, symptoms, and their associations. Image-based data utilizes deep learning-driven image processing algorithms to segment and identify lesion regions and extract quantitative indicators such as texture features, morphological features, and grayscale features. Sequence-based data requires bioinformatics processing such as sequence alignment, variant detection, genotyping, and expression quantification. The processing of raw metabolomics data is particularly complex, involving a series of steps including baseline correction, noise filtering, accurate peak detection, multi-sample peak alignment, retention time correction, isotope peak deconvolution, and peak area or height normalization. Subsequently, preliminary metabolite identification and quantification are performed using internationally standardized metabolite libraries, generating structured relative or absolute quantitative data matrices of metabolites. This preprocessing work employs a combination of automated workflows and manual quality control to ensure high data quality and reliability.

[0088] Furthermore, in step S1, the text data is segmented, part-of-speech tagging is performed, and named entity recognition is performed using natural language processing technology, and then converted into structured information;

[0089] Feature extraction is performed on the image data using image processing algorithms;

[0090] Genotyping and variant identification are performed on the sequence data;

[0091] The raw metabolomics data were processed by peak extraction, peak alignment, retention time correction, isotope deconvolution, and peak area normalization, and the metabolites were preliminarily identified and quantified using a standard metabolite library.

[0092] Furthermore, in S2, the automated text mining uses a deep learning-based Named Entity Recognition (NER) model to identify entities and a Relation Extraction (RE) model to identify semantic relationships between entities, thus constructing triples.

[0093] The knowledge graph represents the extracted entities and their relationships in the form of nodes and edges, forming a structured knowledge network;

[0094] The knowledge base integrates known physical interactions, signaling pathway regulation relationships, enzyme-catalyzed reaction relationships, and gene expression regulation relationships between metabolites and myocardial-related protein targets.

[0095] In this embodiment, this step aims to structurally integrate knowledge scattered across massive amounts of medical literature and databases, providing rich prior knowledge for subsequent feature engineering and model training. The construction of this knowledge base relies on automated text mining and knowledge graph technologies. By utilizing advanced text mining engines from publicly available medical literature databases, disease databases, and pathway databases such as PubMed, OMIM, KEGG, DrugBank, and Reactome, risk factors, biomarkers, metabolic pathways, protein targets, and known drug-target interactions related to myocardial injury are automatically identified and extracted. These risk factors include not only traditional clinical indicators such as hypertension, diabetes, hyperlipidemia, smoking history, age, gender, body mass index (BMI), and family history, but also in-depth mining of specific genetic susceptibility gene loci. The core of the text mining technology is a deep learning-based Named Entity Recognition (NER) model. For example, a pre-trained biomedical language model, BERT, combined with a Conditional Random Field (CRF) layer, is used to accurately identify medical entities such as diseases, genes, proteins, metabolites, and drugs. Subsequently, relation extraction (RE) models, such as those based on Bi-LSTM combined with attention mechanisms, are used to identify semantic relationships between these entities, such as "gene A is related to disease B" and "drug C acts on target D," thus constructing triples of the form (entity 1, relation, entity 2). These triples are further used to construct a knowledge graph, representing entities as nodes and relations as edges, forming a highly structured knowledge network, which is stored and managed using graph databases such as Neo4j. Simultaneously, this knowledge base deeply integrates known physical interactions, signaling pathway regulation relationships, enzyme-catalyzed reaction relationships, and gene expression regulation relationships between metabolites and myocardial-related protein targets, forming a comprehensive metabolite-target interaction network. For example, information is extracted from protein interaction databases such as STRING and BioGRID, and metabolic pathway regulation relationships are obtained from pathway databases such as KEGG and Reactome. This knowledge base is not only a static knowledge store but also equipped with a knowledge update and management unit, regularly acquiring information from the latest research literature and public databases for incremental updates and verification to ensure its timeliness and accuracy.

[0096] Furthermore, in S3, the comprehensive feature set includes:

[0097] Clinical risk characteristics include physiological and biochemical indicators, demographic information, and structured risk factors extracted from clinical records;

[0098] Genomic characteristics, including variation information at specific gene loci and gene expression levels associated with myocardial injury;

[0099] Metabolomics features include the relative concentrations of identified metabolites, metabolite ratios, and activity fractions of specific metabolic pathways obtained through pathway enrichment analysis.

[0100] Temporal characteristics are extracted by performing differential and moving average processing on historical physiological and biochemical data and metabolomics data to identify trend and fluctuation features.

[0101] Features based on graph structure embedding extracted from the knowledge base are used to embed the knowledge graph into a graph neural network.

[0102] In this embodiment, this step involves fusing the preprocessed data obtained in the previous two steps with structured knowledge, and extracting a comprehensive feature set that has the greatest information gain for predicting the risk of myocardial injury. This comprehensive feature set is carefully designed as a multi-dimensional, multi-level representation:

[0103] Firstly, clinical risk characteristics are directly derived from the physiological and biochemical indicators, demographic information, and risk factors extracted from clinical records, such as smoking status and history of diabetes. These characteristics reflect the patient's basic health status and past medical history.

[0104] Secondly, genomic characteristics, including variation information of specific gene loci obtained through genotyping, and gene expression levels related to myocardial injury obtained through RNA sequencing or gene chips.

[0105] Third, metabolomics characteristics, including the relative or absolute concentrations of identified metabolites, the ratios of specific metabolites, and the activity fractions of specific metabolic pathways obtained through pathway enrichment analysis, directly reflect the metabolic state of cardiomyocytes.

[0106] Fourth, temporal characteristics: By performing differential, moving average, and exponential smoothing on the patient's historical physiological and biochemical data and metabolomics data sequences, dynamic change patterns such as trends, fluctuations, and periods can be extracted. For example, the long-term fluctuation of blood pressure or the postprandial peak change rate of blood glucose. These characteristics are crucial for capturing dynamic information on disease progression.

[0107] Fifth, the graph-based embedding features extracted from the knowledge base are used to embed the knowledge graph constructed in the second step through graph neural networks or graph attention networks. This process maps entity nodes and relation edges in the knowledge graph to a low-dimensional vector space, thereby capturing complex association information, structural information, and semantic similarity between entities, generating distributed feature vectors that represent the position and connections of individuals in the knowledge graph. For the resulting large and comprehensive feature set, a feature selection method based on recursive feature elimination (RFE) is adopted. This aims to identify the key feature subset with the highest contribution and relevance to myocardial injury risk prediction, while reducing the dimensionality and complexity of the model and improving its generalization ability. The RFE method works as follows: First, a basic machine learning model, such as a support vector machine or ridge regression model, is selected and trained on a complete dataset containing all original features. After training, the feature with the lowest importance is identified based on the model's evaluation of feature importance. Then, this feature with the lowest importance is removed from the feature set. Next, the basic model is retrained on the remaining feature subset, and the feature with the lowest current importance is identified and removed again. This iterative process continues until the number of features reaches a preset threshold, or until the model's performance metrics are optimal on the validation set. RFE (Reference-Based Evaluator) systematically eliminates redundant and low-relevance features, ultimately yielding a concise and efficient subset of key features.

[0108] Furthermore, in S4, the AdaBoost algorithm iteratively trains weak classifiers, adjusts the weights of training samples in each iteration, and assigns weights to weak classifiers based on their classification performance. The final strong classifier is composed of a weighted vote of all weak classifiers.

[0109] The GBT model constructs multiple decision trees sequentially, with each tree attempting to correct the prediction error of the previous tree, and optimizes the loss function using the gradient descent algorithm during training.

[0110] In this embodiment, this step aims to quantify and assess the identified key risk factors and, based on this, to classify the patient into a preliminary static risk level.

[0111] First, an adaptive reinforcement learning algorithm is used to calculate the weights of the key risk factors selected in the third step to construct a weighted risk factor model. The AdaBoost algorithm is an iterative ensemble learning method that trains a series of weak classifiers sequentially. In each iteration, the AdaBoost algorithm adjusts the weights of the training samples based on the prediction results of the previous weak classifier: the weights of misclassified samples are increased to receive more attention in subsequent iterations, while the weights of correctly classified samples are decreased. After each weak classifier is trained, it is assigned a weight based on its classification performance, with better-performing classifiers receiving higher weights. The final strong classifier is composed of a weighted vote from all the weak classifiers. The AdaBoost model outputs the contribution weights for each key risk factor, which objectively reflect the relative impact and predictive importance of different risk factors on the risk of myocardial injury. Subsequently, the weights output by the weighted risk factor model are combined with the key feature subset selected in the third step and input into a gradient boosting tree model to construct a static risk level classification model. The GBT model is also a powerful ensemble learning model. It builds multiple decision trees sequentially, with each subsequent tree aiming to correct the prediction error of the previous tree.

[0112] Specifically, the GBT model utilizes gradient descent to optimize a predefined loss function during training. Each new decision tree does not directly fit the output value, but rather fits the negative gradient of the predicted residuals of the previous tree. This cumulative learning allows the GBT model to progressively approach the optimal solution and handle complex nonlinear relationships. After receiving weighted key risk factors and other non-temporally specific features as input, the GBT model ultimately outputs the patient's myocardial injury risk level at the current time point. This risk level is divided into three discrete categories: low risk, medium risk, and high risk. The classification results reflect the likelihood of myocardial injury in a patient under static physiological conditions, providing an initial benchmark for subsequent dynamic adjustments.

[0113] Subsequently, a dynamic optimization model for time-series data is constructed. This step aims to capture the dynamic evolution trend of myocardial injury risk in patients, compensating for the shortcomings of static models in handling time-dependent information. This invention utilizes a Long Short-Term Memory (LSTM) network to perform time-series analysis on the preprocessed individual historical physiological and biochemical data and metabolomics data sequences from the first step, in order to construct a dynamic optimization model for time-series data. The LSTM network is a special type of recurrent neural network (RNN) structure, cleverly designed to solve the gradient vanishing or gradient explosion problems that traditional RNNs encounter when processing long-sequence data, effectively learning and memorizing dependencies over long time intervals. The core of LSTM lies in its unique gating mechanism and cell states: the input gate controls the extent to which new information from the current time step is written into the cell state; the forget gate controls how much information from the cell state of the previous time step should be forgotten or retained; and the output gate controls how much information from the cell state of the current time step is output to the hidden state. The cell state is the memory unit of the LSTM, running throughout the entire sequence and responsible for storing and transmitting information over long time sequences. The LSTM model receives physiological and biochemical index sequences and metabolomics data sequences from consecutive time points of the patient as input. Through these gating mechanisms, LSTM networks can effectively capture and learn long-term dependencies and complex dynamic patterns in time-series data. For example, they can identify whether specific patterns of blood pressure fluctuations predict myocardial function deterioration, or whether trends in certain metabolite concentrations are related to the progression of myocardial injury. After deep learning processing the sequence data, the LSTM model outputs a dynamic adjustment factor. This dynamic adjustment factor is a continuous value that reflects the dynamic trend of the patient's myocardial injury risk relative to a static assessment, based on the patient's latest time-series data. For example, a positive value indicates an increased risk, a negative value indicates a decreased risk, and a value close to zero indicates a stable risk, thus providing a quantitative basis for the dynamic adjustment of risk levels.

[0114] Furthermore, in S5, the LSTM network internally includes an input gate, a forget gate, an output gate, and a cell state. The input gate controls when and to what extent new information is written into the cell state. The forget gate controls how much information in the cell state of the previous time step is forgotten. The output gate controls how much information in the cell state is output to the hidden state of the current time step. The cell state is responsible for storing and transmitting information in long-term sequences.

[0115] In this embodiment, this step is one of the core innovations of the present invention. It combines static risk assessment with dynamic trends and further predicts molecular-level interactions associated with high-risk states. Specifically, the static risk level output by the GBT model in step S4 is fused with the dynamic adjustment factor output by the LSTM model in step 5, thereby achieving dynamic and precise adjustment of the myocardial injury risk level. The fusion mechanism can be implemented in various ways. For example, the dynamic adjustment factor output by the LSTM model can be used as an independent predictor variable and weighted linearly combined with the predicted probability of the static risk level, or it can be applied as a correction coefficient to the classification boundary of the static risk level. More specifically, when the LSTM model predicts an upward trend in risk, even if the GBT model initially assesses it as medium risk, the dynamically adjusted risk level will be raised to high risk to reflect the progression of the disease; conversely, if the LSTM model predicts a downward trend in risk, the medium risk will be lowered to low risk. This dynamically adjusted risk level can more accurately reflect the actual risk status of myocardial injury at different time points because it considers both the patient's baseline risk and the dynamic changes in recent physiological and biochemical indicators. Based on this dynamic risk assessment, this invention further predicts metabolite target interactions associated with an individual's high-risk state. The prediction method comprises two complementary aspects:

[0116] Firstly, reasoning is based on metabolomics characteristics and a knowledge base. For individuals identified as high-risk, the metabolomics data are first analyzed to identify significantly abnormal metabolites, with concentrations either higher or lower than normal physiological ranges. Subsequently, combined with the metabolite target knowledge base constructed in the second step, the metabolic pathways affected by these abnormal metabolites and related protein targets are identified. This process can reveal which metabolic imbalances directly participate in the pathophysiological process of myocardial injury.

[0117] Secondly, predictions are based on machine learning models. A specialized machine learning model, such as a graph neural network or ensemble learning model, is trained. This model takes a comprehensive feature set of high-risk individuals as input and known metabolite-target interactions as training labels. By learning these complex association patterns, the model can identify which metabolites exhibit abnormal binding, activation, or inhibition of which protein targets under specific physiological and pathological conditions, and even predict new, previously unknown metabolite-target interactions. The prediction results will clearly indicate which metabolites and their corresponding protein targets have undergone significant dynamic changes or abnormal interactions under specific myocardial injury risk states, providing molecular-level target information for subsequent precise interventions.

[0118] Furthermore, in S6, the fusion mechanism is to use the dynamic adjustment factor output by the LSTM model as an independent predictor variable or as a weighting coefficient to correct the static risk level output by the GBT model.

[0119] The reasoning based on metabolomics features and the knowledge base analyzes significantly abnormal metabolites in individual metabolomics data and identifies the metabolic pathways and protein targets affected by them in conjunction with the knowledge base; the prediction based on machine learning models trains graph neural networks or random forest models and uses the individual comprehensive feature set as input to predict existing metabolite-target interactions.

[0120] In this embodiment, this step is the ultimate goal of the method of the present invention, aiming to transform complex prediction results into actionable clinical guidance. Based on the dynamically adjusted myocardial injury risk level in S6 and the predicted metabolite target interaction information, the present invention can generate personalized prevention and intervention strategy recommendations. These recommendations are highly specific and targeted, including but not limited to lifestyle interventions, such as suggesting specific dietary adjustments based on individual metabolite profiles, such as limiting the intake of high-sugar or high-fat foods and increasing dietary fiber; providing customized exercise recommendations, such as weekly aerobic exercise duration and intensity; optimizing drug treatment regimens, such as recommending specific targeted drugs or adjusting the dosage of existing drugs based on predicted abnormal metabolites and protein targets to accurately correct metabolic imbalances or target pathological pathways; and subsequent monitoring protocols, such as suggesting regular testing of specific metabolite or target levels and imaging follow-ups to promptly assess the intervention effect and adjust the strategy. These recommendations directly target the individual's identified risk factors, abnormal metabolic pathways, and perturbed targets, aiming to minimize the incidence or severity of myocardial injury, achieving true precision medicine.

[0121] According to a second embodiment of the present invention, referring to Figure 2 This invention claims protection for a metabolite target interaction prediction system for myocardial injury, comprising:

[0122] Data acquisition and preprocessing module: used to perform multi-source data acquisition and preprocessing functions in step S1;

[0123] Knowledge base construction module: Used to perform the knowledge base construction function in step S2;

[0124] Metabolomics data analysis module: used for spectral processing, metabolite identification and quantification of raw metabolomics data;

[0125] Feature engineering module: Used to perform feature engineering functions of S3;

[0126] Risk Factor Weighting and Selection Module: Used to perform the feature selection function in step S3 and the weighted risk factor model construction function in step S4;

[0127] Risk level classification module: Used to perform the static risk level classification model construction function in step S4;

[0128] Time series data dynamic optimization module: used to perform the time series data dynamic optimization model construction function in step S5;

[0129] Risk assessment and target prediction module: used to perform dynamic adjustment of risk levels and prediction of metabolite-target interactions in step S6; and

[0130] Intervention strategy recommendation module: This module is used to generate personalized intervention strategy recommendations for step S6.

[0131] Furthermore, the data acquisition and preprocessing module is configured with multiple data interfaces, which support DICOM, HL7, mzML, and FASTQ data formats. The data acquisition and preprocessing module internally includes a data cleaning unit, a standardization unit, a missing value imputation unit, an outlier processing unit, and a data transformation unit.

[0132] In this embodiment, the first component is the data acquisition and preprocessing module. This module serves as the system's entry point, equipped with multiple data interfaces to ensure seamless integration with data streams from various medical information systems and biosample analysis devices. These interfaces support multiple industry-standard data formats, such as DICOM for medical imaging, HL7 for clinical information exchange, mzML for mass spectrometry data, and FASTQ for gene sequencing data. Internally, this module integrates a data cleaning unit, a standardization unit, a missing value imputation unit, an outlier handling unit, and a data transformation unit. The data cleaning unit employs rule-based or machine learning-based algorithms to identify and correct data errors, ensuring the accuracy of the original data. The standardization unit uses methods such as Z-score standardization or Min-Max normalization to unify the dimensions of the data, eliminating the influence of different dimensions. The missing value imputation unit employs various strategies to complete missing values. The data transformation unit is responsible for transforming raw heterogeneous data into a unified, structured data format. For example, it can use a pre-trained word embedding model to transform text data into word vectors or parse it into entity relation triples; and it can extract image data into high-dimensional feature vectors through a deep convolutional neural network.

[0133] Secondly, there is the knowledge base construction module. This module is crucial to the system's intelligence and includes a text mining engine, a knowledge graph construction engine, and a knowledge update and management unit. The text mining engine employs advanced deep learning models to achieve Named Entity Recognition (NER) and Relation Extraction (RE) in medical literature. Specifically, the NER model is fine-tuned based on a pre-trained biomedical language model to identify medical entities such as "myocardial infarction," "hypertension," "statins," "ACE inhibitors," "L-carnitine," and "mitochondria." The RE model uses a self-attention mechanism to capture deep semantic relationships between entities; for example, it identifies the "inhibition" relationship between "β-blockers" and "heart rate." The knowledge graph construction engine stores the extracted entities and relationships as a graph structure in the form of nodes and edges, managed using high-performance graph databases capable of efficiently storing and querying complex graph structure data. The knowledge base stores various entities related to myocardial injury and their interrelationships, forming a vast and sophisticated knowledge network. The knowledge update and management unit regularly retrieves incremental information from the latest research literature and public databases, automatically updating and expanding the knowledge base to ensure its timeliness and completeness.

[0134] Next is the metabolomics data analysis module. This module focuses on processing raw mass spectrometry data and is equipped with a raw spectrum processing unit, a metabolite identification unit, and a metabolite quantification unit. The raw spectrum processing unit performs a series of complex operations on LC-MS / GC-MS raw data, including baseline correction, noise filtering, precise peak detection, multi-sample peak alignment, isotope peak deconvolution, and data redundancy removal. The metabolite identification unit uses a standard spectral library and combines various methods such as precise mass matching, fragment ion matching, and retention time matching to perform high-confidence metabolite structure identification on the extracted peaks. The metabolite quantification unit performs relative quantification of the identified metabolites or absolute quantification by adding internal standards.

[0135] Following this is the feature engineering module. This module serves as a bridge between the raw data and the machine learning model, and is equipped with a feature generation unit, a feature fusion unit, and a feature transformation unit. The feature generation unit extracts and generates various features from the outputs of the data acquisition and preprocessing module, the knowledge base construction module, and the metabolomics data analysis module. These features include, but are not limited to, numerical features, categorical features, time-series features, and graph embedding features. The feature fusion unit employs multimodal data fusion technology to integrate features from different modalities into a unified feature vector. Fusion strategies include early fusion, late fusion, or intermediate fusion. The feature transformation unit performs dimensionality reduction processing on the features, such as normalization, standardization, principal component analysis (PCA), or t-SNE, to optimize the model input, reduce redundancy and collinearity among features, and improve the model's training efficiency and generalization ability.

[0136] Furthermore, there is the risk factor weighting and selection module. This module aims to identify and quantify the most important myocardial injury risk factors and is configured with a Recursive Feature Eradication (RFE) engine and an Adaptive Enhancement (AdaBoost) engine. The RFE engine receives the comprehensive feature set output by the feature engineering module and iteratively trains the base model while systematically removing the features with the lowest weights to identify the subset of key features optimal for predicting myocardial injury risk. This process dynamically determines the optimal number of features through cross-validation. The AdaBoost engine receives the subset of key features selected by the RFE engine and trains a series of weak classifiers. By iteratively adjusting the weights of the training samples and the weights of the weak classifiers, the AdaBoost model outputs the contribution weight of each key risk factor. These weights are represented in floating-point form, directly reflecting the relative importance of the factor in predicting myocardial injury risk. The output of this module is a set of key risk factors with clearly defined weight ranking, providing accurate input for subsequent risk level classification.

[0137] Next is the risk level classification module. This module is responsible for the preliminary assessment of the patient's static myocardial injury risk based on weighted key risk factors, and it is configured with a gradient boosting tree (GBT) model. The GBT model receives the weighted key risk factor set output by the risk factor weighting and selection module, as well as other non-temporal features provided by the feature engineering module, as input. The GBT model classifies and predicts the static risk level of myocardial injury by integrating multiple decision trees and using a gradient descent optimization method to minimize prediction error. Its loss function can use multi-class log loss. The classification result is a discrete category label among low risk, medium risk, and high risk. The GBT model outputs the patient's static risk assessment result at the current time point, serving as a benchmark for subsequent dynamic adjustments.

[0138] Following this is the time-series data dynamic optimization module. This module focuses on processing longitudinal patient data to capture dynamic changes in risk, and it is configured with a Long Short-Term Memory (LSTM) network model. The LSTM network model receives individual historical physiological and biochemical data sequences and metabolomics data sequences provided by the data acquisition and preprocessing module. The LSTM network model performs deep learning on the time-series data through its internal input gates, forget gates, output gates, and cell state mechanisms. Its network structure typically contains multiple LSTM layers followed by fully connected layers to capture long-term dependencies and dynamic change patterns. The model learns how to predict future risk trends based on historical data. The LSTM network model outputs a dynamic adjustment factor, a floating-point number between -1 and 1, reflecting the dynamic trend of the patient's myocardial injury risk: a positive value indicates an increased risk, a negative value indicates a decreased risk, and a zero value indicates stable risk.

[0139] Furthermore, there is the risk assessment and target prediction module. This module is the core analysis unit of the entire system, integrating static and dynamic information and delving into the molecular level. This module receives the static risk level output by the risk level classification module and the dynamic adjustment factor output by the time-series data dynamic optimization module. The module applies the dynamic adjustment factor to the static risk level in a weighted fusion manner to generate a dynamically adjusted myocardial injury risk level. For example, the dynamic adjustment factor is mapped to the [0,1] interval using the sigmoid function as a regulation weight, and then fused with the predicted probability of the static risk. Further, the module combines the dynamically adjusted risk level, the metabolite profile provided by the metabolomics data analysis module, and the metabolite target knowledge base provided by the knowledge base construction module to predict specific metabolite target interactions associated with high-risk states. This prediction is accomplished by performing pathway enrichment analysis, network topology analysis, and interaction prediction models based on graph neural networks. For example, GNNs can learn potential connection patterns between metabolite nodes and target nodes, thereby predicting new or perturbed interactions.

[0140] Next is the intervention strategy recommendation module. This module aims to translate complex predictions into actionable clinical recommendations. This module receives dynamic risk levels and predicted metabolite target interaction information from the risk assessment and target prediction module. Internally, this module includes a rule-based expert system and a knowledge graph reasoning engine. The expert system encodes pre-defined clinical guidelines and treatment pathways, such as, "If the patient's risk level is high and an abnormal cholesterol metabolism pathway is predicted, statin therapy is recommended." The knowledge graph reasoning engine utilizes entities and relationships in a knowledge base, combined with the specific risk characteristics of individual patients, to perform logical reasoning and generate more refined and personalized recommendations. These recommendations cover lifestyle interventions, optimization of drug treatment regimens, and follow-up monitoring plans.

[0141] Finally, there are the data storage module and the computing resource management module. The data storage module is configured with a relational database for storing structured clinical data, feature data, and model parameters; a non-relational database for storing semi-structured or unstructured data, such as raw mass spectrometry data and text analysis results; and a graph database for storing knowledge graphs. The databases employ a distributed storage architecture to ensure high availability, scalability, and disaster recovery capabilities. The computing resource management module is responsible for scheduling and allocating the system's computing resources, including CPU resources, GPU resources, and memory resources. This module supports a distributed computing framework to meet the parallel computing needs of large-scale data processing and complex model training, ensuring stable system operation under high concurrency and large data volumes.

[0142] The working principle of this invention is achieved through the close collaboration of the aforementioned modules: First, the data acquisition and preprocessing module integrates and standardizes heterogeneous biomedical data from multiple sources. Simultaneously, the knowledge base construction module utilizes text mining and knowledge graph technologies to construct a structured knowledge base of risk factors and metabolite targets. The metabolomics data analysis module refines the raw mass spectrometry data, extracting metabolomics features. The feature engineering module integrates multimodal data and knowledge base information to generate a unified comprehensive feature set. Subsequently, the risk factor weighting and selection module uses RFE to screen key features and employs the AdaBoost algorithm to calculate the weights of these key risk factors, providing more accurate input for subsequent predictions. The risk level classification module uses the GBT model to perform preliminary static myocardial injury risk level classification for patients. Meanwhile, the time-series data dynamic optimization module uses an LSTM network to perform deep learning on the patient's historical longitudinal data, capturing the dynamic evolution trend of myocardial injury risk and outputting dynamic adjustment factors. The risk assessment and target prediction module integrates static risk assessment with dynamic adjustment factors to generate a more accurate dynamic risk level. Based on this, and combined with individual metabolomics data and a knowledge base, it predicts specific metabolite target interactions associated with high-risk states. Finally, the intervention strategy recommendation module generates personalized prevention and treatment recommendations based on the predicted dynamic risk level and metabolite target interaction information. Through the collaborative work of the various modules in this system, this invention achieves early, dynamic, and accurate prediction of myocardial injury risk and identification of personalized intervention targets, providing comprehensive and in-depth decision support for clinical practice.

[0143] In one specific embodiment, the method for predicting metabolite target interactions for myocardial injury according to this invention was applied to a cohort study comprising 2000 patients, including 1000 patients diagnosed with myocardial injury, 1000 healthy controls, and 500 high-risk individuals such as those with hypertension, diabetes, or hyperlipidemia. Data collection was conducted over a three-year period, with physiological and biochemical indicators and metabolomics data collected quarterly, and comprehensive genomic sequencing and medical imaging examinations performed annually.

[0144] Specifically, during the data acquisition and preprocessing phase, we obtained 20 routine physiological and biochemical indicators from patients, including systolic blood pressure, diastolic blood pressure, fasting blood glucose, total cholesterol, LDL-C, HDL-C, triglycerides, and heart rate. Clinical records were analyzed using a BioBERT-CRF model for named entity recognition, extracting diagnoses for 15 cardiovascular diseases, including angina, heart failure, and coronary atherosclerosis, as well as 30 commonly used cardiovascular drugs such as aspirin, metoprolol, and losartan. Cardiac MRI data from medical imaging was analyzed using a U-Net deep learning model for automated measurement of key indicators such as left ventricular ejection fraction (LVEF) and left ventricular mass (LVM), and texture features such as entropy and contrast of myocardial fibrosis areas were extracted. Genomic data from genomics were analyzed using whole-exome sequencing, identifying 200 SNP loci associated with myocardial injury and quantifying the expression levels of 50 core genes. Raw metabolomics data were acquired using an LC-MS / MS platform. Peak alignment, retention time correction, and isotope deconvolution were performed using Progenesis QI software. 750 plasma metabolites were identified and relatively quantified using the HMDB and METLIN databases.

[0145] In the knowledge base construction module, we extracted over 100,000 entities and 500,000 relationships from PubMed, KEGG, and DrugBank databases using BERT-CRF and Bi-LSTM-Attention models, constructing a knowledge graph encompassing myocardial injury, risk factors, genes, proteins, metabolites, and drugs. This graph is stored in the Neo4j graph database and updated regularly. For example, it explicitly states that "hypertension" is a risk factor for "myocardial hypertrophy," "angiotensin-converting enzyme" (ACE) is a target of "captopril," and "L-carnitine" is involved in the "fatty acid β-oxidation" metabolic pathway.

[0146] The feature engineering module integrates the aforementioned multi-source data. Clinical features include 50 numerical indicators and 10 categorical indicators. Genomic features include 200 SNP loci encoding and 50 gene expression levels. Metabolomics features include 750 metabolite concentrations and activity scores of 20 core metabolic pathways obtained through pathway enrichment analysis using MetaboAnalyst. Temporal features are generated by calculating the mean, standard deviation, and linear trend slope of various physiological and biochemical indicators over the past year, resulting in 100 temporal features. Furthermore, embedding learning is performed on the constructed knowledge graph using a graph-based knowledge network (GCN), generating a 256-dimensional graph embedding feature vector for each patient. This ultimately forms a comprehensive feature set containing 1185 features.

[0147] The risk factor weighting and selection module uses Support Vector Machine (SVM) as the basic model for RFE (Research-Based Factor Evaluation). Through 5-fold cross-validation, 250 key features with the highest predictive power are finally selected. The AdaBoost algorithm weights these 250 features, with LDL-C having a weight of 0.12, a history of hypertension having a weight of 0.09, cardiac troponin T having a weight of 0.15, the specific SNP site rs10046384 having a weight of 0.07, and plasma L-carnitine concentration having a weight of 0.08.

[0148] The risk classification module employs a gradient boosting tree (GBT) model, receiving 250 weighted key features as well as some non-temporal features. This model achieved a classification accuracy of 92.5% on the training set and made preliminary predictions of low, medium, and high static risk levels of myocardial injury in patients on an independent test set.

[0149] The time-series data dynamic optimization module uses a three-layer LSTM network, with each layer containing 128 units. The input sequence length is 12 time steps, representing physiological, biochemical, and metabolomics data from the past 12 quarters. After the LSTM model is trained, it outputs a dynamic adjustment factor that quantifies the real-time changes in a patient's myocardial injury risk. For example, for a patient with a static assessment of intermediate risk, if their LDL-C has been consistently rising for the past two quarters and the concentration of plasma oxidative stress metabolites such as malondialdehyde has significantly increased, the LSTM model will output a positive dynamic adjustment factor, such as +0.35.

[0150] The risk assessment and target prediction module fuses the static risk probabilities output by the GBT model (e.g., 0.6 for medium risk and 0.3 for high risk) with the dynamic adjustment factor output by the LSTM model. The fusion function is set as: Dynamic risk probability = Static risk probability + Dynamic adjustment factor * 0.2. If the adjusted high-risk probability exceeds 0.5, the final risk level is upgraded to high risk. For example, the adjusted high-risk probability for the aforementioned medium-risk patient becomes 0.3 + 0.35 * 0.2 = 0.37, still classified as medium risk. If another patient has a static high-risk probability of 0.4 and a dynamic adjustment factor of +0.8, the adjusted high-risk probability becomes 0.4 + 0.8 * 0.2 = 0.56, and the final risk level is upgraded to high risk.

[0151] Building upon this foundation, metabolomics data were systematically analyzed for individuals dynamically assessed as high-risk. For example, this high-risk patient was found to have significantly elevated plasma free fatty acid (FFA) concentrations and reduced activity in glucose metabolism pathways such as glycolysis. Using a knowledge base, the system inferred that high FFA levels lead to lipotoxicity in cardiomyocytes and affect mitochondrial function through the PPARα pathway. Simultaneously, a GNN model predicted abnormal binding or activity regulation of FFA with CD36 receptor fatty acid transporters and ACS acyl-CoA synthetase. The prediction results clearly indicate that myocardial injury in this patient is closely related to fatty acid metabolism disorders and their abnormal interactions with downstream CD36 / ACS targets.

[0152] Based on the above predictions, the intervention strategy recommendation module generates personalized recommendations for this high-risk patient: Lifestyle interventions are recommended, including strict restriction of saturated and trans fatty acid intake and increasing consumption of foods rich in Omega-3 fatty acids; 150 minutes of moderate-intensity aerobic exercise per week is also recommended. For optimized drug treatment, the module suggests considering the use of PPARα agonists to regulate fatty acid metabolism, combined with cholesterol absorption inhibitors to lower plasma FFA levels. Follow-up monitoring recommends MRI monitoring of plasma FFA and cardiomyocyte lipid deposition every 3 months, and regular re-examination of cardiac troponin levels.

[0153] Comparative example:

[0154] To further illustrate the non-obviousness and technical superiority of this invention, we have designed a comparative example. In this comparative example, we adopt a traditional static risk assessment method, the main features of which are:

[0155] The data sources are singular and the processing is simplified: only routine physiological and biochemical indicators and some clinical text information are used, without in-depth knowledge base construction or multimodal feature fusion. Metabolomics data are only subjected to simple peak area normalization, without metabolite identification or pathway enrichment analysis. Genomics data are not included in the analysis.

[0156] Feature engineering simplification: No extraction of knowledge graph embedding features is performed, no in-depth mining of temporal features is performed, and only features based on statistical indicators are used.

[0157] Model simplification: Instead of constructing a weighted risk factor model, a standard support vector machine (SVM) is used directly for static risk classification, and no dynamic adjustment factor is introduced.

[0158] Metabolite target interaction prediction without molecular-level prediction: Only macroscopic risk levels can be provided.

[0159] Generalized intervention recommendations: Based on general clinical guidelines, provide non-personalized, non-targeted intervention recommendations.

[0160] Specifically, the comparative method was evaluated using the same high-risk population cohort as in the previous embodiment. Regarding data, only 20 physiological and biochemical indicators and 5 major clinical risk factors were extracted from the patients. No detailed text mining or knowledge graph construction was performed. Metabolomics data were directly input into the model after total peak area normalization, without metabolite identification or pathway analysis. In the feature engineering stage, 30 features were directly input into the model. The classification model was trained using a linear support vector machine with optimized parameters via grid search, outputting low, medium, and high-risk classification results. This model was evaluated on an independent test set, and its accuracy, F1 score, and early myocardial injury event recognition rate were significantly lower than those of the previous embodiment. Furthermore, the comparative method cannot provide information on metabolite target interactions, nor can it adjust risk levels based on the patient's dynamic physiological changes; its intervention recommendations are only general "control blood pressure, blood sugar, and blood lipids," lacking specificity.

[0161] During the feature engineering phase, the variation information of specific gene loci involved may include, but is not limited to, the following SNP loci associated with myocardial injury risk:

[0162] rs10757274 (located at locus 9p21);

[0163] rs7412 and rs429358 (together determine APOE ε2 / ε3 / ε4 genotypes);

[0164] rs708272 (located in the CETP gene);

[0165] rs11591147 (located in the PCSK9 gene);

[0166] rs1799752 (located in the ACE gene, I / D polymorphism);

[0167] And specific variations in genes such as AGT and NOS3 at sites closely associated with other cardiovascular risk factors such as hypertension and diabetes.

[0168] The genotypic information of these loci is extracted and encoded using bioinformatics processes and incorporated as part of a subsequent comprehensive feature set. Those skilled in the art can update and expand the list of specific gene loci based on the latest research findings and public databases.

[0169] Table 1 compares the key performance indicators of the embodiments of the present invention and the comparative examples in terms of myocardial injury prediction performance and the accuracy of intervention strategies. The data are from the same independent validation set, which includes 300 patients, of whom 50 experienced myocardial injury events during the follow-up period.

[0170] Table 1 Evaluation Comparison Table

[0171]

[0172] As clearly shown in the data tables above, the embodiments of the present invention significantly outperform the traditional static SVM method in terms of prediction accuracy, F1 score, and early identification rate for myocardial injury events. Particularly in terms of early identification rate, the embodiments of the present invention can identify 85% of myocardial injury events three months in advance, compared to only 40% in the comparative method. This fully demonstrates the advantages of the present invention in time-series data analysis, multimodal feature fusion, and knowledge graph utilization. Furthermore, the present invention can provide refined prediction results of metabolite target interactions and highly personalized intervention plans, and simulates a higher average risk reduction rate after intervention, which directly translates into superior clinical application value. Although the computational resource consumption of the present invention during the model training phase is relatively high, the significant improvement in predictive performance and accuracy of clinical decision support fully demonstrates its technological advancement and practical value.

[0173] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0174] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0175] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A metabolite target interaction prediction method for myocardial injury, characterized by, The method comprises the following steps: S1, obtaining individual physiological and biochemical data, clinical diagnosis and treatment records, medical image data, genomics data and metabolomics raw data, and performing cleaning, format unification, standardization, missing value filling, outlier processing and noise reduction operation on the collected heterogeneous data; S2, constructing a myocardial injury risk factor and metabolite target knowledge base, extracting risk factors, biomarkers, metabolic pathways, protein targets and known drug-target interaction information related to myocardial injury through automatic text mining and knowledge graph construction technology, and forming a structured knowledge network; S3, fusing the multi-source data preprocessed in S1 and the knowledge base information constructed in S2 to generate a comprehensive feature set containing clinical risk features, genomics features, metabolomics features, time series features and graph structure embedded features extracted from the knowledge base; and using a recursive feature elimination based feature selection method to select a key feature subset; S4, using an adaptive boosting AdaBoost learning algorithm to calculate the weights of the key risk factors selected in S3 to construct a weighted risk factor model, and inputting the weights output by the weighted risk factor model and the key feature subset into a gradient boosting tree GBT model to construct a static risk level classification model, and the GBT model outputs the myocardial injury risk level of the patient at the current time point; S5, using a long short-term memory LSTM model to perform time series analysis on the individual historical physiological and biochemical data and metabolomics data preprocessed in S1 to construct a time series data dynamic optimization model, and the LSTM model outputs a dynamic adjustment factor; S6, fusing the static risk level output by the GBT model in S4 and the dynamic adjustment factor output by the LSTM model in S5, predicting the metabolite target interaction related to the individual high-risk state through reasoning based on metabolomics features and the knowledge base and / or prediction based on a machine learning model, and generating personalized prevention and intervention strategy suggestions according to the dynamically adjusted myocardial injury risk level and the predicted metabolite target interaction.

2. The method of claim 1, wherein, In S1, the physiological and biochemical data includes numerical data of blood pressure, blood glucose, blood lipids and heart rate; The clinical diagnosis and treatment records include text data of medical history, diagnosis, medication and treatment plan; The medical image data includes echocardiogram, magnetic resonance imaging MRI and computed tomography CT image data; The genomics data includes single nucleotide polymorphism SNP and gene expression profile sequence data; and the metabolomics raw data is obtained by liquid chromatography-mass spectrometry LC-MS or gas chromatography-mass spectrometry GC-MS technology.

3. The method of claim 2, wherein, In S1, the text data is processed by natural language processing technology for word segmentation, part-of-speech tagging, named entity recognition and conversion into structured information; The image data is processed by image processing algorithm for feature extraction; The sequence data is genotyped and variation recognized; The metabolomics raw data is processed by peak extraction, peak alignment, retention time correction, isotope deconvolution and peak area normalization, and a standard metabolite library is used for preliminary identification and quantification of metabolites.

4. The method of claim 1, wherein, In S2, the automatic text mining uses a named entity recognition (NER) model based on deep learning to identify entities and a relation extraction (RE) model to identify semantic relationships between entities, thereby constructing triples. The knowledge graph represents the extracted entities and their relationships in the form of nodes and edges, forming a structured knowledge network. The knowledge base integrates known physical interactions between metabolites and myocardial-related protein targets, signal pathway regulation relationships, enzymatic reaction relationships, and gene expression regulation relationships.

5. The method of claim 1, wherein, In S3, the comprehensive feature set includes: Clinical risk features, including physiological and biochemical indicators, demographic information, and structured risk factors extracted from clinical diagnosis and treatment records; Genomics features, including variation information at specific genetic loci and gene expression levels related to myocardial injury; Metabolomics features, including the relative concentrations of identified metabolites, metabolite ratios, and specific metabolic pathway activity scores obtained through pathway enrichment analysis; Temporal features, which are extracted by differentiating and moving average processing of historical physiological and biochemical data and metabolomics data to capture trend and volatility characteristics; Graph structure embedding features extracted from the knowledge base, which are learned through graph neural networks.

6. The method of claim 1, wherein, In S4, the AdaBoost learning algorithm iteratively trains weak classifiers, adjusts the weights of training samples at each iteration, and assigns weights to weak classifiers based on their classification performance. The final strong classifier is composed of the weighted votes of all weak classifiers. The GBT model constructs multiple decision trees sequentially, with each tree attempting to correct the prediction error of the previous tree and optimizing the loss function using gradient descent during training.

7. The method of claim 1, wherein, In S5, the LSTM network contains input gates, forget gates, output gates, and cell states. The input gate controls when and how much new information is written to the cell state. The forget gate controls how much information from the previous time step is forgotten. The output gate controls how much information in the cell state is output to the hidden state at the current time step. The cell state is responsible for preserving and transmitting information over long time series.

8. The method of claim 1, wherein, In S6, the fusion mechanism modifies the static risk level output by the GBT model by using the dynamic adjustment factor output by the LSTM model as an independent prediction variable or as a weight coefficient. The reasoning based on metabolomics features and the knowledge base identifies the metabolic pathways and protein targets affected by significantly abnormal metabolites in individual metabolomics data, combined with the knowledge base. The prediction based on machine learning models trains graph neural networks or random forest models to predict the existence of metabolite-target interactions using the individual comprehensive feature set as input.

9. A metabolite target interaction prediction system for myocardial injury, characterized by, It includes: Data acquisition and preprocessing module: used to perform the multi-source data acquisition and preprocessing function of step S1 in claim 1; Knowledge base construction module: used for performing the knowledge base construction function of step S2 in claim 1; Metabolomics data analysis module: used for performing the spectrum processing, metabolite identification and quantification of metabolomics raw data; Feature engineering module: used for performing the feature engineering function of S3 in claim 1; Risk factor weighting and selection module: used for performing the feature selection function of S3 and the weighted risk factor model construction function of step S4 in claim 1; Risk level classification module: used for performing the static risk level classification model construction function of step S4 in claim 1; Time series data dynamic optimization module: used for performing the time series data dynamic optimization model construction function of step S5 in claim 1; Risk assessment and target prediction module: used for performing the risk level dynamic adjustment and metabolite target interaction prediction function of step S6 in claim 1; and Intervention strategy suggestion module: used for performing the individualized intervention strategy suggestion generation function of step S6 in claim 1.

10. The system of claim 9, wherein, The data acquisition and preprocessing module is configured with a plurality of data interfaces supporting DICOM, HL7, mzML and FASTQ data formats; and the data acquisition and preprocessing module internally comprises a data cleaning unit, a standardization unit, a missing value filling unit, an outlier processing unit and a data conversion unit.

Images