Metabolic age assessment system based on plasma metabolic markers and storage medium

Abstract

This invention provides a system and storage medium for assessing metabolic age based on plasma metabolic biomarkers. The system includes: a data acquisition module, a preprocessing and batch correction module, a feature screening module, a stacked integrated modeling module, a GAM nonlinear correction module, a validation module, and a testing module. This invention can screen a set of metabolic biomarkers from plasma that systematically cover the core pathways of metabolic aging, focus on the current state of metabolic physiology, and have been validated by a large-scale cohort throughout the entire life cycle. These biomarkers can be used for in vitro data-assisted assessment of plasma metabolic age and for health reference stratification.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical fields of metabolomics, mass spectrometry analysis, artificial intelligence-assisted health assessment, in vitro detection and chronic disease risk assessment, and particularly relates to a metabolic age assessment system and storage medium based on plasma metabolic markers. Background Technology

[0002] With the increasing aging of the global population, the burden of age-related chronic diseases (cardiovascular disease, diabetes, malignant tumors, neurodegenerative diseases, etc.) is becoming increasingly heavy. Currently, the clinical assessment of an individual's aging status mainly relies on chronological age (cAge), supplemented by traditional biochemical indicators such as blood glucose, blood lipids, and liver and kidney function. However, chronological age cannot reflect the significant heterogeneity of biological aging among individuals—individuals of the same age can have biological aging processes that differ by more than ten years due to differences in genetic background, lifestyle, and environmental exposure. Summary of the Invention

[0003] The purpose of this invention is to provide a metabolic age assessment system and storage medium based on plasma metabolic markers.

[0004] To address the above problems, this invention provides a metabolic age assessment system based on plasma metabolic biomarkers, comprising: The data acquisition module is used to collect plasma samples and time-series age (cAge) of the sample subjects; based on the collected plasma samples, the corresponding raw mass spectrometry matrix is ​​obtained. The preprocessing and batch calibration module is used to obtain standardized calibration parameters and the corresponding standardized metabolic feature matrix based on the original mass spectrometry matrix. The feature selection module is used to obtain a list of plasma metabolic biomarkers and a corresponding quantitative feature subset matrix of metabolic biomarkers based on the standardized and corrected metabolic feature matrix and the corresponding time-series age cAge. The stacked ensemble modeling module is used to obtain a trained and parameter-frozen base learner model based on the quantitative feature subset matrix of metabolic biomarkers and the time-series age cAge of the corresponding healthy training set. After training convergence, the optimal fusion weights and initial predicted ages of each base learner are obtained. The GAM nonlinear correction module is used to obtain the fitted GAM mapping function g(cAge), the corresponding corrected metabolic age, and the metabolic age acceleration based on the initial predicted age. The validation module is used to validate the effectiveness of the metabolic age assessment system composed of the solidified components. The solidified components include: standardized calibration parameters, a list of plasma metabolic biomarkers, a base learner model that has been trained and whose parameters are frozen, the optimal fusion weights corresponding to each base learner obtained after training convergence, and the fitted GAM mapping function g(cAge). The testing module is used to acquire the sample to be tested; based on the valid metabolic age assessment system, it obtains the corresponding corrected metabolic age and metabolic age acceleration of the sample to be tested.

[0005] The beneficial effects of the present invention are as follows: 1) Solid foundation for large-scale validation and clinical translation: The biomarker combination was validated in a large cohort of 18,476 participants (aged 1–96 years, 6 centers) through feature screening using three algorithms and a 5.5-year longitudinal follow-up, and has sufficient statistical power.

[0006] 2) Full life cycle coverage: The biomarkers showed a significant correlation with age from children (1 year old) to the very old (96 years old), and can be used to assess metabolic aging in people of all ages.

[0007] 3) Clear functional positioning: The 30 metabolites focus on core pathways of amino acids / central carbon / renal function / sulfur-containing compounds that reflect the "current status of metabolic physiology", forming a clear complement to systemic mortality risk markers (which are separately applied for protection by the applicant), making the biological interpretation of mAge / mAA more accurate.

[0008] 4) High accuracy: The stacked ensemble model significantly outperforms any single base learner on the healthy training set and the independent healthy validation set, demonstrating the robustness of multi-model sparse fusion in mass spectrometry metabolic data scenarios.

[0009] 6) Minimize age group bias: GAM correction significantly reduces systematic bias, especially for systematic overestimation bias in the 70-year-old age group (average overestimation of 4.8 years before correction → 1.2 years after correction in the 70–80 age group; 7.1 years before correction → 2.3 years after correction in the 80-year-old age group).

[0010] 7) Significant incremental predictive value: mAAA outperformed most of the 27 routine clinical biochemical indicators in predicting all-cause mortality; adding mAAA to any single biochemical indicator-adjusted model significantly improved the C-index for all-cause mortality prediction (median ΔC-index = +0.025, all LRTs P < 10). -6 This demonstrates that the combination of markers has predictive information independent of existing indicators.

[0011] 8) Dual cohort validation: mAA demonstrated stable metabolic age assessment performance in both healthy individuals (n=10708) and diseased individuals (n=7768); the outcome risk assessment was primarily validated in baseline healthy individuals with 5.5 years of follow-up data.

[0012] 9) Robustness across batches and centers: The architecture maintains stable performance on independent healthy validation sets and data from different acquisition centers and different mass spectrometry batches; as a supplement, it also achieves good performance on UK Biobank proteomics data, NHANES routine biochemical indicators and gut microbiome data, indicating that the fusion mechanism has the potential for expansion.

[0013] 10) Convenient and high-throughput: When used with the MALDI-TOF-MS platform, only 100 nL of plasma is required, and a single sample can be tested in about 30 seconds, supporting large-scale screening deployment.

[0014] 11) The mechanism is explainable: 30 metabolites cover multiple known aging-related core pathways, including amino acid metabolism, central carbon metabolism, kidney function and muscle metabolism, and sulfur-containing compound metabolism, with a clear biological basis.

[0015] 12) Modular system implementation: This method can be encapsulated as a system, electronic device and computer-readable storage medium, which is suitable for deployment in scenarios such as physical examination centers, research institutions and health management platforms, and supports batch high-throughput metabolic age assessment. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a two-layer stacked integrated architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a two-layer stacked integrated architecture according to an embodiment of the present invention, showing 12 base learners and a 5-fold cross-validation out-of-fold generation method, as well as the weighted fusion process of Lasso meta-learners; Figure 3 This is a scatter plot and residual distribution of mAge and cAge in a healthy training set (n=5,252) and an independent healthy validation set (n=5,456) according to an embodiment of the present invention, showing the effect of deviation improvement before and after GAM correction. Figure 4 This is a schematic diagram of the progressive ablation experiment results of a base learner based on SHAP value according to an embodiment of the present invention, showing the relationship between the number of learners and the RMSE of the validation set, and verifying the necessity of the stacked integrated architecture. Figure 5 This is a schematic diagram of a mAA three-class stratification system according to an embodiment of the present invention, showing the mAA distribution of the healthy cohort and the Q1 / Q3 dividing line; Figure 6This is an example ROC result graph of a panel of 30 markers in mAAA three-class classification according to an embodiment of the present invention; Figure 7 This is a 5.5-year cumulative morbidity / mortality risk curve (Kaplan–Meier) of the mAA acceleration / deceleration group according to an embodiment of the present invention, covering all-cause mortality (n=291), new-onset fatty liver (n=405), new-onset hypertension (n=323), new-onset diabetes (n=196), new-onset coronary heart disease (n=180), new-onset stroke (n=137), new-onset hyperlipidemia (n=137), and new-onset hyperuricemia (n=64) (all P<0.001); Figure 8 This is a forest plot of the association between mAA and lifestyle / exposure factors according to an embodiment of the present invention, showing the regression coefficients (95% CI) of mAA with factors such as smoking, secondhand smoke, hair dye, shift work, dietary structure (whole grains / fried / processed meat / seafood, etc.), exercise participation, and family history of cardiovascular disease. Figure 9 This is a comparison chart of mAA with comprehensive clinical health indicators (LE8 cardiovascular health score, frailty index) and conventional blood biochemical indicators in all-cause mortality prediction according to an embodiment of the present invention, demonstrating the incremental predictive value of mAA in multiple baseline models. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] The concept of a biological aging clock can be introduced into aging research: multidimensional biomarkers based on DNA methylation (epigenetic clock), proteomics (protein clock), and metabolomics have been used to construct biological age models. Among them, metabolomics has unique advantages because it directly reflects the end products of gene-environment interactions, can sensitively capture the effects of exposures such as diet, exercise, and disease on the body's physiological state, and its samples are easily obtained (peripheral plasma). Research on metabolomics aging biomarkers has validated their association with mortality risk and chronic disease incidence in several large-scale population cohort studies.

[0019] Existing metabolomics-based aging biomarkers and methods for assessing metabolic age have the following main shortcomings: (1) Lack of panel design for biomarkers that specifically reflect the “current status of metabolic physiology”: Existing schemes often mix biomarkers of “physiological status” and “systemic risk”, which limits the interpretability of metabolic age output; (2) Lack of standardized output schemes directly coupled with test kits: Only a single metabolite or a single statistical result is given, making it difficult to form a deployable standardized product; (3) Insufficient throughput of detection platforms: Most studies rely on traditional LC-MS / MS or NMR platforms, which are complex to process and time-consuming to detect a single sample, making it difficult to support large-scale population cohort studies. (4) Batch effects and center effects were not systematically corrected: Significant technical variations exist in multicenter and multi-batch data, affecting the cross-center repeatability of the model; (5) Mass spectrometry metabolic features are high-dimensional, sparse, and redundant: The original mass spectrometry metabolic features are numerous (hundreds to thousands), with a large number of low-abundance, missing, and highly correlated variables. The single feature screening method has poor stability and is prone to overfitting. (6) Single models are insufficient for characterizing complex interactions: Existing models mostly use LASSO linear regression or a single machine learning algorithm, which makes it difficult to simultaneously characterize the linear relationship, local neighborhood structure and complex nonlinear interaction modes between mass spectrometry metabolites. (7) Age-dependent systematic bias has not been eliminated: The existing metabolic age clock has significant systematic overestimation or underestimation biases in different age groups (especially in the elderly over 70 years old), which affects the accuracy of health assessment results; (8) The detection throughput, reagent kit composition, quality control and result interpretation templates for large-scale population screening scenarios are still not clear enough, which limits the translation and application.

[0020] The aforementioned problems have led to challenges in the application of existing metabolic age clocks in health management and scientific research, including insufficient accuracy, poor reproducibility, limited applicable populations, and inadequate productization. There is an urgent need to develop a new metabolic aging assessment solution that is high-throughput, low-bias, robust across batches, and can be deployed as a standardized reagent kit.

[0021] This invention aims to provide a complete technical solution for assessing metabolic physiological status based on plasma metabolomics, with specific objectives including: 1) Provide a set of plasma metabolic biomarkers for auxiliary assessment of systemic aging and chronic disease risk. The set focuses on core metabolic pathways that reflect the current metabolic physiology, such as amino acid metabolism, central carbon metabolism, renal function-related metabolites, and sulfur-containing compounds, and includes a total of 30 metabolites. 2) A method for assessing metabolic age based on stacked ensemble learning and nonlinear calibration for mass spectrometry metabolic data (especially MALDI-TOF-MS data) is provided. A two-layer stacked ensemble architecture with 12 base learners across 5 major machine learning families is constructed, and a generalized additive model (GAM) nonlinear bias correction is superimposed to output interpretable metabolic age (mAge) and metabolic age acceleration (mAA) indices, achieving high accuracy, low systematic bias and batch robustness. 3) Provide a test kit containing the above-mentioned biomarker combination and supporting calculation module, so that it can output mAge, mAA and three-category results of accelerated, normal and decelerated, and use the results for auxiliary stratification of all-cause mortality and multiple chronic disease risks; 4) Provide systems, electronic devices, and computer-readable storage media corresponding to the above methods to achieve standardized deployment from metabolic data and time-series age input, feature screening, stacked ensemble modeling, GAM correction, mAA stratification to result output.

[0022] like Figure 1 As shown, this invention provides a metabolic age assessment system based on plasma metabolic biomarkers, which can be a metabolic age assessment system based on stacked ensemble learning and GAM nonlinear calibration. The system may include: a data acquisition module, a preprocessing and batch correction module, a feature selection module, a stacked ensemble modeling module (12 base learners + Lasso meta-learner), a GAM nonlinear correction module, a mAA hierarchical evaluation module, a validation module, a detection module, and a comprehensive report output module. The system can perform the following steps: Step S1: The data acquisition module is used to collect plasma samples from the subjects. Each subject has a corresponding collection center number, time-series age (cAge), and sample queue grouping label. Mass spectrometry is performed on the plasma samples of each subject to obtain the corresponding raw mass spectrometry matrix. The sample queue grouping label includes: a healthy training set, an independent healthy validation set, and a disease validation set. Simultaneously, the four types of metadata corresponding to each row of samples in the raw mass spectrometry matrix—namely, the collection center number, mass spectrometry target plate number, sample queue grouping label, and time-series age (cAge)—are bound row by row to the raw mass spectrometry matrix to generate a corresponding metadata table.

[0023] Specifically, step S1 may include: Step S1.1, Subject screening and sample baseline metadata entry: 1) Inclusion and exclusion criteria (refer to Example 1 below) Inclusion criteria: Healthy individuals without diabetes, cardiovascular or cerebrovascular diseases, chronic kidney disease, malignant tumors, COPD, or other chronic diseases; disease cohorts included any of the above-mentioned chronic diseases; age range: 1–96 years. Exclusion criteria: acute infection within 3 weeks prior to sampling, pregnancy / lactation, and use of metabolic interference drugs within the past 3 months; 2) Baseline metadata synchronization and entry operation: Before blood is collected from subjects for enrollment, three types of static metadata are simultaneously entered and fixed in the clinical cohort database: a. Corresponding to the numbers of 6 data acquisition centers; b. Subject's chronological age cAge, i.e., actual chronological age; c. Pre-assigned sample queue group labels, including: healthy training set, independent healthy validation set, and disease validation set; Three types of static metadata can be permanently bound to the subject's unique sample ID and transferred synchronously with the plasma sample to the mass spectrometry detection stage.

[0024] Here, regarding the pre-assigned sample queue grouping labels, the fixed partitioning rule for the sample subject dataset could be, for example, the total healthy sample p=10708, and the disease validation set q=7768; The total healthy sample p is divided into: a healthy training set n=5252 and an independent healthy validation set m=5456. The disease validation set q is used only for external model validation, and the disease cohort q does not participate in feature selection or model training; Step S1.2, Standardized collection, aliquoting, and cryopreservation of fasting plasma: Blood collection guidelines: Subjects should fast for ≥8 hours, and 1 mL of EDTA-anticoagulated peripheral venous whole blood should be collected. Centrifugation: Within 2 hours of blood collection, centrifuge at 4℃ and 2000×g for 10 minutes to separate the supernatant plasma. Aliquoting and storage: Plasma is aliquoted into cryovials and frozen at -80°C away from light, avoiding repeated freeze-thaw cycles; Sample ID binding: Each plasma cryopreservation tube is printed with a unique sample ID, which is associated with the center number, cAge, and queue group label entered in S1-1 to form a sample traceability link.

[0025] Step S1.3: Record sample dilution, target plate spotting, and target plate numbering before MALDI-TOF mass spectrometry: Sample dilution: Thaw 100 nL of plasma and dilute it with deionized water with a resistivity ≥18.2 MΩ·cm at a volume ratio of 1:10 and mix well; Target plate spotting procedure: Add 500 nL of diluted plasma to a MALDI metal target plate and air dry at room temperature; then cover with 500 nL of iron-based nanoparticle matrix solution with a concentration of 1 mg / mL and a particle size of 5–15 nm, and dry again before collection; Mass spectrometry target plate number synchronous archiving: Each target plate is assigned a unique digital mass spectrometry target plate number before the instrument is set up. The sample ID of all spotted samples on the target plate is uniformly registered with the corresponding target plate number and written into the mass spectrometry operation log. The target plate number is bound to the sample ID and is exported synchronously with the original spectrum file, serving as a permanent metadata attachment to the spectrum matrix.

[0026] Step S1.4, MALDI-TOF mass spectrometry acquisition, synchronously outputting the raw mass spectrum matrix and a complete set of bound metadata: Mass spectrometry fixed acquisition parameters (refer to Example 2): positive ion reflectance mode, 355nm laser, 1000Hz repetition rate, 20kV accelerating voltage, 250ns delayed extraction; 2000 laser strikes per sample; mass-to-charge ratio acquisition range m / z 100–500Da; only peaks with SNR > 3 are retained for subsequent analysis; Quality control specifications: Each target plate adopts a nine-square grid QC mixed sample layout, with intra-batch QC peak intensity RSD < 20% and inter-batch quality control CV < 25%; Data Synchronization and Packaging Output: After the mass spectrometer completes acquisition, it automatically outputs the raw mass spectrum matrix; at the same time, it binds the four types of metadata (6 acquisition center numbers, mass spectrometer target plate number, queue group label, and time-series age cAge) corresponding to each row of samples to the spectrum matrix row by row, generating a matching metadata table. The two are combined as the complete output data package S1, which can be imported into step S2 later.

[0027] Here, fasting peripheral venous plasma is collected from subjects, preferably through matrix-assisted collection of plasma metabolic profiles, requiring only a nanoliter sample volume (100 nL), with a single sample collection time of approximately 30 seconds, achieving high-throughput metabolomics detection; for example, fasting peripheral venous plasma can be collected from subjects, and mass spectrometry detection can be performed using an iron-based nanoparticle matrix MALDI-TOF-MS platform to output the raw mass spectrometry matrix; this invention can also be adapted to metabolic feature matrices output by other mass spectrometry platforms, after extraction, alignment, and normalization of similar peaks.

[0028] The raw mass spectrometry matrix obtained in step S1 contains the raw data of m / z peak intensities of plasma samples from all subjects.

[0029] Specifically, the row index of the original mass spectrometry matrix is: sample index, with each row corresponding to all mass spectrometry detection signals of one subject's plasma; the column index of the original mass spectrometry matrix is: mass spectrometry m / z characteristic peak index, with each column corresponding to a fixed mass-to-charge ratio metabolic peak. Each row of the original mass spectrometry matrix represents a single plasma sample from a single subject; therefore, the complete set of metadata attributes bound to this row is fixed. For the same row (same sample), the acquisition center number, mass spectrometry target plate number, queue group label, and time-series age cAge remain constant; among them, The collection center number indicates that each sample comes from only one clinical collection center, and there is only one center code per line; multiple centers will not appear. The mass spectrometry target plate number indicates that one plasma sample is spotted on only one MALDI target plate for detection, and each row corresponds to only one target plate number. The cohort grouping label indicates that the subjects were pre-assigned to a healthy training set, an independent healthy validation set, or a disease validation set upon enrollment, with each row corresponding to only one grouping label. Time-series age cAge: The age of the subjects in the same sample is unique, with only one age value in a single row.

[0030] Step S2, the preprocessing and batch correction module, is used to perform mass spectrometry data preprocessing, hierarchical batch correction, sign-preserving logarithmic transformation and normalization scaling on the original mass spectrometry spectrum matrix of the healthy training set samples output in step S1, and to fit and solidify the normalization correction parameters, including: peak processing parameters corresponding to mass spectrometry data preprocessing, reference batch correction rules corresponding to hierarchical batch correction, reference distribution parameters corresponding to reference distribution mapping and normalization scaling parameters corresponding to normalization scaling, so as to obtain the normalized and corrected metabolic feature matrix corresponding to the healthy training set. Preferably, the preprocessing and batch correction module is used to perform mass spectrometry data preprocessing, reference batch correction, reference distribution mapping, and normalization scaling on the raw mass spectrometry matrix of the healthy training set samples. Specifically, peak preprocessing is performed on the raw mass spectrometry matrix of the healthy training set samples to obtain a coarsely extracted metabolic feature matrix; based on the mass spectrometry target plate number, acquisition center number, and quality control sample information that are matched row by row with the coarsely extracted metabolic feature matrix, a hierarchical empirical Bayesian batch correction strategy is adopted to establish reference batch correction rules at the target plate level and the acquisition center level in sequence, thereby obtaining the batch-corrected feature matrix corresponding to the healthy training set; a sign-preserving logarithmic transformation is performed on the batch-corrected feature matrix, and the mean μ and standard deviation σ of each metabolic feature are calculated based on the healthy training set to form a set of normalization scaling parameters, thereby obtaining the normalized and corrected metabolic feature matrix corresponding to the healthy training set.

[0031] Subsequent validation or testing modules can perform batch anchoring correction and reference distribution alignment on independent health validation sets, disease validation sets, or samples to be tested, based on synchronously detected quality control samples, internal standard samples, or reference samples. Then, they can be standardized using the standardized scaling parameters obtained from the health training set to generate the standardized and corrected metabolic feature matrix of the corresponding sample set.

[0032] Here, subsequent independent health validation sets, disease validation sets, and external samples to be tested can undergo a consistent preprocessing procedure based on the fixed peak processing parameters, reference batch correction rules, reference distribution parameters, and standardized scaling parameters. Specifically, for new target plates, new batches, or new collection centers that did not appear during the training phase, batch anchoring correction is performed based on the synchronously detected quality control samples, internal standard samples, or reference samples, and then mapped to the reference distribution of the health training set, ultimately obtaining the standardized and corrected metabolic feature matrix of each group of samples.

[0033] Peak detection, peak alignment, baseline correction, and noise filtering (preserving peaks with SNR>3) were performed on the original mass spectrometry matrix. A hierarchical empirical Bayesian batch correction strategy (ComBat algorithm) was adopted, first correcting the experimental batch level (mass spectrometry target plate), and then correcting the acquisition center level (6 centers), systematically eliminating non-biological technical variations. Sign-preserving logarithmic transformation x′ =sign(x) × log(1 + |x|) and StandardScaler normalization were performed to adapt to common skewed distributions, sparsity intensities, and batch differences in mass spectrometry data.

[0034] Step S2.1, perform peak preprocessing, including: performing Gaussian smoothing (σ=1), TopHat baseline correction, SNR>3 peak detection, multi-sample peak alignment (m / z tolerance 0.01Da) on the original spectra of the healthy training set samples, filtering invalid low signal-to-noise ratio peaks, and outputting a coarsely extracted metabolic feature matrix; Here, a unified peak processing threshold and calculation parameters are used in this step (Gaussian smoothing σ=1, SNR threshold 3, m / z alignment tolerance 0.01Da, TopHat baseline subtraction coefficient, peak alignment offset). This set of parameters is fixed for peak preprocessing and is solidified throughout the process, so all samples can be reused uniformly.

[0035] Step S2.2 includes: Step S2.2.1: Obtain the mass spectrometry target plate number that matches the coarse extracted metabolic feature matrix output in step S2.1 row by row; call the pycombat single-layer ComBat empirical Bayesian correction function and quality control samples, using the mass spectrometry target plate number as the unique batch stratification variable, perform deviation fitting on the peak intensity of each column of the coarse extracted metabolic feature matrix in the healthy training set, solve and solidify the target plate stratification reference batch correction rule; and based on the solidified target plate stratification reference batch correction rule, obtain the intermediate feature matrix after target plate correction; the target plate stratification reference batch correction rule is used to constrain the target plate batch anchoring correction of subsequent samples to reduce the technical variation caused by the differences in spotting, laser intensity and target plate substrate between different target plates within the same center. Here, subsequent independent health validation sets, disease validation sets, and external samples to be tested can undergo a consistent preprocessing procedure based on the fixed peak processing parameters, reference batch correction rules, reference distribution parameters, and standardized scaling parameters. Specifically, for new target plates, new batches, or new collection centers that did not appear during the training phase, batch anchoring correction is performed based on the synchronously detected quality control samples, internal standard samples, or reference samples, and then mapped to the reference distribution of the health training set, ultimately obtaining the standardized and corrected metabolic feature matrix of each group of samples.

[0036] The mass spectrometer target plate number is recorded synchronously in step S1, during mass spectrometry spotting and acquisition, and stored in the same mass spectrometer output file as the original mass spectrometry spectrum matrix.

[0037] Subsequent independent health validation sets, disease validation sets, and external samples to be tested do not retrain the metabolic age prediction model; for differences in target plate level technology, the target plate level reference batch correction rules formed based on the health training set are combined with the quality control samples, internal standard samples, or reference samples in the current testing batch to complete the target plate level batch anchoring correction.

[0038] Step S2.2.2: Obtain the acquisition center number that matches the intermediate feature matrix after target plate calibration row by row; call the pycombat single-layer ComBat calibration function again, using the acquisition center number as the global hierarchical batch variable, and independently perform empirical Bayes bias fitting on each column of metabolic features of the target plate calibrated intermediate feature matrix of the healthy training set output in step S2.2.1 to solve the acquisition center hierarchical reference batch calibration rule; based on the acquisition center hierarchical reference batch calibration rule, obtain the center-calibrated feature matrix; wherein, the acquisition center hierarchical reference batch calibration rule is used to constrain the center hierarchical batch anchoring calibration of subsequent samples to reduce the global technical variation caused by differences in instrument models, experimental operations and reagent consumables across acquisition centers; record the target plate hierarchical reference batch calibration rule and the acquisition center hierarchical reference batch calibration rule as the reference batch calibration rule.

[0039] Here, a two-level stratified ComBat batch effect correction is performed. The first level of correction uses the mass spectrometer target plate number as the batch variable to correct for technical variations in a single batch of target plates. The second level of correction uses the numbers of the six acquisition centers as the batch variables to correct for differences in instruments and operations across multiple centers. The correction tool is the pycombat package, which provides a stratified correction algorithm and outputs a correction feature matrix that eliminates batch / center variations. The mass spectrometer target plate number is recorded synchronously in step S1, during mass spectrometry spotting and acquisition, and stored in the same mass spectrometer output file as the original mass spectrometry spectrum matrix.

[0040] Independent health validation sets, disease validation sets, and external samples to be tested are processed based on the collection center-level reference batch correction rules formed by the health training set. When the sample to be tested comes from a collection center already included in the training phase, it is corrected according to the corresponding center reference distribution. When the sample to be tested comes from a new center not included in the training phase, the center-level batch anchoring and reference distribution alignment are completed based on the quality control samples, internal standard samples, or reference samples detected simultaneously.

[0041] Preferably, the preprocessing and batch correction module is further configured to, when performing mass spectrometry data correction on an independent healthy validation set, a disease validation set or a to-be-detected sample, synchronously collect supporting reference samples, including: matching the independent healthy validation set with a corresponding quality control sample, matching the disease validation set with a corresponding internal standard sample, and matching the to-be-detected sample with a corresponding reference sample; the subsequent validation module or test module performs anchoring correction and distribution alignment on the metabolic feature intensity of the current detection batch based on the reference batch correction rule and reference distribution parameters formed by the healthy training set, and maps the corrected metabolic features to the reference distribution of the healthy training set; then perform sign-preserving logarithmic transformation and standardized scaling processing to respectively obtain the standardized and corrected metabolic feature matrix of the independent healthy validation set and the standardized and corrected metabolic feature matrix of the disease validation set; wherein the reference distribution parameters are global statistical reference parameters fitted and solidified based on the healthy training set, including: the standardized mean μ and standard deviation σ corresponding to each metabolic feature, the ComBat empirical Bayesian prior distribution hyperparameters at the target plate level and the collection center level, and the signal reference distribution parameters of the quality control samples supporting the healthy training set; the subsequent validation module or test module reuses the reference distribution parameters to complete the distribution alignment of the metabolic features of the current batch, and uniformly maps the corrected metabolic features to the reference reference distribution of the healthy training set. The preprocessing and batch correction module is further configured to, when performing mass spectrometry data correction on an independent healthy validation set, a disease validation set or a to-be-detected sample, synchronously collect supporting reference samples, including: matching the independent healthy validation set with a corresponding quality control sample, matching the disease validation set with a corresponding internal standard sample, and matching the to-be-detected sample with a corresponding reference sample; the subsequent validation module or test module performs anchoring correction and distribution alignment on the metabolic feature intensity of the current detection batch based on the reference batch correction rule and reference distribution parameters formed by the healthy training set, and maps the corrected metabolic features to the reference distribution of the healthy training set; then perform sign-preserving logarithmic transformation and standardized scaling processing to respectively obtain the standardized and corrected metabolic feature matrix of the independent healthy validation set and the standardized and corrected metabolic feature matrix of the disease validation set; wherein the reference distribution parameters are global statistical reference parameters fitted and solidified based on the healthy training set, including: the standardized mean μ and standard deviation σ corresponding to each metabolic feature, the ComBat empirical Bayesian prior distribution hyperparameters at the target plate level and the collection center level, and the signal reference distribution parameters of the quality control samples supporting the healthy training set; the subsequent validation module or test module reuses the reference distribution parameters to complete the distribution alignment of the metabolic features of the current batch, and uniformly maps the corrected metabolic features to the reference reference distribution of the healthy training set.

[0042] Here, the independent health validation set, disease validation set, and external samples to be tested are processed based on the collection center-level reference batch correction rules formed by the health training set; when the sample to be tested comes from a collection center already included in the training phase, it is corrected according to the corresponding center reference distribution; when the sample to be tested comes from a new center not included in the training phase, the center-level batch anchoring and reference distribution alignment are completed based on the quality control sample, internal standard sample, or reference sample detected simultaneously.

[0043] Step S2.3 includes: Step S2.3.1, Perform sign-preserving logarithmic transformation, including: substituting each value x in the center-corrected feature matrix corresponding to the healthy training set in step S2.2.2 into the fixed formula: x′ = sign (x) × log (1 + |x|) to calculate, and obtain the intermediate feature matrix after logarithmic transformation, where sign (x) preserves the original intensity sign, and log (1 + |x|) compresses the mass spectrometry skewness and large-scale sparse intensity to adapt to the distribution characteristics of mass spectrometry data; Here, the logarithmic transformation compensation constant in this step is fixed at 1, which is a fixed parameter that can be used uniformly for all subsequent samples without recalculation.

[0044] Step S2.3.2, perform standardization scaling, including: StandardScaler standardization processing, scaling all feature values ​​to a mean of 0 and a variance of 1. Interval: Divide the sample rows corresponding to the healthy training set from the intermediate feature matrix after the logarithmic transformation in step S2.3.1, group them by feature columns, and calculate the mean μ and standard deviation σ for each column to form a fixed set of standardization scaling parameters such as {μ1,σ1,μ2,σ2,…,μ n ,σ n Based on the mean μ and standard deviation σ of each column, substitute each element into the formula column by column. After conversion, the standardized and corrected metabolic feature matrix corresponding to the healthy training set is obtained; Here, the output product of step S2 is a standardized and corrected metabolic feature matrix containing only the quantitative features of the 30 biomarkers of this invention.

[0045] The mass spectrometry target plate number is the unique identifier of the mass spectrometry metal target plate carrying the diluted plasma sample and matrix in the MALDI-TOF-MS detection step S1. It belongs to the supporting metadata recorded synchronously in the mass spectrometry detection step S1, and is collected from the same source as the original mass spectrometry matrix, without any variables created out of nothing. The collection center number is a unique numerical code for the 6 clinical sample collection centers. It is the queue metadata that is archived synchronously during the subject enrollment phase and is matched one-to-one with the coarsely extracted metabolic feature matrix. Each row of samples corresponds to a unique center number. The pycombat hierarchical correction algorithm is a hierarchical iterative correction procedure based on the ComBat empirical Bayesian correction framework. The hierarchical logic is to first correct local batch bias and then correct multi-center global bias.

[0046] Subsequently, based on the fixed standardized scaling parameters obtained from fitting the healthy training set, the standardized conversion formula can be uniformly substituted into each column and element of the healthy training set, independent healthy validation set, disease validation set, and external samples to be tested, and finally the standardized and corrected metabolic feature matrix of each group of samples can be obtained.

[0047] The preprocessing parameters fitted and solidified in step S2 include three categories: peak processing fixed parameters, reference batch correction rules and reference distribution parameters, and normalized scaling parameters. The peak processing fixed parameters and normalized scaling parameters are determined based on the healthy training set; the reference batch correction rules are established based on the healthy training set and its quality control samples, target plate numbers, and acquisition center numbers. Subsequent independent healthy validation sets, disease validation sets, and samples to be tested do not require retraining of the feature selection model, stacked ensemble model, and GAM correction model, but they must undergo batch anchoring correction based on synchronous quality control samples, internal standard samples, or reference samples before being mapped to the reference distribution of the healthy training set using the aforementioned normalized scaling parameters.

[0048] Two-level layered ComBat batch calibration parameters: target plate layer, acquisition center layer corresponding to each layer batch offset, batch scaling factor, baseline correction item; Standardized scaling parameters: Specific mean μ and standard deviation σ for each metabolic feature column in the healthy training set; Peak preprocessing, sign-preserving logarithmic transform fixed parameters: Gaussian smoothing σ=1, SNR peak filtering threshold 3, m / z peak alignment offset tolerance 0.01Da, TopHat baseline subtraction coefficient, logarithmic transform compensation constant 1.

[0049] Step S3, the feature screening module, is used for joint feature screening and functional classification using three algorithms: Based on the standardized and corrected metabolic feature matrix and the corresponding time-series age cAge corresponding to the healthy training set, joint feature screening using three algorithms is performed to lock in a list of 30 plasma metabolic biomarkers; Based on the list of plasma metabolic biomarkers, the corresponding feature columns are extracted from the standardized and corrected metabolic feature matrix of the healthy training set obtained in step S2 to generate a subset matrix of quantitative features of metabolic biomarkers in the healthy training set, which can be used as input for the stacked ensemble modeling in the subsequent step S4; Here, the same list is reused for subsequent independent health validation sets, disease validation sets, and external samples to be tested to extract corresponding feature columns and generate their respective corresponding metabolic biomarker quantitative feature subset matrices. The output standardized and corrected metabolic feature matrix and the corresponding time-series age cAge in step S2 are used as inputs. The intersection and union of three feature screening algorithms, namely random forest, Lasso regression and mutual information, are combined to screen and obtain a fixed list of 30 plasma metabolic biomarkers. Preferably, the list of 30 plasma metabolic markers includes: L-cysteine, leucine, glyceric acid, 4-hydroxyproline, L-asparagine, serine, imidazole-4-acetaldehyde, creatinine, O-phosphothreonine, itaconic acid, betaine, 2-aminobenzoic acid, DL-O-phosphoserine, pyroglutamic acid, L-arginine, 3-hydroxyisovaleric acid, L-histidine, and L-glutamic acid. The following acids are listed: taurine, selenocystine, 2,3-dihydroxybutanoic acid, acetoacetic acid, lactic acid, succinic acid, ribonic acid, phosphate hydroxypyruvic acid, propionylcarnitine, L-carnitine, thiocyanate, and 2-hydroxyethanesulfonate.

[0050] The 30 plasma metabolic biomarkers focus on core metabolic pathways that reflect the current state of metabolic physiology, covering: 1) Amino acid metabolism and derivatives (sulfur-containing amino acids, branched-chain amino acids, aromatic amino acids, hydroxylated amino acids, phosphorylated amino acids, etc., a total of 17 kinds); 2) Central carbon metabolism (glycolysis intermediates, TCA cycle-related organic acids, ketone bodies, a total of 8 types); 3) Renal function and muscle metabolism markers (creatinine, carnitine, and acylcarnitine, a total of 3); 4) Sulfur-containing compounds (thiocyanates and sulfonates, a total of 2 types).

[0051] A complete list of 30 plasma metabolic markers is shown in Table 1, including their Chinese and English names, molecular formulas, and HMDB numbers.

[0052] Table 1

[0053] Step S3.1: From the standardized and corrected metabolic feature matrix output in step S2.3, based on the healthy training set in the queue grouping label in step S1, filter out the sample row feature submatrix corresponding to the healthy training set. Step S3.2: Call the configured RandomForestRegressor model with the first fixed hyperparameters, take the sample row feature submatrix corresponding to the healthy training set as the input of the independent variable, and use the time-series age cAge supervision vector corresponding to the healthy training set as the prediction target to complete the model training; after training, output the global importance score of each column of metabolic features in the sample row feature submatrix corresponding to the healthy training set through the built-in algorithm of the model; sort all metabolic features from high to low according to the importance score, and extract the top 30% of metabolic features according to the sorting position to obtain the RF initial feature set; Here, the independent variable matrix is: the healthy training set sample rows within the standardized and corrected metabolic feature matrix; The supervision label vector is: the time-series age cAge corresponding to the healthy training set; The first fixed hyperparameter is: RandomForestRegressor (n_estimators=100, random_state=42).

[0054] The importance score for random forest regression is: RandomForestRegressor, n_estimators=100, random_state=42. The features are extracted, ranked by importance, and the top 30% of features are selected. Step S3.3: Call the configured LassoCV model with the second fixed hyperparameter, take the sample row feature submatrix corresponding to the healthy training set as the input of independent variables, take the time-series age cAge supervision vector corresponding to the healthy training set as the prediction target, and complete the model training using 5-fold cross-validation; after training, extract the regression coefficients of metabolic features in each column of the sample row feature submatrix of the healthy training set, calculate the absolute value of the coefficients, remove invalid metabolic features with an absolute value of 0, sort the remaining metabolic features in descending order of absolute value of coefficients, and select the top 30% of metabolic features to obtain the Lasso initial feature set; Here, the second fixed hyperparameter is: LassoCV (cv=5, random_state=42).

[0055] The absolute value score of the Lasso regression coefficients is: LassoCV, 5-fold cross-validation, random_state=42, selecting the top 30% of features with non-zero coefficients. Step S3.4: Set a fixed random seed random_state=42 as the parameter of the mutual_info_regression function, and calculate the mutual information regression score between each metabolic feature in the feature submatrix of the sample row corresponding to the healthy training set and the time-series age cAge supervision vector corresponding to the healthy training set column by column; sort all metabolic features according to the mutual information score from high to low, and extract the top 30% of metabolic features to obtain the initial mutual information feature set.

[0056] The fixed random seed is: random_state=42.

[0057] The mutual information regression score is: mutual_info_regression, random_state=42, and the top 30% of features with mutual information scores are selected; In steps S3.1 to S3.4, using time-series age cAge as the supervised variable, the following calculations are performed in the healthy training set: RF: Random Forest Feature Importance (RandomForestRegressor, n_estimators=100, random_state=42). Lasso: Absolute values ​​of Lasso regression coefficients (LassoCV, 5-fold cross-validation, random_state=42); MI: Mutual information regression score (mutual_info_regression, random_state=42).

[0058] Step S3.5, intersection and union selection rules: Let RF, Lasso, and mutual information initial feature set be RF, Lasso, and MI, respectively; successively calculate the pairwise feature intersections RF∩Lasso, RF∩MI, and Lasso∩MI to obtain three sets of common feature subsets; then take the union of the three intersections, and select and retain 30 stable age-related metabolic candidates from the union to form a fixed list of 30 metabolic biomarkers. Based on this list, extract the corresponding columns to generate a quantitative feature subset matrix of metabolic biomarkers.

[0059] Here, each method independently selects the top 30% of features, calculates three pairwise intersections (RF∩Lasso, RF∩MI, Lasso∩MI), and takes the union of the three intersections to obtain 30 robust age-related metabolic candidates (amino acids / central carbon / renal function / sulfur-containing compounds), which serve as the biomarker panel of this invention.

[0060] The behavior of the quantitative feature subset matrix of metabolic biomarkers: Sample (all subjects in the healthy training set); The columns are as follows: 30 columns, each column uniquely corresponding to the standardized quantitative value of one fixed metabolic biomarker of the present invention; The study targeted 30 metabolites to cover four core pathways reflecting metabolic homeostasis, without contamination by disease-specific biomarkers. Step S4, stacked ensemble modeling module, used for training the two-layer stacked ensemble model, outputs the initial predicted age pred_age: The inputs are: the biomarker quantitative feature subset matrix from step S3 and the time-series age cAge of the corresponding healthy training set; The output is the initial predicted age pred_age corresponding to the healthy training set, which serves as the input for subsequent S5.

[0061] The first layer contains 12 base learners across 5 major machine learning families, and uses 5-fold cross-validation to generate out-of-fold predictions; The second layer uses Lasso as the meta-learner to perform sparse weighted fusion of the outputs of the 12 base learners to obtain the initial predicted age pred_age; Step S4 involves two-layer stacked ensemble modeling to calculate the initial predicted age pred_age: This step uses the feature submatrix formed by the healthy training set sample rows in the metabolic biomarker quantitative feature subset matrix output from S3, and the time-series age cAge supervision vector of the healthy training set corresponding to each row of this submatrix as input to construct a first-layer model containing 12 base learners across 5 major machine learning classes; an out-of-fold prediction matrix is ​​generated through 5-fold cross-validation, and then the initial predicted age pred_age is obtained and output using a Lasso meta-learner for weighted fusion; The 12 base learners are: Linear Regression (LR), Lasso Regression, Ridge Regression, ElasticNet, Support Vector Machine (SVM), KNN (Distance Weighted Nearest Neighbor), XGBoost, CatBoost, LightGBM, ExtraTrees, Random Forest (RF), and a four-layer Neural Network (MLP).

[0062] Step S4.1, Fixed architecture of the first layer of 12 base learners: Construct 12 base learners across 5 major machine learning families, and preset fixed hyperparameters for the 12 base learners. The 12 base learners include: linear models (LR, Lasso, Ridge, ElasticNet), support vector machines (SVM, where C=10, rbf kernel), nearest neighbor models (KNN, where n_neighbors=3, distance weighted), tree ensemble models (XGBoost, CatBoost, LightGBM, ExtraTrees, RF), and neural networks (MLP, 4 hidden layers, 128 / 256 / 512 / 1024 neurons); Based on the biomarker quantitative feature subset matrix of the healthy training set obtained in step S3 and the corresponding time-series age cAge of the healthy training set obtained in step S1, call the 12 base learners with fixed hyperparameters respectively, and use KFold (n_splits=5, shuffle=True, random_state=42) to execute 5 on the healthy training set. Folded cross-validation: Within each fold, the corresponding base learner is fitted using the training subset within the fold. After all 5 folds are completed, 12 fully trained and parameter-frozen base learner models are retained for forward inference on the validation set, disease validation set, and external samples to be tested. Simultaneously, each base learner generates a one-dimensional prediction vector based on the out-of-fold validation samples. The out-of-fold prediction vectors corresponding to the 12 base learners are horizontally concatenated column-wise to generate an out-of-fold prediction matrix with 12 columns and a number of rows equal to n (the total number of samples in the healthy training set). This out-of-fold prediction mechanism can avoid leakage of training labels.

[0063] Here, 5-fold cross-validation (KFold, n_splits=5, shuffle=True, random_state=42) is performed on the healthy training set to generate out-of-fold prediction matrices (shape: n × 12) for each base learner, preventing label leakage.

[0064] The hyperparameters of the 12 base learners and the coefficients of the meta-learners are shown in Table 2.

[0065] Table 2

[0066] Step S4.2, Sparse Weighted Fusion of the Second-Layer Lasso Meta-Learner (Input: n×12-dimensional prediction matrix from Step S4.1): Using the n×12-dimensional out-of-fold prediction matrix output from S4.1 as the input feature matrix, and combining it with the time-series age cAge supervision vector corresponding to the archived and row-by-row matched healthy training set from Step S1 as the fitting label, the meta-learner is trained using a fixed-configuration LassoCV (5-fold CV, random_state=42). After training convergence, the optimal fusion weights corresponding to each of the 12 base learners are obtained. During meta-learner training, the prediction components corresponding to base learners with low contribution or strong noise interference are assigned weights close to 0. This characteristic achieves noise suppression in subsequent weighting operations. A linear weighted summation operation is performed, including: taking the entire n×12-dimensional out-of-fold... The prediction matrix is ​​generated by multiplying each predicted component in each column of the matrix by its corresponding optimal fusion weight, and then summing all the product terms in a single sample row to output the initial predicted age pred_age corresponding to the healthy training sample.

[0067] Using the out-of-fold prediction matrices of 12 base learners as the input feature matrix, a Lasso meta-learner (LassoCV, 5-fold CV, random_state=42) is trained, outputting the optimal fusion weights of each base learner, automatically suppressing model components with large noise contributions, and obtaining the initial predicted age pred_age corresponding to healthy training samples.

[0068] Train the LassoCV meta-learner (5-fold cross-validation, random_state=42), and automatically output the prediction results of all base learners corresponding to the optimal fusion weights of the 12 base learners in the healthy training set; linearly weighted fusion of all base learners.

[0069] Step S5, GAM nonlinear correction module, used for GAM nonlinear systematic bias correction, calculates mAge and mAA: In the healthy training set, the generalized additive model pred_age = g(cAge) + ε is fitted using the pygam library (LinearGAM) to obtain the fitted GAM mapping function g(cAge), where ε is the model random error term (residual); Based on the fitted g(cAge), the systematic bias term b(cAge) is extracted: b(cAge) = g(cAge) - cAge; The systematic bias term b(cAge) is used to correct the initial predicted age, and the corrected metabolic age mAge corresponding to the healthy training set is: mAge = pred_age - b(cAge) = pred_age - g(cAge) + cAge; The difference between the corrected metabolic age and the time-series age is used to calculate the metabolic age acceleration mAA corresponding to the healthy training set: mAA = mAge - cAge = pred_age - g(cAge).

[0070] Here, the input for step S5 is: the initial predicted age pred_age corresponding to the healthy training sample output in step S4 and the time-series age cAge corresponding to the healthy training sample in step S1. The output of step S5 is: the corrected mAge and mAA dual indicators, used for mAA stratification and downstream input of the entire risk assessment process); the mAA output by S5 can be used as an input indicator for subsequent steps such as metabolic aging stratification, all-cause mortality risk assessment, chronic disease incidence risk stratification, and quantitative evaluation of lifestyle interventions.

[0071] Step S6, the validation module, is used to perform forward inference based on independent health validation sets and disease validation sets, and to calculate the validation set mAge and mAA (model generalization performance validation): The input for step S6 is: Step S2: Fit and solidify the complete set of correction coefficients and hierarchical ComBat correction parameters for the standardized correction process; The list of 30 locked metabolic biomarkers in step S3; Step S4.1 The 12 base learners are trained and their parameters are frozen; Step S4.2 After training converges, the optimal fusion weights corresponding to each of the 12 base learners are obtained; Step S5: Fit the completed GAM mapping function g(cAge); Step S1 outputs the original mass spectrometry matrix of the independent healthy validation set and the original mass spectrometry matrix of the disease validation set, along with the corresponding time-series age cAge.

[0072] The processing flow is as follows: Step S6.1: Based on the peak processing parameters in step S2, peak preprocessing is performed on the original mass spectrometry matrix of the independent health validation set and the original mass spectrometry matrix of the disease validation set, respectively. Based on the quality control samples matched to the independent health validation set and the internal standard samples matched to the disease validation set, and based on the reference batch correction rules and reference distribution parameters formed by the health training set, the metabolic feature intensity of the independent health validation set and the disease validation set is first anchored and corrected, and then distribution alignment is performed based on the reference distribution parameters. The distribution alignment process is achieved through numerical translation and scaling. Finally, all anchored and corrected metabolic features are uniformly mapped to the reference distribution of the health training set to obtain the mapped metabolic features. The mapped metabolic features are subjected to sign-preserving logarithmic transformation, and then standardized scaling is performed based on the fixed standardized scaling parameters to generate the standardized and corrected metabolic feature matrix of the independent health validation set and the standardized and corrected metabolic feature matrix of the disease validation set. Step S6.2: Based on the locked list of 30 metabolic biomarkers from Step S3, extract 30 columns of features from the standardized and corrected metabolic feature matrix of the independent healthy validation set and the standardized and corrected metabolic feature matrix of the disease validation set, respectively, according to the quantitative feature subset matrix of the 30 biomarkers. The order of the extracted feature columns is strictly aligned with that of the healthy training set, resulting in a dimension of n. val The feature submatrix of the independent health validation set and the feature submatrix of the disease validation set are both 30 × 10; n val This refers to the total number of samples in the independent health validation set or the total number of samples in the disease validation set.

[0073] Step S6.3: Input the feature sub-matrix of the independent health validation set and the feature sub-matrix of the disease validation set into the 12 base learners trained and with parameters frozen in S4.1; for each validation set in the independent health validation set and the disease validation set, each base learner outputs a one-dimensional vector predicting the age of all samples in that validation set. Concatenate the prediction vectors output by the 12 base learners in the same validation set horizontally along the columns to obtain the independent health validation set with dimension n. val The prediction matrix is ​​×12, and the disease validation set has an n-dimensional dimension. val A prediction matrix of ×12; Step S6.4: After reusing the convergence of S4.2, the optimal fusion weights corresponding to the 12 base learners are obtained. Linear weighted summation is performed on the prediction matrix of the independent health validation set and the prediction matrix of the disease validation set to obtain the initial predicted age pred_age of the independent health validation set and the initial predicted age pred_age of the disease validation set. Step S6.5: Substitute the time-series age cAge corresponding to each sample in the independent health validation set and the disease validation set into the GAM mapping function g(cAge) fitted and permanently fixed based on the health training set in step S5, and calculate the system bias term b(cAge) = g(cAge) - cAge; solve the corresponding intermediate correction amount pred_age - g(cAge) based on the initial predicted age pred_age of the independent health validation set and the initial predicted age pred_age of the disease validation set, respectively. Step S6.6: Based on the intermediate correction value output from S6.5 and the time-series age cAge of each sample, substitute them into the following formulas to solve for the corrected metabolic age mAge and metabolic age acceleration mA for each sample in the independent healthy validation set and the disease validation set: mAge = pred_age - b(cAge) = pred_age - g(cAge) + cAge; mAA: mAA = mAge - cAge = pred_age - g(cAge) .

[0074] Step S6.7: Based on the corrected metabolic age mAge and metabolic age acceleration mAA of the independent health validation set and disease validation set, evaluate the age prediction generalization accuracy of the metabolic age assessment system composed of solidified components, analyze the distribution of metabolic aging characteristics in the disease population, and complete the effectiveness validation of the metabolic age assessment system.

[0075] Here, the fixed components of the metabolic age assessment system include: Mass spectrometry preprocessing and parameter system (step S2): Two-stage stratified ComBat batch calibration parameters, standardization of μ, σ, peak preprocessing and logarithmic transformation to fix thresholds for each feature column; Fixed biomarker screening rules and a list of 30 metabolic biomarkers (step S3): a set of fixed features obtained by joint screening using RF, Lasso, and mutual information; The stacked ensemble prediction model (step S4 stacked two-layer model) includes: first layer: a base learner with 12 parameters frozen; second layer: a Lasso meta-learner and 12 fixed optimal fusion weights; GAM nonlinear bias correction function (step S5): GAM mapping function g(cAge) fitted and solidified only by the healthy training set.

[0076] Specifically, based on the mAge of the independent healthy validation set samples and the known true time-series age cAge, the prediction error indices R2, MAE, and RMSE can be calculated, and the prediction stability can be verified by age group to quantitatively evaluate the generalization accuracy of the model's age prediction. Based on the mAA of the disease validation set samples and the baseline mAA of the independent healthy validation set, inter-group statistical tests, distribution characteristic comparisons, and statistics on the proportion of abnormally aged individuals can be carried out to analyze the distribution pattern of accelerated metabolic aging in the disease population.

[0077] If the age prediction error on the independent healthy validation set meets the preset accuracy standard, and the disease validation set shows a significant positive mAA shift compared to the healthy baseline, then the model demonstrates that it possesses both stable age generalization prediction ability and biological efficacy in distinguishing pathological metabolic aging, thus completing the effectiveness verification of the entire method.

[0078] Step S7, the testing module, based on the validated components of the metabolic age assessment system, obtains inference from the plasma sample of the subject to be tested and outputs the clinical test mAge and mAl. The input for step S7 is: Step S2: Fit and solidify the complete set of correction coefficients and hierarchical ComBat correction parameters for the standardized correction process; Step S3: List of 30 metabolic biomarkers; Step S4.1 The 12 base learners are trained and their parameters are frozen; Step S4.2 After training converges, the optimal fusion weights corresponding to each of the 12 base learners are obtained; Step S5: Fit the completed GAM mapping function g(cAge).

[0079] The processing flow for step S7 is as follows: Step S7.1: Based on the peak processing parameters obtained from fitting the healthy training set in step S2, peak preprocessing is performed sequentially on the original mass spectrometry matrix of the sample to be tested; based on the reference batch correction rules and reference distribution parameters formed by matching the reference sample to the reference sample with the reference sample from the healthy training set, the metabolic feature intensity of the sample to be tested is first anchored and corrected, and then distribution alignment is performed based on the reference distribution parameters; the distribution alignment process is achieved through numerical translation and scaling, and finally all anchored and corrected metabolic features are uniformly mapped to the reference distribution of the healthy training set to obtain the mapped metabolic features; a sign-preserving logarithmic transformation is performed on the mapped metabolic features, and then normalization scaling is performed based on the fixed normalization scaling parameters to generate the normalized and corrected metabolic feature matrix of the sample to be tested; Step S7.2: Based on the list of 30 metabolic biomarkers locked in step S3, extract 30 columns of features from the standardized and corrected metabolic feature matrix of the test samples. The order of the extracted feature columns is strictly aligned with the order of the feature columns in the healthy training set, resulting in a dimension of n. test A feature submatrix of the sample to be detected, which is 30 × 10⁻³⁰. Where n test This represents the number of samples to be tested.

[0080] Step S7.3: Input the feature sub-matrices of the samples to be detected into the 12 base learners that have been trained and whose parameters are frozen. Each base learner outputs a one-dimensional vector predicting the age of all samples to be detected. Concatenate the prediction vectors output by the 12 base learners horizontally along their columns to obtain a vector of dimension n. test The prediction matrix of the sample to be detected is 12 × 12.

[0081] Step S7.4 reuses the optimal fusion weights from step S4.2 to perform a linear weighted summation to calculate the initial predicted age pred_age of the sample to be detected: Based on the 12 optimal fusion weights that are permanently fixed after training convergence in step S4.2, a linear weighted summation operation is performed row by row on the prediction matrix of the sample to be detected output in S7.3 to obtain the initial predicted age pred_age of the sample to be detected.

[0082] Step S7.5: Substitute the temporal age cAge of each sample to be tested into the solidified GAM mapping function g(cAge) to calculate the system bias term b(cAge) = g(cAge) - cAge; and solve the correction intermediate quantity pred_age - g(cAge) based on the initial predicted age pred_age of the sample to be tested. Step S7.6: Based on the intermediate correction value of the sample to be tested output in S7.5 and the time-series age cAge of each sample, substitute them into the following formulas to solve for the corrected metabolic age mAge and metabolic age acceleration mAge corresponding to the sample to be tested: mAge = pred_age - b(cAge) = pred_age - g(cAge) + cAge; mAA: mAA = mAge - cAge = pred_age - g(cAge) .

[0083] Preferably, the present invention further includes step S8, which includes: The mAA stratification assessment module is used to determine the stratification threshold corresponding to the healthy training set based on metabolic age acceleration mAA; and to obtain metabolic aging stratification labels based on the stratification threshold; the metabolic aging stratification labels include: accelerated metabolic aging, normal metabolic age, and decelerated metabolic aging.

[0084] Here, based on the metabolic age acceleration (mAA) output in step S8, the stratification thresholds can be determined according to the mAA distribution of the healthy reference cohort (n=5252): Q1 = -1.687 years, Q3 = 3.152 years; accordingly, individuals are divided into: 1) Accelerated metabolic aging (mAA>3.152 years): Located above the upper quartile of the mAA distribution in healthy individuals, indicating that metabolic aging progresses significantly faster than in healthy individuals of the same age, and is associated with an increased risk of chronic disease onset and death; 2) Normal metabolic age (-1.687 ≤ mAA ≤ 3.152): in the middle 50%, the rate of metabolic aging is within the normal range; 3) Slower metabolic aging (mAA < -1.687 years): Located below the lower quartile, it indicates that the metabolic aging process is significantly slower than that of the same age group, and has a better health prognosis.

[0085] According to another aspect of the present invention, an integrated MALDI-TOF-MS detection kit is also provided. The kit includes: A. Plasma sample pretreatment reagents; B. Standards, isotope internal standards, or calibrators corresponding to 30 metabolites; C. Quality control materials (QC pool); D. The matrix, mobile phase, or derivatization reagent required for mass spectrometry analysis (in a preferred embodiment, it is an iron-based inorganic nanoparticle matrix with a concentration of 1 mg / mL, adapted to the MALDI-TOF-MS detection platform). E. Results calculation module (software / storage medium, with built-in mAge prediction model and GAM correction parameters based on stacked ensemble learning). F. Report generation module (automatically outputs mAge, mAA values, hierarchical levels, and auxiliary assessments); G. Instruction manual.

[0086] The stratification criteria for the kit are: accelerated metabolic aging (mAA>3.152 years), normal metabolic age (-1.687 ≤mAA ≤ 3.152), and decelerated metabolic aging (mAA<-1.687 years).

[0087] The application of plasma metabolic biomarker-based metabolic age assessment systems and kits in the preparation of the following products can be as follows: (1) All-cause mortality risk auxiliary assessment product: used for non-diagnostic in vitro risk assessment of an individual's all-cause mortality risk within 5.5 years; among which, mAAA, as an independent risk indicator, is significantly associated with all-cause mortality (HR = 1.50 / 1SD, 95% CI 1.34–1.69, P = 6.14 × 10⁻⁶). -12 The mean α (C-index = 0.7351) showed better predictive ability than comprehensive clinical health indicators such as the frailty index and the LE8 cardiovascular health score. Even after incorporating baseline models built from 27 routine clinical biochemical indicators, including blood glucose, blood lipids, and liver and kidney function, mAAA maintained independent predictive ability (HR range 1.50–1.53, all P < 0.001). Furthermore, after adjusting for all routine blood biochemical indicators, the association between mAAA and all-cause mortality remained robust (HR = 1.55, 95% CI 1.38–1.75, P = 3.03 × 10⁻⁶). -13 Furthermore, after adding mAAA to the baseline model, the C-index increased from 0.7402 to 0.7499 (Likelihood ratio test LRT P = 1.23 × 10⁻⁶). -9 The median ΔC-index = +0.025 proves that mAA has predictive information independent of existing indicators.

[0088] (2) Chronic disease incidence risk prediction product: used to predict the risk of new incidence of the following chronic diseases (validated by 5.5 years prospective follow-up): fatty liver (n=405), hypertension (n=323), diabetes (n=196), coronary heart disease (n=180), stroke (n=137), hyperlipidemia (n=137), hyperuricemia / gout (n=64), malignant tumors, etc.; the cumulative incidence risk of the above chronic diseases was significantly increased in the accelerated group vs. the decelerated group (log-rank all P<0.001); it can also be extended to atrial fibrillation, heart failure and other outcomes.

[0089] (3) Lifestyle and Early Risk Management Products: Based on the quantitative association between mAAA and modifiable lifestyle factors, these products are used for individualized lifestyle assessment and health management recommendations. Specific associations include (adjusted for age and gender in a large-scale health cohort): Risk factors for accelerated metabolic aging: smoking (β = +1.19 years, 95% CI 0.91–1.48, P<0.001), secondhand smoke exposure (β = +0.64 years, P<0.001), hair dye use (β = +0.99 years, P<0.001), shift work history (β = +0.92 years, P = 0.001), intake of sugary foods, animal organs, fried foods, processed meats and seafood (all P<0.05), and family history of premature cardiovascular disease in one's father or mother (all P<0.01). Protective factors for slowing down metabolic aging: intake of whole grains (β = -0.34 years, P = 0.002) and regular physical exercise (β = -0.48 years, P<0.001).

[0090] The above results support mAA as a quantitative indicator for monitoring the effects of early, reversible lifestyle interventions, and it can be used in the development of precision health management products.

[0091] Figure 1 This is a schematic diagram of a two-layer stacked integrated architecture. The left side is the sample input area, the middle is the core area for aging clock modeling with 12 base learners stacked in two layers, and the right side is the hierarchical visualization area for correction effect and metabolic age acceleration.

[0092] Figure 2 A schematic diagram of a two-layer stacked integrated architecture is shown, illustrating the generation method of 12 base learners and 5-fold cross-validation out-of-fold, as well as the weighted fusion process of Lasso meta-learners; (See attached diagram) Figure 2 The complete closed-loop chain is fully demonstrated, showing the classification of the first layer of 12 base learners, the logic of generating the prediction matrix through 5-fold cross-validation out-of-fold, and the weighted fusion output of the second layer Lasso meta-learner to produce pred_age. The 12 base learners are arranged in five categories: linear model, SVM, KNN, tree ensemble, and neural network. Arrows clearly indicate the direction of data flow: the feature matrix of 30 markers is input to each base learner, the prediction vector is output in sections, the matrix is ​​concatenated into an n×12 prediction matrix, input to the Lasso meta-learner, and the output is pred_age. All data flow links match step S4.

[0093] Figure 3The scatter plots and residual distributions of mAge and cAge in the healthy training set (n=5,252) and the independent healthy validation set (n=5,456) demonstrate the improvement effect of GAM correction on the bias. The scatter plots and residual distributions of mAge and cAge before and after GAM correction in the healthy training set and independent healthy validation set are shown. The plots are divided into two main sections: the left side represents the healthy training set (n=5252), and the right side represents the independent healthy validation set (n=5456). Each section contains two sets of comparative data: an uncorrected linear fitting curve and a GAM-corrected fitting curve. Complete fitting indices R, R², RMSE, MAE, and MAPE are labeled. The residual distribution curve below the plot visually shows the positive bias of the elderly population before correction and the uniform distribution of residuals after correction, quantifying the technical effect of the GAM module in step S5 in eliminating age-related systematic bias. All numerical parameters are completely consistent with the correction effect data in Example 5.

[0094] Figure 4 The results of progressive ablation experiments on base learners based on SHAP values ​​are presented to demonstrate the relationship between the number of learners and the RMSE of the validation set, thus verifying the necessity of the stacked integration architecture. Figure 4 A schematic diagram showing the results of the progressive ablation experiment, ranking the contributions of the 12 base learners based on SHAP values; Figure 3 The inner bar chart arranges the models KNN, RF, XGBoost, CatBoost, and LightGBM in descending order of SHAP contribution, and marks the fixed SHAP weight values ​​of each model; the line chart shows the trend of RMSE change on the validation set after gradually removing low-contribution learners, proving that the complete 12-model stacked architecture can achieve the lowest prediction error, supporting the creative design of the 12-base learner architecture of this invention, and corresponding to the SHAP ablation verification content of feature 5.

[0095] Figure 5 A schematic diagram of the mAA three-category stratification system is provided, showing the mAA distribution of the healthy cohort and the Q1 and Q3 dividing lines. Figure 5 The horizontal axis represents the mAA value, and the vertical axis represents the population density distribution curve. Two fixed threshold lines, Q1=-1.687 and Q3=3.152, are vertically marked. The curve is divided into three regions, labeled as decelerated metabolic aging (mAA<Q1, accounting for 25% of the population), normal metabolic age (Q1≤mAA≤Q3, accounting for 50% of the population), and accelerated metabolic aging (mAA>Q3, accounting for 25% of the population). The stratification threshold is completely consistent with the fixed quartile parameters of sub-protocol C, intuitively demonstrating the standardized stratification judgment logic and providing a visual standard for reagent kit risk assessment.

[0096] Figure 6The figure presents example ROC results for a panel of 30 biomarkers in mAAA tri-class classification, and compares the ROC curves of the panel of 30 biomarkers in the mAAA tri-class classification task. The figure includes two sets of control curves: one set represents the complete 30 biomarker combination of this invention, and the other set represents the reduced control biomarker set. Each set of curves plots the tri-class ROC curves for the decelerating population, the normal population, and the accelerating population, with AUC values ​​labeled for each group. The micro-average AUC reaches 0.878, visually demonstrating that the 30 biomarker combination has excellent discriminative ability for stratifying metabolic aging, corresponding to the technical advantages of the biomarker combination in sub-scheme A.

[0097] Figure 7 The 5.5-year cumulative morbidity / mortality risk curves for the mAAA acceleration / deceleration groups (Kaplan–Meier) cover all-cause mortality (n=291), new-onset fatty liver (n=405), new-onset hypertension (n=323), new-onset diabetes (n=196), new-onset coronary heart disease (n=180), new-onset stroke (n=137), new-onset hyperlipidemia (n=137), and new-onset hyperuricemia (n=64) (all P<0.001); Figure 7 The study is divided into eight subplots, corresponding to eight clinical outcomes: all-cause mortality, new-onset coronary heart disease, new-onset stroke, new-onset hypertension, new-onset diabetes, new-onset hyperlipidemia, new-onset non-alcoholic fatty liver disease, and new-onset hyperuricemia. Each subplot has two curves representing the population with accelerated metabolic aging and the population with decelerated metabolic aging, respectively. The log-rank test was uniformly labeled, and the p-value was <0.001, which intuitively verified the stable predictive value of mAAA stratification for long-term adverse clinical outcomes and matched the application data of sub-plan E for chronic disease and all-cause mortality risk assessment.

[0098] Figure 8 Forest plot of association between mAAA and lifestyle / exposure factors, showing the regression coefficients (95% CI) of factors such as smoking, secondhand smoke, hair dye, shift work, dietary structure (whole grains / fried / processed meat / seafood, etc.), exercise participation, and family history of cardiovascular disease with mAAA. Figure 8 The internal structure lists all influencing factors, including smoking, secondhand smoke, hair dye, shift work, whole grain intake, regular exercise, fried food, processed meat, and family history of early-onset cardiovascular disease. The horizontal structure labels the regression coefficient β and 95% confidence interval for each factor. Positive coefficients represent increased mAAA (accelerated aging), and negative coefficients represent decreased mAAA (slowed aging). All coefficient values ​​are completely consistent with the lifestyle-related parameters in Example 6, supporting the application direction of the lifestyle intervention assessment product of this invention.

[0099] Figure 9This chart compares mAA with comprehensive clinical health indicators (LE8 cardiovascular health score, frailty index) and routine blood biochemical indicators in all-cause mortality prediction, demonstrating the incremental predictive value of mAA in various baseline models. Figure 9 The data is divided into three comparative subplots: the first group compares the univariate HR and C-index of mAAA, LE8 cardiovascular score, and frailty index; the second group compares the increase in C-index before and after adding mAAA to the baseline model of "LE8 + frailty index"; and the third group compares the changes in discrimination of the baseline model of 27 routine biochemical indicators before and after adding mAAA. The data are fully labeled with HR, ΔC-index, and LRT test p-value, quantitatively demonstrating that mAAA has incremental predictive value independent of traditional clinical indicators, corresponding to the beneficial effect of sub-protocol E on all-cause mortality risk assessment.

[0100] Example 1: Subject Recruitment and Plasma Sample Collection Inclusion criteria: Healthy individuals are defined as those who do not suffer from any of the following diseases: diabetes, coronary heart disease, stroke, atrial fibrillation, heart failure, peripheral vascular disease, gout, thyroid disease, chronic bronchitis, asthma, COPD, tuberculosis, chronic gastritis, peptic ulcer, viral hepatitis, fatty liver, cirrhosis, chronic kidney disease, fracture, or malignant tumor; anyone suffering from any of these diseases is included in the disease cohort.

[0101] Exclusion criteria: Individuals with acute or infectious symptoms within three weeks prior to sampling were excluded; pregnant and lactating women were excluded; and individuals who had used drugs that may affect metabolomics profiles within the past three months were excluded.

[0102] Plasma collection: Subjects fasted for at least 8 hours, and 1 mL of peripheral venous blood was collected and placed in an EDTA anticoagulant tube; centrifugation (2000×g, 10 min, 4°C) was completed within 2 hours after blood collection, and the supernatant plasma was collected, aliquoted and stored at -80°C.

[0103] Study cohort size: Healthy training set (from centers A, B, C) n=5252, aged 1-96 years; Healthy validation set (from centers D, E, F) m=5456, aged 2-88 years; Disease validation cohort (from centers D, E, F) q=7768, aged 59-62 years; Total 18476 participants.

[0104] Example 2, MALDI-TOF-MS detection of plasma metabolites: A1. Instruments and reagents: Nanoparticle-enhanced laser desorption / ionization time-of-flight mass spectrometer (MALDI-TOF-MS, Bruker GmbH, Germany); iron-based inorganic nanoparticles (particle size approximately 5–15 nm, 1 mg / mL deionized water); deionized water (resistivity ≥ 18.2 MΩ·cm).

[0105] A2, Sample dilution: Take 100 nL of plasma sample, dilute with deionized water at a volume ratio of 1:10, mix well and set aside.

[0106] A3, Target plate preparation: Take 500 nL of diluted plasma sample and spot it onto a MALDI target plate, and let it air dry at room temperature; then cover it with 500 nL of iron nanoparticle matrix solution.

[0107] A4, mass spectrometry acquisition: positive ion reflection mode, laser wavelength 355 nm, laser repetition rate 1000 Hz, accelerating voltage 20 kV, delayed extraction time 250 ns, 2000 laser emissions per acquisition, mass-to-charge ratio range m / z 100–500.

[0108] A5, Quality Control: Each target board is equipped with a 9-grid quality control layout; peak signals with a signal-to-noise ratio (SNR) > 3 are retained; intra-batch QC RSD < 20%; inter-batch QC CV < 25%.

[0109] Example 3: Data preprocessing and screening of 30 metabolites: 1. Spectrum processing and feature extraction (Python 3.11.4): Limit the detection range m / z to 100–500 Da, Gaussian smoothing (σ=1); TopHat baseline correction; peak detection (SNR=3); multi-sample peak alignment (m / z tolerance=0.01 Da); retain peaks with SNR>3, initially obtaining 232 metabolic features.

[0110] 2. Batch effect correction: The ComBat algorithm in the pycombat package is used for stratified batch correction - step one uses the mass spectrometry target plate number as the batch variable, and step two uses the acquisition center number (6 centers) as the batch variable.

[0111] 3. Feature transformation and standardization: Perform sign-preserving logarithmic transformation x′ = sign(x) × log(1 + |x|); StandardScaler standardization.

[0112] 4. Joint feature selection using three algorithms: 1) Dataset partitioning: The healthy samples (p=10708) were divided into a training set (n=5252) and an independent validation set (m=5456). 2) RandomForestRegressor(n_estimators=100, random_state=42): Takes the top 30% of features; 3) LassoCV(cv=5, random_state=42): Takes the top 30% of features with non-zero coefficients; 4) `mutual_info_regression(random_state=42)`: Retrieves the top 30% of features; 5) The union of the intersections of the three pairs of samples yielded 30 age-related metabolic candidates.

[0113] 5. Metabolite annotation: m / z library matching was performed using the Human Metabolome Database (HMDB); high-confidence structural annotation was further performed on the MALDI-TOF-MS features of 30 targets.

[0114] Example 4: Construction and Ablation Validation of Stacked Integration Model 1. Training of the first base learner: Perform 5-fold cross-validation (KFold, n_splits=5, shuffle=True, random_state=42) on the healthy training set (30 features, n=5, 252) for each of the 12 base learners (see Table 2) to generate the out-of-fold prediction matrix of each base learner for all training samples.

[0115] 2. Key hyperparameter optimization: XGBoost: n_estimators=700, max_depth=5, learning_rate=0.05, booster='gbtree'.

[0116] CatBoost: iterations=700, depth=10, learning_rate=0.05, l2_leaf_reg=7.

[0117] KNN: n_neighbors=3, weights='distance', algorithm='kd_tree', leaf_size=50, p=1.

[0118] MLP: 4 hidden layers (128→256→512→1024 neurons), each layer followed by ReLU+BatchNorm+Dropout(0.2); Adam(lr=0.001); early stopping (patience=20); maximum 500 rounds.

[0119] 3. Second-layer meta-learner fusion: Train Lasso meta-learners (LassoCV, cv=5, random_state=42); KNN has the highest weight (0.3429), followed by RF (0.2537), and XGBoost is third (0.1713); the weights of the four linear models are all close to or equal to zero.

[0120] 4. SHAP Ablation Validation: The SHAP value of each base learner for the final ensemble output was calculated using the shap library (TreeExplainer). Progressive pruning experiments showed that the first 3–5 learners could capture most of the predicted signals, validating the rationality of the ensemble architecture design.

[0121] Example 5: GAM Nonlinear Correction and mAA Calculation: 1. GAM model fitting: On the healthy training set (n=5, 252), the pygam library (LinearGAM) is used to fit pred_age = β0 + s(cAge) + ε, and the optimal smoothing parameter λ is selected through cross-validation.

[0122] 2. System bias extraction and correction: mAge = pred_age - b(cAge) = pred_age - g(cAge) + cAge; mAA = mAge - cAge = pred_age - g(cAge).

[0123] 3. Health reference cohort mAA statistics: Q1 = -1.687 years; Q3 = 3.152 years; three-category thresholds—accelerating (>3.152), normal (-1.687 to 3.152), and decelerating (<-1.687).

[0124] 4. Evaluation of correction effect: by age group - 70-80 years old group: average overestimation before correction was 4.8 years → 1.2 years after correction; 80 years and above group: 7.1 years before correction → 2.3 years after correction.

[0125] Example 6: 5.5-year longitudinal follow-up clinical validation (the ability of cross-sectional mAAA to assist in the assessment of long-term outcomes) Study population: Among baseline healthy participants, 6,169 provided 5.5 years of mortality follow-up (all-cause mortality events n=291 as of January 2025); 3,745 completed the 2023 follow-up questionnaire and reported new chronic disease status—new fatty liver (n=405), hypertension (n=323), diabetes (n=196), coronary heart disease (n=180), stroke (n=137), hyperlipidemia (n=137), hyperuricemia (n=64), etc.

[0126] Main results: (1) Prediction of all-cause mortality risk: The cumulative risk of all-cause mortality was significantly higher in the accelerated group compared to the decelerated group (log-rank P<0.001); Univariate Cox model: For every 1 SD increase in mAA, HR = 1.50 (95% CI 1.34–1.69, P = 6.14 × 10⁻⁶). -12 The mean squared error (C-index = 0.7351) was superior to the LE8 cardiovascular health score (HR = 0.79, 95% CI 0.70–0.89; C-index = 0.7336) and the frailty index (HR = 1.17, 95% CI 1.06–1.29; C-index = 0.7298). In a multivariate Cox model incorporating mAAA, LE8, and the frailty index, mAAA maintained independent predictive power (HR = 1.51, P = 1.64 × 10⁻⁶). -12 Adding mAAA to the "LE8+ Weakness Index" baseline model increased the C-index from 0.7352 to 0.7453 (LRT P = 8.01 × 10⁻⁶). -9 Compared with 27 routine blood biochemical indicators: mAAA remained independently significant after adjusting for any single biochemical indicator (HR range 1.50–1.53, all P < 0.001); after adjusting for all biochemical indicators, the association between mAAA and all-cause mortality remained robust (HR = 1.55, 95% CI 1.38–1.75, P = 3.03 × 10⁻⁶). -13 Furthermore, the C-index increased from 0.7402 to 0.7499 (LRT P = 1.23 × 10⁻⁶). -9 Based on comprehensive evidence, mAA has a better predictive ability for all-cause mortality than LE8, frailty index and most routine blood biochemical indicators, and has significant incremental predictive value.

[0127] (2) Prediction of multiple new chronic diseases: The cumulative incidence of new chronic diseases was significantly higher in the accelerated group compared with the decelerated group, including new fatty liver (n=405), new hypertension (n=323), new diabetes (n=196), new coronary heart disease (n=180), new stroke (n=137), new hyperlipidemia (n=137), and new hyperuricemia (n=64) (all log-rank P<0.001); the same pattern was seen in new atrial fibrillation, new malignant tumors, and new heart failure, demonstrating the wide application value of mAAA as an auxiliary assessment indicator for multiple outcomes.

[0128] (3) Lifestyle-related factors (suggesting that mAA can reflect the burden of modifiable lifestyle): Accelerating factors: smoking (β = +1.19 years, 95% CI 0.91–1.48, P<0.001), secondhand smoke exposure (β = +0.64 years, P<0.001), hair dye use (β = +0.99 years, P<0.001), shift work (β = +0.92 years, P = 0.001); intake of sugary foods, animal offal, fried foods, processed meats, and seafood (all P<0.05); family history of premature cardiovascular disease in fathers <55 years / mothers <65 years (all P<0.01); Mitigating factors: intake of whole grains (β = -0.34 years, P = 0.002), regular physical exercise (β = -0.48 years, P<0.001).

[0129] The above results indicate that mAA is sensitive to modifiable lifestyle factors such as smoking, exercise, and diet, and can be used as a metabolic quantitative indicator of the effectiveness of lifestyle interventions.

[0130] Table 3 compares mAAA with generalized clinical health indicators in mortality prediction. A systematic comparison of mAAA with generally accepted generalized clinical health assessment indicators demonstrates the translational value of this invention. Table 3

[0131] Conclusion: mAA has better univariate predictive ability for all-cause mortality than LE8 and frailty index; when mAA is added to a baseline model that already includes LE8, frailty index or all 27 blood biochemical indicators, the C-index is significantly improved, suggesting that mAA carries prognostic information independent of these existing indicators, can significantly enhance the ability of mortality risk stratification, and has clinical translation and health management application value.

[0132] Example 7, Extended Supplementary Verification: 1) UK Biobank Proteomics (n=2,922): R=0.941, R²=0.886, RMSE=7.2 years after application of stacked integrated architecture.

[0133] 2) NHANES clinical biochemical indicators (66 routine blood biochemical indicators): After applying the stacked integrated architecture, R=0.941, R²=0.885.

[0134] 3) Gut microbiome (relative abundance of OTUs at the genus level): R=0.732, R²=0.535 after applying the stacked integrated architecture.

[0135] Example 8: Implementation of the system, electronic device, and computer-readable storage medium: System architecture includes a data acquisition module, a preprocessing and batch correction module, a feature selection module, a stacked ensemble modeling module (12 base learners + Lasso meta-learner), a GAM nonlinear correction module, a mAA hierarchical evaluation module, and a comprehensive report output module.

[0136] Electronic device implementation: The processor executes program instructions stored in memory and can be deployed in hospital information systems (HIS), health management platform servers, physical examination centers, or research workstations.

[0137] Computer-readable storage medium: stores the above program instructions and pre-training parameters, including 12 base learner hyperparameters, Lasso meta-learner coefficients (see Table 1), fitting parameters of the GAM correction function g(cAge), and mAA stratification thresholds (Q1=-1.687, Q3=3.152).

[0138] Example 9: Composition and Use of the Detection Kit: Kit components: A. Plasma sample pretreatment reagents: deionized water (resistivity ≥ 18.2 MΩ·cm); B. Detection reagents: Iron-based inorganic nanoparticle matrix (particle size 5–15 nm, 1 mg / mL, compatible with MALDI-TOF-MS); C. Standards / Internal Standards: Standards or isotopic internal standards corresponding to the 30 metabolites mentioned above; D. Calibrators and quality control samples (QC pool); E. Computational software module (storage medium): Built-in stacked ensemble learning model parameters, GAM correction parameters, and mAA stratification threshold; F. Report Generation Module: Automatically outputs mAge, mAA values, acceleration / normal / deceleration stratification results, and reference reports; G. Instruction manual.

[0139] Procedure: Fasting blood collection from subjects → Plasma separation → 100 nL sample dilution → Spotting onto MALDI target plate → Nanoparticle matrix coverage → MALDI-TOF-MS mass spectrometry acquisition → Preprocessing and batch calibration → Stacked integrated model calculation of mAge → GAM calibration → Output mAge and three-part classification results → Generate individualized report.

[0140] This invention includes: a set of 30 age-related metabolic biomarkers (based on plasma metabolomics screening) reflecting metabolic physiological status; and a system for assessing metabolic age (mAge) and metabolic age acceleration (mAA) using high-dimensional, sparse, batch-effect-prone plasma metabolic feature matrices generated by matrix-assisted laser desorption / ionization mass spectrometry (MALDI-TOF-MS) and other mass spectrometry platforms. This invention can screen a set of metabolic biomarkers from plasma that systematically covers the core pathways of metabolic aging, focuses on the current state of metabolic physiology, and has been validated by large-scale cohorts throughout the entire life cycle. Based on this set of biomarkers in high-dimensional, sparse, batch-effect-prone mass spectrometry metabolic data scenarios, a highly accurate, age-group-biased, and robust cross-batch and cross-center metabolic age assessment system and storage medium are constructed for aging assessment, all-cause mortality risk assessment, and early warning of various chronic disease risks.

[0141] The core technical features of this invention are as follows: Feature 1: A functionally focused portfolio of metabolic biomarkers (biomarker level): Thirty age-related metabolites, identified through plasma metabolomics screening and focusing on the "current metabolic physiology," served as the core information carriers. These 30 metabolites cover key pathways reflecting the current metabolic physiology, including amino acid metabolism and derivatives, central carbon metabolism and organic acids, renal function and muscle metabolism, and sulfur-containing compounds. These 30 metabolites were not simply random combinations, but rather an effective set obtained through comprehensive screening based on time-series age associations, mAAA stratification ability, model contribution, and combinatorial stability in a large-scale population cohort.

[0142] Feature 2, Joint Feature Selection Using Three Algorithms (Modeling Level): To address the characteristics of high-dimensionality, sparseness, and strong correlation of mass spectrometry metabolic peaks, three methods are combined: Random Forest (RF) importance score, absolute value of Lasso regression coefficient, and mutual information (MI) score. Each method independently selects the top 30% of features, calculates the pairwise intersections of the three methods (RF∩Lasso, RF∩MI, Lasso∩MI), and then takes the union to form a robust set of age-related metabolic candidates (30 in total).

[0143] Feature 3: Two-layer stacked integrated architecture of 12 base learners: The first layer comprises 12 base learners across five major machine learning families: four linear models (LR / Lasso / Ridge / ElasticNet), five tree ensemble models (XGBoost / CatBoost / LightGBM / ExtraTrees / RF), one nearest neighbor model (KNN), one support vector machine (SVM), and one neural network (MLP). The linear models characterize the overall linear trend, the tree ensemble model captures higher-order nonlinear splitting relationships, KNN represents the neighborhood structure of locally similar samples, SVM fits complex boundaries through kernel mapping, and MLP extracts multi-layer nonlinear representations. Out-of-fold predictions are generated through 5-fold cross-validation to prevent information leakage. The second layer uses Lasso as the meta-learner to sparsely weight and fuse the outputs of the 12 base learners, automatically retaining stable contributing models and compressing redundant or noisy model weights to obtain the initial predicted age, pred_age.

[0144] Feature 4, GAM nonlinear system bias correction: A generalized additive model (GAM) is used to fit the nonlinear relationship between the initial predicted age pred_age and the time-series age cAge: pred_age = g(cAge) + ε. The systematic bias term b(cAge) = g(cAge) - cAge is extracted and subtracted from the initial prediction, outputting the corrected metabolic age mAge = pred_age - b(cAge). Based on this, the metabolic age acceleration mA = mAge - cAge is calculated. This step significantly reduces the systematic overestimation bias in the 70+ age group.

[0145] Feature 5, SHAP-driven progressive ablation verification: The contributions of the 12 base learners were ranked based on their SHAP values. The minimum effective learner combination was determined through progressive pruning experiments. The contribution ranking was as follows: KNN (0.3429), RF (0.2537), XGBoost (0.1713), CatBoost (0.1580), LightGBM (0.1032), etc. Progressive ablation experiments verified the necessity and stability of the integrated architecture.

[0146] Feature 6, mAAA three-level stratification and multiple outcome-assisted assessment: Based on the lower quartile Q1 = -1.687 years and upper quartile Q3 = 3.152 years of the mAA distribution in the healthy reference cohort, subjects were divided into three categories: accelerated metabolic aging (mAA > 3.152), normal metabolic age (-1.687 ≤ mAA ≤ 3.152), and decelerated metabolic aging (mAA < -1.687). The output results can be further used for auxiliary assessment of all-cause mortality risk, and auxiliary risk stratification for chronic diseases such as malignant tumors, hypertension, type 2 diabetes, stroke, hyperlipidemia, hyperuricemia / gout, and non-alcoholic fatty liver disease; it can also be extended to outcomes such as coronary heart disease, atrial fibrillation, heart failure, peripheral vascular disease, chronic bronchitis, COPD, gallstones, colon polyps, cirrhosis, and kidney stones.

[0147] Feature 7, cross-batch and cross-center robustness and compatibility with high-throughput detection platforms: The stacked integrated architecture maintains stable performance across different mass spectrometry batches, different acquisition centers, and independent health validation sets. It is preferably used in conjunction with a matrix-assisted laser desorption / ionization mass spectrometry (MALDI-TOF-MS) detection platform, employing an iron-based nanoparticle matrix (1 mg / mL). The plasma sample consumption does not exceed 100 nL, and the single sample detection time is approximately 30 seconds, which can support large-scale high-throughput screening deployment.

[0148] Feature 8, Integrated design of the test kit: The kit includes sample pretreatment reagents, standards or internal standards for the 30 metabolites, mass spectrometry matrix or ionization auxiliary reagents, calibrators, quality control materials, accompanying analytical instructions, and calculation software module. The kit can be used in either direct detection mode based on mass spectrometry or targeted detection mode after derivatization. The accompanying software module has a built-in metabolic age prediction model based on stacked ensemble learning and GAM correction parameters. It accepts quantitative data and time-series age of the 30 metabolites as input and outputs personalized reports with mAge, mAA, and three-category labels.

[0149] Obviously, those skilled in the art can make various modifications and variations to the invention without departing from the spirit and scope of the invention. Therefore, if these modifications and variations fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A metabolic age assessment system based on plasma metabolic biomarkers, characterized in that, include: The data acquisition module is used to collect plasma samples and time-series age (cAge) of the sample subjects; Based on the plasma samples collected from the subjects, the corresponding original mass spectrometry matrix was obtained; The preprocessing and batch calibration module is used to obtain standardized calibration parameters and the corresponding standardized metabolic feature matrix based on the original mass spectrometry matrix. The feature selection module is used to obtain a list of plasma metabolic biomarkers and a corresponding quantitative feature subset matrix of metabolic biomarkers based on the standardized and corrected metabolic feature matrix and the corresponding time-series age cAge. The stacked ensemble modeling module is used to obtain a trained and parameter-frozen base learner model based on the quantitative feature subset matrix of metabolic biomarkers and the time-series age cAge of the corresponding healthy training set. After training convergence, the optimal fusion weights and initial predicted ages of each base learner are obtained. The GAM nonlinear correction module is used to obtain the fitted GAM mapping function g(cAge), the corresponding corrected metabolic age, and the metabolic age acceleration based on the initial predicted age. The validation module is used to validate the effectiveness of the metabolic age assessment system composed of the solidified components. The solidified components include: standardized calibration parameters, a list of plasma metabolic biomarkers, a base learner model that has been trained and whose parameters are frozen, the optimal fusion weights corresponding to each base learner obtained after training convergence, and the fitted GAM mapping function g(cAge). The testing module is used to acquire the sample to be tested; based on the valid metabolic age assessment system, it obtains the corresponding corrected metabolic age and metabolic age acceleration of the sample to be tested. The preprocessing and batch correction module is used to perform peak preprocessing, two-layer rule solidification, sign-preserving logarithmic transformation, and normalization scaling.

2. The metabolic age assessment system based on plasma metabolic markers as described in claim 1, characterized in that, The peak preprocessing includes: performing Gaussian smoothing, TopHat baseline correction, SNR>3 peak detection and multi-sample peak alignment on the original mass spectrometry matrix of the healthy training set samples, filtering invalid low signal-to-noise ratio peaks, and outputting a coarsely extracted metabolic feature matrix. The two-layer rule solidification process includes: obtaining mass spectrometry target plate numbers that match the coarse extracted metabolic feature matrix row by row; calling the pycombat single-layer ComBat empirical Bayesian correction function and quality control samples, using the mass spectrometry target plate number as the unique batch stratification variable, performing deviation fitting on the peak intensity of each column of the coarse extracted metabolic feature matrix in the healthy training set, solving and solidifying the target plate stratification reference batch correction rule; and obtaining the intermediate feature matrix after target plate correction based on the solidified target plate stratification reference batch correction rule; the target plate stratification reference batch correction rule is used to constrain the target plate batch anchoring correction of subsequent samples to reduce technical variations caused by differences in spotting, laser intensity, and target plate substrate between different target plates within the same center; obtaining the acquisition center number and center-level quality control information that match the intermediate feature matrix after target plate correction row by row; and calling pycombat single-layer ComBat again. The correction function, using the acquisition center number as a global hierarchical batch variable, performs empirical Bayesian bias fitting on each column of the metabolic features in the intermediate feature matrix after target plate correction of the healthy training set, and solves the acquisition center-level reference batch correction rule; based on the acquisition center-level reference batch correction rule, the center-corrected feature matrix is ​​obtained; wherein, the acquisition center-level reference batch correction rule is used to constrain the center-level batch anchoring correction of subsequent samples, so as to reduce the global technical variation caused by differences in instrument models, experimental operations and reagent consumables across acquisition centers; the target plate-level reference batch correction rule and the acquisition center-level reference batch correction rule are recorded as the reference batch correction rule; The execution of the sign-preserving logarithmic transformation includes: substituting each value x in the center-corrected feature matrix corresponding to the healthy training set into a fixed formula: x′ = sign (x) × log (1 + |x|) to calculate the intermediate feature matrix after logarithmic transformation, where sign (x) preserves the original intensity sign, and log (1 + |x|) compresses the mass spectrometry skewness and large-scale sparse intensity to adapt to the distribution characteristics of mass spectrometry data; The standardization scaling process includes: dividing the sample rows corresponding to the healthy training set from the intermediate feature matrix after logarithmic transformation, grouping them by feature columns, calculating the mean μ and standard deviation σ of each column respectively, forming a fixed set of standardization scaling parameters; and substituting each column and element into the formula based on the mean μ and standard deviation σ of each column. After conversion, the standardized and corrected metabolic feature matrix corresponding to the healthy training set is obtained.

3. The metabolic age assessment system based on plasma metabolic markers as described in claim 2, characterized in that, The data acquisition module is used to collect plasma samples from subject participants. Each subject participant has a corresponding collection center number, time-series age (cAge), and sample queue grouping label. Mass spectrometry is performed on the plasma samples of each subject participant to obtain the corresponding raw mass spectrometry matrix. The sample queue grouping label includes: a healthy training set, an independent healthy validation set, and a disease validation set. The collection center number, mass spectrometry target plate number, sample queue grouping label, and time-series age (cAge) are bound row-by-row to the raw mass spectrometry matrix. The preprocessing and batch correction module is used to perform mass spectrometry data preprocessing, hierarchical batch correction, sign-preserving logarithmic transformation, and normalization scaling on the original mass spectrometry spectrum matrix of the healthy training set samples. It fits and solidifies the normalization correction parameters, including: peak processing parameters corresponding to mass spectrometry data preprocessing, reference batch correction rules corresponding to hierarchical batch correction, reference distribution parameters corresponding to reference distribution mapping, and normalization scaling parameters corresponding to normalization scaling, so as to obtain the normalized and corrected metabolic feature matrix corresponding to the healthy training set. The feature filtering module is used to filter features based on the standardized and corrected metabolic feature matrix and the corresponding time-series age cAge corresponding to the healthy training set, and lock in the list of plasma metabolic biomarkers; based on the list of plasma metabolic biomarkers, the module extracts the corresponding feature columns from the standardized and corrected metabolic feature matrix of the healthy training set to generate a subset matrix of quantitative features of metabolic biomarkers in the healthy training set. The stacked ensemble modeling module is used to obtain 12 base learner models with trained and frozen parameters based on the quantitative feature subset matrix of metabolic biomarkers and the time-series age cAge of the corresponding healthy training set. After training convergence, the optimal fusion weights of each of the 12 base learners and the initial predicted age corresponding to the healthy training set are obtained. The GAM nonlinear correction module is used to obtain the fitted GAM mapping function g(cAge), the corrected metabolic age, and the metabolic age acceleration based on the initial predicted age corresponding to the healthy training set. The validation module is used to validate the effectiveness of the metabolic age assessment system composed of fixed components based on independent health validation sets and disease validation sets. The fixed components include: standardized calibration parameters, a list of plasma metabolic biomarkers, 12 base learner models that have been trained and whose parameters are frozen, the optimal fusion weights corresponding to each of the 12 base learners obtained after training convergence, and the fitted GAM mapping function g(cAge).

4. The metabolic age assessment system based on plasma metabolic markers as described in claim 3, characterized in that, The preprocessing and batch correction module is further configured to, when performing mass spectrometry data correction on an independent health validation set, a disease validation set or a sample to be detected, synchronously acquire supporting reference samples, including: matching the independent health validation set with a corresponding quality control sample, matching the disease validation set with a corresponding internal standard sample, and matching the sample to be detected with a corresponding reference sample; the subsequent validation module or test module performs anchoring correction and distribution alignment on the metabolic feature intensity of the current detection batch based on the reference batch correction rule and the reference distribution parameter formed by the health training set, and maps the corrected metabolic features to the reference distribution of the health training set; then performing sign-preserving logarithmic transformation and standardized scaling processing to respectively obtain a standardized corrected metabolic feature matrix of the independent health validation set and a standardized corrected metabolic feature matrix of the disease validation set; wherein the reference distribution parameters are global statistical reference parameters fitted and solidified based on the health training set, comprising: the standardized mean μ and standard deviation σ corresponding to each metabolic feature, the ComBat empirical Bayesian prior distribution hyperparameters at the target plate level and the collection center level, and the signal reference distribution parameters of the quality control samples matched with the health training set; the subsequent validation module or test module reuses the reference distribution parameters to complete the distribution alignment of the metabolic features of the current batch, and uniformly maps the corrected metabolic features to the reference reference distribution of the health training set.

5. The metabolic age assessment system based on plasma metabolic markers as described in claim 4, characterized in that, A feature screening module, configured to screen out a sample row feature submatrix corresponding to the health training set from the standardized corrected metabolic feature matrix; invoking a RandomForestRegressor model configured with a first fixed hyperparameter, taking the sample row feature submatrix corresponding to the health training set as an independent variable input, and completing model training with the chronological age cAge supervision vector corresponding to the health training set as a prediction target; after the training is completed, outputting a global importance score corresponding to each column of metabolic features in the sample row feature submatrix corresponding to the health training set through a built-in algorithm of the model; sorting all the metabolic features in descending order according to the importance scores, and intercepting the first 30% of the metabolic features according to the sorting rank to obtain an RF primary selection feature set; invoking a LassoCV model configured with a second fixed hyperparameter, taking the sample row feature submatrix corresponding to the health training set as an independent variable input, taking the chronological age cAge supervision vector corresponding to the health training set as a prediction target, and performing 5-fold cross-validation to complete model training; after the training is completed, extracting the regression coefficient of the metabolic feature of each column in the sample row feature submatrix of the health training set, calculating the absolute value of the coefficient, eliminating invalid metabolic features with the absolute value of the coefficient equal to 0, sorting the remaining metabolic features in descending order according to the absolute value of the coefficient, and intercepting the first 30% of the metabolic features in the sorted order to obtain a Lasso primary selection feature set; Set a fixed random seed random_state=42 as the parameter of the mutual_info_regression function, calculate the mutual information regression score between each metabolic feature in the feature submatrix of the corresponding sample row of the healthy training set and the time-series age cAge supervision vector corresponding to the healthy training set column by column; sort all metabolic features from high to low according to mutual information score, and select the top 30% of metabolic features to obtain the initial mutual information feature set; Let RF, Lasso, and MI be the initial feature set, the initial feature set, and the initial feature set based on mutual information, respectively. Then, sequentially calculate the pairwise feature intersections RF∩Lasso, RF∩MI, and Lasso∩MI to obtain three sets of common feature subsets. Next, take the union of these three intersections and select 30 stable age-related metabolic candidates from this union to form a fixed list of 30 metabolic biomarkers. Based on this list of 30 metabolic biomarkers, extract the corresponding columns to generate a quantitative feature subset matrix of metabolic biomarkers.

6. The metabolic age assessment system based on plasma metabolic markers as described in claim 5, characterized in that, The list of plasma metabolic markers includes: L-cysteine, leucine, glyceric acid, 4-hydroxyproline, L-asparagine, serine, imidazole-4-acetaldehyde, creatinine, O-phosphothreonine, itaconic acid, betaine, 2-aminobenzoic acid, DL-O-phosphoserine, pyroglutamic acid, L-arginine, 3-hydroxyisovaleric acid, L-histidine, L-glutamic acid, taurine, selenocysteine, 2,3-dihydroxybutyric acid, acetoacetic acid, lactic acid, succinic acid, ribonucleic acid, phosphohydroxypyruvate, propionylcarnitine, L-carnitine, thiocyanate, and 2-hydroxyethanesulfonic acid.

7. The metabolic age assessment system based on plasma metabolic markers as described in claim 6, characterized in that, The stacked ensemble modeling module is used to construct 12 base learners across 5 major machine learning families. Fixed hyperparameters are preset for these 12 base learners, which include: LR, Lasso, Ridge, and ElasticNet linear models; SVM (Support Vector Machine); KNN (Kearn Nearest Neighbors); XGBoost, CatBoost, LightGBM, ExtraTrees, RF tree ensemble models, and MLP neural networks. Based on the quantitative feature subset matrix of biomarkers in the healthy training set and the corresponding time-series age cAge, the 12 base learners with fixed hyperparameters are invoked, and KFold is used to perform 5-fold cross-validation on the healthy training set. Within each fold, the corresponding base learner is fitted using the training subset within the fold. After all 5-fold cycles are completed, the fully trained and parameter-frozen 12 base learner models are retained. Simultaneously, each base learner generates a prediction vector based on the out-of-fold validation samples. The out-of-fold prediction vectors corresponding to the 12 base learners are horizontally concatenated column-wise to generate a prediction matrix with 12 columns and a number of rows equal to the total number of samples in the healthy training set (n). The prediction matrix of the healthy training set is used as the input feature matrix, and the time-series age cAge supervision vector corresponding to the healthy training set is matched row by row as the fitting label. The meta-learner is trained by calling LassoCV with a fixed configuration. After the training converges, the optimal fusion weights corresponding to the 12 base learners are obtained. In the meta-learner training, the prediction components corresponding to the base learners with low contribution or strong noise interference are assigned weights close to 0. This characteristic achieves noise suppression in subsequent weighting operations. The linear weighted summation operation is performed, including: taking the entire n×12 dimensional prediction matrix, multiplying each column of the prediction component of the matrix by the corresponding optimal fusion weight, and then summing all the product terms in the row of a single sample to output the initial predicted age pred_age corresponding to the healthy training sample.

8. The metabolic age assessment system based on plasma metabolic markers as described in claim 7, characterized in that, The GAM nonlinear correction module is used to fit a generalized additive model pred_age = g(cAge) + ε using the pygam library on a healthy training set, obtaining a fitted GAM mapping function g(cAge), where ε is the model's random error term. Based on the fitted g(cAge), a systematic bias term b(cAge) is extracted: b(cAge) = g(cAge) - cAge. The initial predicted age is corrected using the systematic bias term b(cAge), resulting in the corrected metabolic age mAge corresponding to the healthy training set: mAge = pred_age - b(cAge) = pred_age - g(cAge) + cAge. The metabolic age acceleration mA corresponding to the healthy training set is calculated by taking the difference between the corrected metabolic age and the time-series age: mAge = mAge - cAge = pred_age - g(cAge).

9. The metabolic age assessment system based on plasma metabolic markers as described in claim 8, characterized in that, Also includes: The mAA stratification assessment module is used to determine the stratification threshold corresponding to the healthy training set based on metabolic age acceleration. Based on the stratification threshold, metabolic aging stratification labels are obtained; The metabolic aging stratification labels include: accelerated metabolic aging, normal metabolic age, and decelerated metabolic aging.

10. A computer-readable storage medium having stored thereon computer-executable instructions, wherein, When the computer-executable instructions are executed by the processor, the processor causes the processor to: perform the system as described in any one of claims 1 to 9.

Images