A method for calculating biological age of diabetes patients based on glycosylation markers and multi-task deep learning framework

Abstract

The application discloses a diabetes patient biological age calculation method based on glycosylation markers and a multi-task deep learning framework, and belongs to the technical field of bioinformatics and medical data processing technology, and the method comprises the following steps: obtaining glycomics feature data and clinical examination data of diabetes patients; preprocessing the data; obtaining a feature set by adopting a multi-method screening and fusion strategy; constructing a deep learning model comprising a multi-modal feature fusion module and a multi-task learning module for training, wherein the multi-task learning module comprises a main task of aging clock regression and at least one auxiliary task; and outputting biological age prediction values and age acceleration evaluation results of individuals by using the trained model. The method realizes deep fusion of multi-modal biological data and exclusive modeling for the diabetes population, can accurately quantify the biological aging process and the deviation degree relative to the calendar age, and provides an objective quantitative tool for clinical evaluation.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and medical data processing technology, and in particular relates to a method for calculating the biological age of diabetic patients based on glycosylation markers and a multi-task deep learning architecture. Background Technology

[0002] As an important indicator of an organism's true physiological state, the aging clock reflects an individual's aging process and future health risks more effectively than calendar age. Currently, methods for predicting biological age mainly include epigenetic-based "methylation clocks," transcriptomic gene expression prediction models, and metabolomics-based biomarker combinations. These methods, by capturing relevant molecular changes during aging, provide quantitative evidence for assessing an individual's aging status and potential health risks, revealing to some extent the molecular mechanisms of aging, and have been applied to aging assessment and related research in the general population.

[0003] However, existing technologies still have significant limitations: First, most methods rely on information at a single glycomic level, failing to comprehensively integrate multidimensional biological data and making it difficult to capture the synergistic effects of different biological pathways in the aging process. Second, the statistical modeling methods used are mostly based on linear assumptions, lacking the ability to identify and characterize complex nonlinear relationships and interaction effects between variables, thus limiting prediction accuracy. Third, most existing models are built for generally healthy populations, lacking dedicated aging clock calculation models for specific disease populations such as diabetes, limiting their application in precision clinical management. It is particularly noteworthy that the fusion modeling of glycomic data and clinical phenotypic data has not been fully explored, and existing methods still have significant shortcomings in feature selection robustness, multimodal information fusion, and disease-specific modeling. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for calculating the biological age of diabetic patients based on glycosylation biomarkers and a multi-task deep learning framework, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, this invention provides a method for calculating the biological age of diabetic patients based on glycosylation biomarkers and a multi-task deep learning framework, comprising: S1. Obtain glycosylation data and clinical examination data of diabetic patients; S2. Preprocess the glycomics feature data and clinical examination data; S3. For the preprocessed feature data, multiple variable screening methods are used in parallel to screen features, and a stable candidate feature set is obtained by weighted rank fusion method. S4. Perform deep feature learning and cross-modal fusion on the glycomic features and clinical examination data in the candidate feature set respectively, and construct and train a multi-task deep learning network model based on shared representation, including the main task of aging clock regression and at least one auxiliary task. S5. The multi-task deep network model is trained and validated using a nested cross-validation strategy, and an early stopping strategy is used to control overfitting. S6. Use the trained model to calculate the biological age prediction value of individual diabetic patients and output the accelerated age assessment result based on the difference between the prediction value and the calendar age.

[0006] Preferably, the glycomics characteristic data in step S1 is obtained by plasma IgG N-glycosylation detection, and the clinical examination data includes at least one of demographic indicators, physiological and biochemical indicators, and metabolic indicators.

[0007] Preferably, the data preprocessing in step S2 includes: processing missing values ​​in the glycomics feature data using an autoencoder reconstruction interpolation method; identifying extreme values ​​and performing tail reduction processing using a comprehensive judgment method combining the standard deviation method and the interquartile range method; and performing batch effect correction on the glycomics feature data using the ComBat algorithm.

[0008] Preferably, the multiple variable screening methods described in step S3 include at least three of the following: univariate statistical test, mutual information screening, adaptive elastic network screening, stability screening, importance screening based on tree model, sparse group lasso screening, and graph regularization screening.

[0009] Preferably, the weighted rank fusion method in step S3 is as follows: assign weights to the feature importance ranking output by each variable screening method, calculate the weighted rank score of each feature, sort them from low to high according to the weighted rank scores, and select the top-ranked features to form the candidate feature set.

[0010] Preferably, in step S4, the process of constructing a multi-task deep network model includes: inputting glycomics feature data from the candidate feature set into a one-dimensional convolutional layer to extract local features, inputting clinical examination data into a fully connected layer for feature transformation, and then using a multi-head self-attention mechanism to fuse the features of the two modalities to form a shared representation; based on the shared representation, constructing a multi-task learning network structure containing a main task branch and at least one auxiliary task branch, wherein the main task branch is used to perform aging clock regression and output the predicted biological age, and the auxiliary task branch is used to perform auxiliary prediction tasks related to the progression of diabetes.

[0011] Preferably, the multi-task learning network structure in step S4 adopts a shared specific structure, including a shared fully connected layer and task-specific layers independent for each task; the output layer of the main task branch is a linear regression layer, and the auxiliary task branches include a binary output layer for classifying the risk of diabetes complications or a linear output layer for regressing inflammation levels.

[0012] Preferably, the comprehensive loss function of the multi-task learning network structure in step S4 is a weighted sum of the main task loss and at least one auxiliary task loss, and the weights are determined adaptively during training using an uncertainty weighting method.

[0013] Preferably, the nested cross-validation strategy in step S5 is as follows: an outer five-fold cross-validation is used to evaluate the model's generalization performance, and an inner three-fold cross-validation is used within each outer training fold for hyperparameter tuning and model selection.

[0014] Preferably, the formula for calculating the accelerated aging assessment result in step S6 is: AgeAccel = BioAge - ChronAge; Wherein, BioAge is the individualized biological age prediction value output by the model; ChronAge is the individual's true age at time; AgeAccel is the deviation of the individual's biological aging rate relative to normal aging time. When AgeAccel is positive, it indicates accelerated biological aging, and when AgeAccel is negative, it indicates delayed aging.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention is the first to integrate glycomics, a molecular marker that directly reflects the metabolic and immune status of diabetes, with clinical phenotypic data in a deep learning framework for modeling. It also utilizes a multi-task learning mode to introduce diabetes-related pathological information (auxiliary task) as regularization, thereby specifically modeling the unique physiological and pathological state of diabetic patients. This allows the final trained model to more accurately capture and quantify the aging patterns unique to this population, and the output bioage and ageaccel values ​​have higher population specificity and accuracy.

[0016] To address the limitations of limited information from a single data modality, this invention allows the model to automatically learn and integrate deep features and complex interactions from two different types of data—glycomics (high-dimensional, continuous, reflecting molecular modifications) and clinical indicators (structured, multidimensional, reflecting physiological states)—at different scales and with different properties. By leveraging their complementary information, the model constructs a more comprehensive and robust representation of an individual's aging state than from a single data source, thereby improving the model's prediction accuracy and generalization ability.

[0017] To address the shortcomings of traditional linear models in fitting performance, this invention employs a deep neural network architecture, focusing on the core feature of "constructing a multi-task deep learning model." This model, through its multi-layered nonlinear transformation structure, can automatically learn the highly complex nonlinear relationships and high-order interactions between input multimodal features and output aging clocks, without relying on manually pre-defined linear or simple nonlinear assumptions. This allows it to more realistically reflect the complexity of biological processes, significantly improving the goodness of fit and prediction accuracy of age prediction.

[0018] To address the issues of noise and redundant variables in high-dimensional data, this invention employs a multi-method screening and fusion strategy to obtain a feature set for model input. This approach combines the advantages of different screening methods (such as linear, nonlinear, sparsity, and structured methods) and achieves a more stable and reliable subset of key features through consensus-based fusion. This reduces the risk of overfitting and enhances the interpretability of the model's biological basis, providing high-quality, high-signal-to-noise ratio input for subsequent deep learning modeling.

[0019] This invention transforms model predictions into an intuitive "AgeAccel" by "outputting a predicted biological age and an age acceleration assessment result based on the difference between the predicted value and the calendar age." This indicator directly quantifies the degree to which an individual's biological aging deviates from the average level of their age group. Positive values ​​represent accelerated aging, while negative values ​​represent delayed aging. It provides clinicians with an objective and quantitative tool that can be used to assess the aging status of individual diabetic patients, stratify complication risks, and monitor the effectiveness of interventions, greatly enhancing the model's clinical applicability. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of 24 chromatographic peaks in an embodiment of the present invention; Figure 2 This is a flowchart of a method for calculating the biological age of diabetic patients based on glycosylation markers and a multi-task deep learning architecture, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of multimodal representation learning and multi-task deep networks according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0023] Example 1 like Figure 2 As shown, this embodiment provides a method for calculating the biological age of diabetic patients based on glycosylation biomarkers and a multi-task deep learning architecture, including: S1. Obtain glycosylation data and clinical examination data of diabetic patients; Furthermore, the glycomics characteristic data mentioned in step S1 is obtained by plasma IgG N-glycosylation detection, and the clinical examination data includes at least one of demographic indicators, physiological and biochemical indicators, and metabolic indicators.

[0024] Specifically, step S1 includes: S101. Collect blood samples; Those being tested are required to eat a light diet, avoid excessive alcohol consumption, and refrain from smoking the day before blood collection. They must not take any hypoglycemic medication on the morning of the blood collection. On the day of collection, a professional nurse will use an EDTA-anticoagulated vacuum blood collection tube to draw 5ml of fasting venous blood from the participant's arm in the morning. The blood sample will be immediately refrigerated and transported to the glucose testing laboratory.

[0025] S102, Blood Sample Processing; Remove the blood collection tube, affix the sample number (keep the tube as stable and vertical as possible during the process), and place it on a test tube rack to stand at room temperature for 1 hour. Then, place the blood collection tube in a centrifuge at 2000 rpm and 4°C for 15 minutes. After centrifugation, use a pipette to aspirate the supernatant plasma sample.

[0026] S103 and IgG N-glycosylation detection; The quantitative detection of plasma IgG glycosyl modification was performed using a hydrophilic-interactive high-performance liquid chromatography-ultra-high-performance liquid chromatography (HPLC-ULLC). The procedure mainly consisted of four steps: separation and purification of IgG using a 96-well Protein G extraction plate; extraction and purification of glycans using PNGase F enzyme; 2-AB fluorescent labeling of the glycans; and quantitative detection using an UPLC fluorescence detector. The specific experimental procedure is as follows: ①IgG isolation IgG was separated using a 96-well Protein G extraction plate.

[0027] ② IgG concentration detection The purity of the isolated IgG molecules was determined by sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE), using Precision Plus Protein All Blue Standards as molecular weight markers, stained with GelCode Blue, and imaged using a VersaDoc imaging system. Finally, the IgG concentration was determined using a BioSpec-nano scanner.

[0028] ③IgG deglycosylation The N-glycosyl group on IgG is removed using glycosidases (N-Glycosidase F, PNGase F).

[0029] ④ Glycosyl labeling and purification The glycosyl groups after enzyme digestion were further labeled with 2-AB (2-Aminobenzamide) fluorescent labeling.

[0030] ⑤UPLC glycosyl detection and structural identification Waters ACQUITY Ultra Performance Liquid Chromatography (UPLC) instrument was used to separate 2-AB fluorescently labeled N-glycosyl groups based on the principle of hydrophilic interaction. Following a conventional integrated algorithm, the detection data was automatically processed, and peaks for each glycosyl group were plotted according to the peak area and retention time of the detected sugar chains. The final chromatogram was divided into 24 peaks (GP1-GP24) using the same method (e.g., Figure 1 ).

[0031] The N-glycosyl structure of IgG is represented by the following letters and numbers commonly used in glycosylation research: F represents a core fucose attached to the core N-acetylglucosamine; Mx represents “x” mannose attached to the core N-acetylglucosamine; Ax represents “x” branches (N-acetylglucosamine) attached to the trimannose core; A2 represents N-acetylglucosamine attached to both branches of the core mannose; B represents bisecting N-acetylglucosamine attached to the core mannose; Gx represents “x” galactose attached to the branches; [3]G1 and [6]G1 represent galactose attached to α1-3 and α1-6 mannose, respectively; Sx represents “x” sialic acid attached to galactose. Table 1 shows the meaning of the 24 glycosyl peaks measured in the experiment and the structural meaning of the glycosyl peaks. See Table 1 for the structural diagram.

[0032] Table 1 S104. Input of glycomics and clinical examination data (gender, age, blood pressure, blood lipids, BMI, liver and kidney function, etc.). In this invention, the clinical examination data includes the following specific indicators and their units: Gender (Male=1, Female=0); Age (unit: years); Body Mass Index (BMI, unit: kg / m²); Systolic blood pressure (SBP, unit: mmHg); Diastolic blood pressure (DBP, unit: mmHg); Total cholesterol (TC, unit: mmol / L); Triglycerides (TG, unit: mmol / L); Low-density lipoprotein cholesterol (LDL-C, unit: mmol / L); High-density lipoprotein cholesterol (HDL-C, unit: mmol / L); Alanine aminotransferase (ALT, unit: U / L); Aspartate aminotransferase (AST, unit: U / L); Total bilirubin (TBIL, unit: μmol / L, optional); Creatinine (Cr, unit: μmol / L); Blood urea nitrogen (BUN, unit: mmol / L); Fasting blood glucose (FBG, unit: mmol / L); Glycated hemoglobin (HbA1c, unit: %) Albumin (ALB, unit: g / L); Total protein (TP, unit: g / L).

[0033] S2. Preprocess the glycomics feature data and clinical examination data; Furthermore, the data preprocessing in step S2 includes: using an autoencoder reconstruction interpolation method to process missing values ​​in the glycomics feature data; using a comprehensive judgment method combining the standard deviation method and the interquartile range method to identify extreme values ​​and perform tail reduction processing; and using the ComBat algorithm to perform batch effect correction on the glycomics feature data.

[0034] Specifically, the data preprocessing method in this embodiment includes: missing value processing, extreme value detection and processing, standardization, batch correction, and feature grouping, dividing glycosyl features into groups according to glycan biosynthetic pathways or peak clusters; (1) Methods for handling missing values; Continuous variables (such as blood pressure, blood lipids, liver and kidney function, peak glucose levels, etc.) are imputed using Multiple Imputation by Chained Equations (MICE). The imputation model is based on a multiple linear regression or Bayesian regression framework, taking into account the correlation between variables.

[0035] For categorical variables (such as gender, whether or not there are complications), mode imputation is used to replace missing values ​​with the most common category of the variable.

[0036] For glycosylated peak data, an autoencoder-based reconstruction imputation method was used to recover missing feature values ​​through unsupervised learning.

[0037] (2) Extreme value detection and processing; Extreme value identification employs a combination of two methods: Standard deviation method: For continuous variables, use the range of mean plus or minus three standard deviations (mean ± 3SD) to determine extreme values; Interquartile Range (IQR) method: Calculate the interquartile range IQR = Q3 - Q1. If the observed value exceeds the range of [Q1 - 1.5×IQR, Q3 + 1.5×IQR], it is considered an outlier.

[0038] The identified extreme values ​​are processed using Winsorization, which replaces values ​​exceeding the upper bound with the upper bound and values ​​below the lower bound with the lower bound, in order to maintain the overall stability of the data distribution.

[0039] (3) Data standardization methods; All continuous features (including glycosylation peak intensity and clinical biochemical indicators) were normalized using the Z-score algorithm, calculated as follows: z = (x - μ) / σ Where μ is the characteristic mean and σ is the standard deviation.

[0040] For non-normally distributed variables, perform a logarithmic transformation (log10 or ln) before standardization to approximate a normal distribution.

[0041] Before model training, the features of the validation and test sets are standardized on the same scale using the mean and standard deviation of the training set to avoid information leakage.

[0042] (4) Batch effect correction method; Systematic biases exist between different experimental batches or instruments in glycomics data. The ComBat algorithm (Empirical Bayes ComBat Correction) is used for correction. This algorithm preserves biological variation while adjusting batch means and variances to eliminate non-biological differences.

[0043] For clinical data where there are differences in testing instruments or sample sources, median centering is used as a supplementary method. The characteristic median of each batch is adjusted to the overall median level to ensure the consistency of distribution across different batches.

[0044] S3. For the preprocessed feature data, multiple variable screening methods are used in parallel to screen features, and a stable candidate feature set is obtained by weighted rank fusion method. Furthermore, the various variable screening methods described in step S3 include at least three of the following: univariate statistical tests, mutual information screening, adaptive elastic network screening, stability screening, tree model-based importance screening, sparse group lasso screening, and graph regularization screening.

[0045] Furthermore, the weighted rank fusion method described in step S3 is as follows: assign weights to the feature importance ranking output by each variable screening method, calculate the weighted rank score of each feature, sort them from low to high according to the weighted rank scores, and select the top-ranked features to form the candidate feature set.

[0046] Specifically, the multi-variable screening method employs multiple advanced filters and obtains a stable candidate set through rank fusion. The specific steps include: S301, Variable selection method type; The "various advanced filters" used in this invention include, but are not limited to, the following types: Univariate t-tests or Spearman correlation screening are used to assess the linear or monotonic correlation between each feature and the target variable (age); Mutual information filtering is used to identify nonlinear correlation features; Adaptive Elastic Net (AENet) filters out redundant variables by combining a linear model with L1 and L2 regularization terms. Stability selection identifies high-frequency inclusion variables through repeated sampling and LASSO modeling. Random Forest Importance evaluates variable contributions based on the feature importance of a tree model. Sparse Group Lasso (SGL) achieves feature sparsity while preserving group structure information; Graph regularization, based on the adjacency relationship between features, constrains the smoothing of adjacent feature weights, preserving structural consistency at the path level.

[0047] S302, Core parameters of each filter; Univariate t-test and mutual information screening: The two-sided significance level P < 0.05 or FDR < 0.05 after Benjamini-Hochberg method correction was used as the retention criteria.

[0048] Adaptive elastic network: The regularization parameter α is taken in the range of 0.5 to 0.8 and is determined by the inner five-fold cross-validation; the weight coefficients are set according to the initial LASSO regression results.

[0049] Stable selection: Using a 50% sampling ratio, 50% of the samples are randomly selected each time, and the number of repetitions is set to 100. Features with a selection frequency greater than 0.7 are considered stable features.

[0050] Random forest importance screening: The forest size is set to 500 decision trees, and the importance threshold is the portion above 1.2 times the average value of all features.

[0051] Sparse group lasso: The ratio of group regularization parameter λ1 to individual sparse parameter λ2 is set to 1:0.5, and the final λ value is determined based on the principle of minimizing cross-validation error.

[0052] Graph regularization filtering: The adjacency matrix is ​​constructed based on the similarity of characteristic biological pathways or peak clusters, and the Laplacian regularization parameter γ is between 0.1 and 0.5.

[0053] S303, the computational logic of rank fusion; To synthesize the results of different filters and improve the robustness of feature selection, this invention employs a weighted rank aggregation method. Its calculation logic is as follows: The feature list output by each filter is ranked, with higher rankings indicating greater feature importance. Each filter is assigned a weight w_i, which is determined based on the robustness and interpretability of the method. For example, linear methods (AENet, SGL) have higher weights, while random methods (Random Forest) have slightly lower weights. Calculate the weighted rank score for each feature: Score_j = Σ (w_i × rank_i(j)) Reordered from lowest to highest weighted score, with lower scores indicating higher overall importance; The top N features are selected as the final candidate feature set.

[0054] S304, Criteria for determining candidate feature sets.

[0055] The number of features N is determined based on the principle of minimizing cross-validation error. The optimal value of N is selected by calculating the mean absolute error (MAE) or coefficient of determination (R²) of the prediction model under different values ​​of N.

[0056] If there is no significant inflection point, the N value is determined based on the point where the cumulative contribution rate of feature importance reaches 90%.

[0057] Meanwhile, an overlap analysis was performed on the results of each screening method. If a feature was selected repeatedly in three or more methods, it was given priority to be included in the candidate set.

[0058] The final candidate feature sets are labeled with glycosyl feature subsets and clinical feature subsets respectively, providing a unified feature matrix for subsequent multimodal deep learning model input.

[0059] S4. Perform deep feature learning and cross-modal fusion on the glycomic features and clinical examination data in the candidate feature set respectively, and construct and train a multi-task deep learning network model based on shared representation, including the main task of aging clock regression and at least one auxiliary task. Further, in step S4, the process of constructing a multi-task deep network model includes: inputting glycomics feature data from the candidate feature set into a one-dimensional convolutional layer to extract local features, inputting clinical examination data into a fully connected layer for feature transformation, and then using a multi-head self-attention mechanism to fuse the features of the two modalities to form a shared representation; based on the shared representation, constructing a multi-task learning network structure containing a main task branch and at least one auxiliary task branch, wherein the main task branch is used to perform aging clock regression and output the predicted biological age, and the auxiliary task branch is used to perform auxiliary prediction tasks related to the progression of diabetes.

[0060] Furthermore, the multi-task learning network structure described in step S4 adopts a shared specific structure, including a shared fully connected layer and task-specific layers independent for each task; the output layer of the main task branch is a linear regression layer, and the auxiliary task branches include a binary output layer for classifying the risk of diabetes complications or a linear output layer for regressing inflammation levels.

[0061] Furthermore, the comprehensive loss function of the multi-task learning network structure described in step S4 is a weighted sum of the main task loss and at least one auxiliary task loss, with the weights determined adaptively during training using an uncertainty weighting method.

[0062] Specifically, such as Figure 3As shown, multimodal representation learning and multi-task deep networks utilize glycosylation and clinical modalities to improve the accuracy and robustness of the main task (age regression) and construct a shared-specific multi-task deep structure. Specific steps include: S401, Implementation of multimodal representation learning; Glycomics Data Processing and Feature Extraction: The glycomics data is a continuous feature matrix, with the input dimension being the number of samples × the number of glycosylation peaks. The model first extracts local features through a one-dimensional convolutional layer to capture the structural patterns and local smoothing features between glycan peaks. The number of convolutional kernels is set to 32, the kernel size is 3, the stride is 1, and the activation function is ReLU. The convolutional output is processed by batch normalization and max pooling layers to obtain a low-dimensional embedding representation.

[0063] Clinical data processing and feature extraction: Clinical examination data consisted of low-dimensional structured numerical features, with the input dimension being the number of samples multiplied by the number of indicators. Feature transformation was performed using two fully connected layers: the first layer had 64 neurons, and the second layer had 32 neurons, both using the ReLU activation function. The output features were then normalized to form a clinical modality embedding representation.

[0064] Modal fusion and interactive learning: Glycosyl embeddings and clinical embeddings exchange information during the fusion phase through an attention mechanism. A multi-head attention structure is employed, with four heads, each with a dimension of 32. The attention mechanism calculates the weighted feature contributions between different modalities, achieving cross-modal feature fusion. The fusion results are then integrated into a shared representation vector via a linear layer, which is used as input for subsequent multi-task learning modules.

[0065] S402, Task definition for multi-task learning; This invention employs a multi-task learning framework, simultaneously optimizing both the main and auxiliary tasks to enhance the model's robustness and feature representation capabilities. The tasks are defined as follows: Main task: Aging clock regression task, using the calendar age of the samples as the supervision signal, the output is a continuous variable, and the loss function is Huber regression loss.

[0066] Auxiliary Task A: Risk Classification of Diabetic Complications. This task determines the presence of diabetic complications based on clinical diagnostic labels, outputting a binary variable with cross-entropy loss as the loss function.

[0067] Auxiliary Task B: Inflammation Level Regression Task. This task uses inflammation-related indicators (such as C-reactive protein levels or white blood cell count) as the target output to predict the degree of inflammation, with the mean squared error as the loss function.

[0068] The main task and auxiliary tasks share some underlying feature representations, but each retains its own task-specific branch to avoid mutual interference between tasks.

[0069] S403, Shared-Specific Multitasking Deep Architecture Design; The overall network architecture consists of two parts: a shared layer and a task-specific layer.

[0070] The shared layer consists of two fully connected layers with 128 and 64 neurons respectively, using ReLU as the activation function. The output of the shared layer serves as the basis for the input features of the multi-task.

[0071] The task-specific layer consists of a fully connected layer (32 neurons) and a regression output layer (1 neuron). The output layer does not use an activation function to enable continuous value prediction.

[0072] The auxiliary task A-specific layer consists of a fully connected layer (32 neurons) and a classification output layer (1 neuron). The classification layer uses the Sigmoid activation function to output probability values.

[0073] The auxiliary task B-specific layer consists of a fully connected layer (32 neurons) and a regression output layer (1 neuron), which uses a linear activation function to output the inflammation level value.

[0074] The loss function of each task branch is dynamically adjusted using uncertainty weighting, and the final comprehensive loss is the weighted sum of the losses of the main task and the auxiliary task.

[0075] S404, Activation Function and Regularization Settings; All hidden layers use the ReLU activation function to speed up convergence and avoid gradient vanishing. The output layer selects either a Sigmoid or a linear function depending on the task type.

[0076] The Dropout layer is used to prevent overfitting. A Dropout operation is added between the shared layer and the task-specific layer, with the ratio set to 0.3.

[0077] The model uses L2 regularization to constrain the weight parameters, with a regularization coefficient of 0.001.

[0078] During the training phase, an early stopping strategy is used. If the validation set error does not improve within 10 consecutive epochs, training is terminated early.

[0079] S405, Training and Optimization Strategies; The optimization algorithm uses the Adam optimizer, with an initial learning rate of 0.001 and a batch size of 64.

[0080] The model training employed a five-fold cross-validation strategy to evaluate performance stability.

[0081] The outputs after training include: predicted aging clock, aging clock acceleration value (predicted age minus calendar age), complication risk score, and inflammation level prediction value.

[0082] S5. The multi-task deep network model is trained and validated using a nested cross-validation strategy, and an early stopping strategy is used to control overfitting. Furthermore, the nested cross-validation strategy described in step S5 is as follows: an outer five-fold cross-validation is used to evaluate the model's generalization performance, and an inner three-fold cross-validation is used within each outer training fold for hyperparameter tuning and model selection.

[0083] Specifically, this embodiment employs nested cross-validation (outer layer evaluation, inner layer parameter tuning) and an early stopping strategy to control overfitting, performing model training, validation, and model selection. The specific steps include: S501, Nested Cross-Validation Design; Outer cross-validation is used to evaluate the model's generalization performance, employing 5-fold cross-validation. All samples are randomly divided into five subsets, with one subset selected each time as the outer test set, and the remaining four subsets used for inner training and validation.

[0084] Inner-layer cross-validation is used for model parameter optimization and hyperparameter tuning, employing 3-fold cross-validation. The inner training set is further divided into three parts, with two parts used for training and one part for validation, obtaining the optimal parameter combination through iterative iteration.

[0085] The nested design of outer and inner layers can effectively avoid data leakage problems and ensure that parameter tuning and performance evaluation processes are independent of each other, thereby obtaining robust model performance estimates.

[0086] S502, Early Stopping Strategy and Model Saving Rules; Early stopping is used to prevent overfitting. During training, if the validation set loss does not decrease for five consecutive training epochs, training is automatically terminated.

[0087] The model is saved using an optimal weight saving mechanism. During training, the validation set loss value is monitored in real time, and the current model parameter file is automatically saved whenever the validation loss reaches a new minimum. The final model parameters used for outer layer evaluation are the version with the minimum validation set loss.

[0088] If the validation set loss fluctuates significantly at the end of training, the average weight of the three most recent optimal points is taken as the final model weight to improve stability.

[0089] S503, Model Evaluation Indicators and Criteria for Determining the Optimal Model; This invention primarily uses the following three indicators to comprehensively evaluate the model performance: (1) Mean Absolute Error (MAE) reflects the average deviation between the predicted value and the actual value; (2) Root Mean Square Error (RMSE) reflects the overall fluctuation of prediction error; (3) Coefficient of determination (R²) is used to measure the extent to which the model explains the variation in aging clock.

[0090] During the model selection process, the above indicators are calculated for each outer test set, and the average value of five folds is taken as the overall performance indicator of the model.

[0091] The criteria for determining the optimal model are: the model with the highest R² value and the smallest MAE is determined as the final model; when multiple models have similar performance, the one with the highest weighted composite score of R² and RMSE is given priority.

[0092] The model's final output includes a predicted aging clock and an age acceleration value (predicted age minus calendar age), and reports the average results of MAE, RMSE, and R² on the validation set to demonstrate the model's overall performance across different tasks.

[0093] S504, Training and Optimization Settings; The model was trained using the Adam optimization algorithm, with an initial learning rate of 0.001, a weight decay coefficient of 0.0001, and a batch size of 64.

[0094] The loss function is a weighted combination of the Huber regression loss of the main task and the cross-entropy or mean squared error loss of the auxiliary task, with the weights being automatically learned through an uncertainty weighting method.

[0095] After each training session, the training loss, validation loss, and validation metric change curves are recorded for subsequent analysis and interpretability assessment.

[0096] S6. Use the trained model to calculate the biological age prediction value of individual diabetic patients and output the accelerated age assessment result based on the difference between the prediction value and the calendar age.

[0097] Furthermore, the formula for calculating the accelerated age assessment result in step S6 is as follows: AgeAccel = BioAge - ChronAge; BioAge is the individualized biological age prediction value output by the multimodal, multi-task deep learning model constructed in this embodiment; ChronAge refers to the individual's true age at time, that is, the natural age calculated from the date of birth; AgeAccel is the deviation of the individual's biological aging rate relative to normal aging time; when AgeAccel is positive, it indicates accelerated biological aging, and when AgeAccel is negative, it indicates delayed aging.

[0098] In this embodiment, the model constructed by the method includes a multi-level model system from simple to complex, and the feature sets input by the models at different levels cover feature subsets of different ranges and complexities.

[0099] Example 2 To verify the feasibility and effectiveness of the multi-task deep learning biological aging prediction method based on glycomics and clinical data described in Example 1, this example selects a batch of clinically diagnosed type 2 diabetes patients for model construction and validation. The specific implementation steps are as follows: I. Sample Source and Grouping; A total of 600 clinically diagnosed patients with type 2 diabetes were collected. All participants signed informed consent forms, and the study was approved by the ethics committee. Based on the sample proportions, 500 patients were selected as the training set for model training and parameter optimization; another 100 patients were selected as the validation set for model performance evaluation and final testing.

[0100] II. Sample inclusion and exclusion criteria; Inclusion criteria: The patient has been clinically diagnosed with type 2 diabetes for at least one year.

[0101] Age range: 30 to 70 years old, gender not limited.

[0102] It has complete clinical testing records and plasma samples, which can be used for glycomics analysis.

[0103] There was no severe liver or kidney failure or acute cardiovascular or cerebrovascular events.

[0104] Exclusion criteria: Individuals with concurrent malignant tumors, autoimmune diseases, or hematological disorders.

[0105] There has been an acute inflammatory or infectious disease within the past three months.

[0106] People who take immunosuppressants or glucocorticoids for a long time.

[0107] Those samples have serious quality defects in collection or testing, making it impossible to guarantee the accuracy of the analysis.

[0108] III. Data Acquisition and Glycosyl Detection; Venous blood samples were collected from all patients. After plasma separation, IgG was extracted for N-glycosylation analysis. Glycosylation was detected using high-performance liquid chromatography (HPLC) to obtain the relative abundance data of each glycan peak. The glycosylation profile of each sample included 20 to 30 major peaks, which were normalized by peak area for subsequent modeling.

[0109] IV. Compilation of Clinical Examination Data; Clinical and biochemical test data were collected for each subject during the same period, including the following indicators: Basic information: gender, age, body mass index (BMI, unit kg / m²). Blood pressure indicators: Systolic blood pressure (SBP, mmHg), Diastolic blood pressure (DBP, mmHg); Blood lipid indicators: total cholesterol (TC), triglycerides (TG), low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C), all in mmol / L; Liver function indicators: alanine aminotransferase (ALT) and aspartate aminotransferase (AST), in units of U / L; Kidney function indicators: creatinine (Cr, unit μmol / L), blood urea nitrogen (BUN, unit mmol / L); Inflammatory marker: C-reactive protein (CRP, mg / L).

[0110] V. Data Preprocessing; Missing value handling: Continuous variables are imputed using multiple interpolation (MICE), and categorical variables are imputed using the mode.

[0111] Extreme value handling: Use the standard deviation method and interquartile range method to judge extreme values, and perform tail reduction processing on outliers.

[0112] Standardization: All continuous variables were logarithmically transformed and then standardized using Z-score.

[0113] Batch correction: The ComBat algorithm was used to eliminate batch effects in glycomics data.

[0114] VI. Multi-method variable selection and fusion; In the combined matrix of glycomic and clinical features, variables were screened using a variety of methods, including univariate t-test, mutual information screening, adaptive elastic net (AENet), stability selection, random forest importance screening, sparse group lasso (SGL), and graph regularization screening.

[0115] The main parameters of each filter are as follows: The stable screening was repeated 100 times, with a sampling ratio of 50% and a feature selection frequency threshold of 0.7. The regularization parameter α of the elastic net is 0.5, and the value of λ is determined through three-fold cross-validation. The random forest has 500 trees, and the importance threshold is 1.2 times the average value.

[0116] The results of each method were combined using a weighted rank fusion algorithm, and the top 80 features with the highest weighted scores were selected as the candidate feature set, of which glycosyl features accounted for approximately 60% and clinical features accounted for 40%.

[0117] VII. Multimodal representation learning and multi-task deep network construction; Glycomics data were processed by a one-dimensional convolutional layer with 32 kernels and a kernel size of 3; clinical data were converted into low-dimensional embedding vectors by a two-layer fully connected network.

[0118] The features of the two modalities are fused through a multi-head self-attention mechanism (4 heads, 32 dimensions) to form a unified multimodal representation.

[0119] The multi-task architecture includes shared layers (two fully connected layers with 128 and 64 neurons respectively) and task-specific layers: The main task is an aging clock regression task, and the output is a continuous variable; Auxiliary task A is a risk classification task for diabetes complications, and the output is a binary variable. Auxiliary task B is a regression task, and its output is a continuous variable.

[0120] The loss function includes the mean squared error loss of the main task and the cross-entropy loss of the auxiliary task. The weights are determined in the range of 0.1 to 1.0 through grid search to minimize the overall loss.

[0121] VIII. Model Training and Validation; Nested cross-validation was used for training and evaluation, with an outer five-fold cross-validation layer and an inner three-fold cross-validation layer. The optimized algorithm was Adam, with an initial learning rate of 0.001 and a batch size of 64.

[0122] The early stopping strategy stops training if the validation set loss does not decrease for five consecutive training epochs, and saves the model parameters with the minimum validation loss.

[0123] Model performance is primarily measured by mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). The optimal model is determined by maximizing R² and minimizing MAE.

[0124] IX. Aging Clock and Accelerated Age Calculation.

[0125] The continuous variable predicted by the model is the individualized aging clock (BioAge). The method for calculating the aging clock acceleration value (AgeAccel) is as follows: AgeAccel = BioAge - ChronAge BioAge is the individualized biological age prediction value output by the multimodal, multi-task deep learning model constructed in this embodiment; ChronAge refers to the individual's true age at time, that is, the natural age calculated from the date of birth; AgeAccel is the deviation of the individual's biological aging rate relative to normal aging time; when AgeAccel is positive, it indicates accelerated biological aging, and when AgeAccel is negative, it indicates delayed aging.

[0126] Example 3 In this embodiment, the sample size is set to be... The glycomics feature matrix is (Each glycosyl peak), the clinical examination feature matrix is ​​as follows (Gender, age, blood pressure, blood lipids, BMI, liver and kidney function, etc.), concatenate and input as follows ( (Based on the calendar age of diabetic patients) The outcome variable; the predicted age output by the model. Defined as the aging clock (BioAge), accelerated aging is defined as... .

[0127] S1 Data preprocessing and correction; 1) Missing values ​​and extreme value handling: Continuous variables are truncated by quantiles (1%–99%), and missing values ​​are handled by multiple imputation or autoencoder-based reconstruction imputation; 2) Normalization: Z-score standardization is performed on each feature. The glycosylation peaks are logarithmically approximated as normal. 3) Batch / instrument effect correction: Batch correction based on empirical Bayes (ComBat) is used, and the batch matrix is ​​denoted as follows. After correction, the result is ; 4) Feature grouping information: Glycosyl features are divided into groups according to glycan biosynthetic pathways or peak clusters. , used for subsequent sparse regularization.

[0128] S2 Multi-method variable selection and fusion; To from higher dimensions Robust and complementary candidate features are obtained through screening. This embodiment employs multiple advanced screeners in parallel and obtains a stable candidate set through rank fusion. .

[0129] S2.1 Univariate and Mutual Information Pre-screening: For each feature Calculation and Age Pearson / Spearman related or mutual information The set of salient features is preserved after BH-FDR correction. .

[0130] S2.2 Adaptive Elastic Net (AENet) Filtering: submatrix Fitting Weight Based on prior knowledge (adaptive). The selected feature set is denoted as by inner-layer cross-validation. .

[0131] S2.3 Stability Selection: Repeatedly selecting random subsamples (e.g.) Perform LASSO (times) and record the frequency of each feature being selected. Thresholding is then performed.

[0132] S2.4 Nonlinear Importance Filtering: Importance Estimation Based on Gradient Boosting Tree / Extreme Random Tree (or the absolute mean of SHAP) As an importance), thresholding is obtained .

[0133] S2.5 Sparse Group Lasso (SGL): Solving considering glycosylation grouping structure. get .

[0134] S2.6 Graph Regularization (Pathway / Peak Cluster Adjacency Constraint): Constructing the Laplacian Matrix of the Feature Map Solve To encourage smoothing of adjacent feature weights in pathways, we obtain .

[0135] S2.7 Rank Fusion and Capacity-Constrained Selection: Ranking of Methods Perform weighted Borda fusion under capacity constraints Please ask: The final candidate feature set was obtained, and the glycosyl and clinical subsets were labeled respectively.

[0136] S3 Multimodal Representation Learning and Multitask Deep Networks; To simultaneously utilize glycosylation and clinical modalities and improve the accuracy and robustness of the primary task (age regression), this embodiment constructs a shared-specific multi-task deep structure.

[0137] S3.1 Feature Embedding and Interaction: Linear embedding and batch normalization are used for continuous features; Multi-head self-attention is then introduced to model higher-order interactions between features.

[0138] S3.2 Shared Trunk Road + Expert Hybrid (MMoE): [This section appears to be incomplete and requires further context.] One expert; Each task gating vector The task indicates: S3.3 Task Header and Losses: Main task (age regression): Using Huber loss Auxiliary Task A (Order classification of age bins for regularization representation): Let the true age bin labels be... Multi-class cross-entropy Auxiliary Task B (Input Reconstruction / Denoising for Robustness): Masked Autoencoder Reconstruction Loss S3.4 Task Adaptive Weighting and Structure Regularization: Employing Uncertainty Weighting (learnable) ) Group sparsity regularization and L1 sparsity are applied to the weights of the first layer to maintain the interpretability of the glycosylation pathway: Overall Objective: S4 Training, Validation, and Model Selection; Nested cross-validation (outer layer evaluation, inner layer parameter tuning) and an early stopping strategy were employed to control overfitting. The main metrics were MAE and... : Model selection criteria: Minimum MAE / Maximum MAE on the outer validation set and examine The correlation between the distribution stability and clinical complications.

[0139] S5 aging clock and aging acceleration calculation and output; Obtain the final model Then, for the new individuals (after simultaneous preprocessing via S1): AgeAccel = BioAge - ChronAge Among them, BioAge is the individualized biological age prediction value output by the multimodal multi-task deep learning model constructed in this embodiment; ChronAge refers to the individual's true age at time, that is, the natural age calculated from the date of birth; AgeAccel is the amount of deviation of the individual's biological aging rate relative to normal aging time.

[0140] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for calculating the biological age of diabetic patients based on glycosylation biomarkers and a multi-task deep learning architecture, characterized in that, Includes the following steps: S1. Obtain glycosylation data and clinical examination data of diabetic patients; S2. Preprocess the glycomics feature data and clinical examination data; S3. For the preprocessed feature data, multiple variable screening methods are used in parallel to screen features, and a stable candidate feature set is obtained by weighted rank fusion method. S4. Perform deep feature learning and cross-modal fusion on the glycomic features and clinical examination data in the candidate feature set respectively, and construct and train a multi-task deep learning network model based on shared representation, including the main task of aging clock regression and at least one auxiliary task. S5. The multi-task deep network model is trained and validated using a nested cross-validation strategy, and an early stopping strategy is used to control overfitting. S6. Use the trained model to calculate the biological age prediction value of individual diabetic patients and output the accelerated age assessment result based on the difference between the prediction value and the calendar age.

2. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, The glycomics characteristic data mentioned in step S1 are obtained by plasma IgG N-glycosylation detection, and the clinical examination data includes at least one of demographic indicators, physiological and biochemical indicators and metabolic indicators.

3. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, The data preprocessing in step S2 includes: using an autoencoder reconstruction interpolation method to process missing values ​​in the glycomics feature data; using a comprehensive judgment method combining the standard deviation method and the interquartile range method to identify extreme values ​​and perform tail reduction processing; and using the ComBat algorithm to perform batch effect correction on the glycomics feature data.

4. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, The various variable screening methods described in step S3 include at least three of the following: univariate statistical tests, mutual information screening, adaptive elastic network screening, stability screening, importance screening based on tree models, sparse group lasso screening, and graph regularization screening.

5. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, The weighted rank fusion method described in step S3 is as follows: assign weights to the feature importance ranking output by each variable screening method, calculate the weighted rank score of each feature, sort them from low to high according to the weighted rank scores, and select the top-ranked features to form the candidate feature set.

6. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, In step S4, the process of constructing a multi-task deep network model includes: inputting glycomics feature data from the candidate feature set into a one-dimensional convolutional layer to extract local features, inputting clinical examination data into a fully connected layer for feature transformation, and then using a multi-head self-attention mechanism to fuse the features of the two modalities to form a shared representation; based on the shared representation, constructing a multi-task learning network structure containing a main task branch and at least one auxiliary task branch, wherein the main task branch is used to perform aging clock regression and output the predicted biological age, and the auxiliary task branch is used to perform auxiliary prediction tasks related to the progression of diabetes.

7. The method for calculating the aging clock of diabetic patients according to claim 6, characterized in that, The multi-task learning network structure described in step S4 adopts a shared specific structure, including a shared fully connected layer and task-specific layers independent for each task; the output layer of the main task branch is a linear regression layer, and the auxiliary task branches include a binary output layer for classifying the risk of diabetes complications or a linear output layer for regressing inflammation levels.

8. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, The comprehensive loss function of the multi-task learning network structure described in step S4 is a weighted sum of the main task loss and at least one auxiliary task loss. The weights are determined adaptively during training using an uncertainty weighting method.

9. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, The nested cross-validation strategy described in step S5 is as follows: outer five-fold cross-validation is used to evaluate the model's generalization performance, and inner three-fold cross-validation is used within each outer training fold for hyperparameter tuning and model selection.

10. The method for calculating the aging clock of diabetic patients according to claim 1, characterized in that, The formula for calculating the accelerated age assessment result in step S6 is as follows: AgeAccel = BioAge - ChronAge; Wherein, BioAge is the individualized biological age prediction value output by the model; ChronAge is the individual's true age at time; AgeAccel is the deviation of the individual's biological aging rate relative to normal aging time. When AgeAccel is positive, it indicates accelerated biological aging, and when AgeAccel is negative, it indicates delayed aging.

Images