Diabetes cognitive impairment method based on metabonomics analysis and prediction

Abstract

The invention discloses a diabetes cognitive impairment method based on metabonomics analysis and prediction, and relates to the technical field of biological information, and the method comprises the following steps: S1, obtaining metabonomics data and immunomics data from a peripheral blood sample of a diabetic patient, extracting relevant time sequence data aiming at glucose metabolism, and calculating the glucose metabolism related time sequence data; processing the sequence data by adopting a time sequence analysis algorithm to obtain time sequence change characteristics; s2, constructing a cross-omics interaction network according to time sequence change characteristics, integrating an incidence relation between metabolite concentration and immune factor expression, and setting a dynamic interaction mode; according to the diabetes cognitive impairment method based on metabonomics analysis and prediction, through multi-omics data integration and dynamic network analysis, the precision and reliability of diabetes cognitive impairment mechanism analysis are remarkably improved, and a theoretical basis is provided for precise intervention.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, specifically to a method for predicting diabetic cognitive impairment based on metabolomics analysis. Background Technology

[0002] Multi-omics analysis, as a cutting-edge field in life sciences, integrates data from multiple biological omics systems to reveal the interaction mechanisms of different biological processes in complex diseases, playing a crucial role in understanding disease occurrence and development, and in developing precision medicine strategies. Particularly in research on chronic diseases such as diabetes, the interaction between the immune and metabolic systems is considered a significant factor influencing disease progression and complications (such as cognitive impairment). Compared to single-omics studies, multi-omics methods can more comprehensively depict the dynamic changes in biological systems, providing important tools for revealing disease mechanisms and identifying therapeutic targets. However, current multi-omics research still faces significant technical challenges in addressing complex diseases, especially in research on how immune-metabolic interactions synergistically affect cognitive function in diabetic states, where breakthroughs in existing methodologies are urgently needed.

[0003] Existing methods often struggle to fully capture the dynamic relationships between different omics datasets when integrating multi-omics data. Traditional analyses typically rely on static data integration at a single time point, neglecting the temporal changes in the immune and metabolic systems during the progression of diabetes. For example, some studies may analyze glycemic metabolic disorders using metabolomics while simultaneously detecting inflammatory cytokine levels using immunoomics, but these methods fail to reveal the synergistic mechanisms of these two processes at different stages of the disease. Furthermore, when processing high-dimensional data, existing technologies often suffer from high data heterogeneity and significant noise interference, leading to the obscuring of key biological signals. These limitations make it difficult for researchers to accurately identify core regulatory factors in immune-metabolic interactions, thereby restricting a deeper understanding of the mechanisms of cognitive impairment.

[0004] At the technical level, the core challenge in integrating metabolomics and immunomics data lies in effectively analyzing the dynamic interaction networks across these omics frameworks. In diabetes, there is a complex interaction between immune system activation (such as the release of inflammatory factors) and metabolic disorders (such as insulin resistance), and this interaction continuously changes as the disease progresses. However, a key problem with current technologies in constructing cross-omics interaction networks is the lack of an effective computational framework to capture the patterns of these dynamic changes. For example, small changes in metabolite concentrations may trigger specific responses in immune cells, but existing methods struggle to accurately distinguish which metabolite changes are the drivers of the immune response.

[0005] The complexity of this dynamic interactive network leads to another technical challenge: how to accurately extract key regulatory factors and signaling pathways from high-dimensional, heterogeneous omics data. Because metabolomics data (e.g., metabolite concentrations) differ significantly from immunomics data (e.g., cytokine expression) in magnitude and time scale, existing analytical tools struggle to identify the core factors truly driving immune-metabolic interactions against a noisy background. For example, in diabetic patients, a certain metabolite may affect the inflammatory response through a specific signaling pathway, leading to neuronal damage and cognitive decline, but current methods struggle to pinpoint this causal relationship from massive datasets. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting cognitive impairment in diabetes based on metabolomics analysis. This method extracts and verifies key regulatory factors and signaling pathways of immune-metabolic interaction from multi-omics data, and studies the synergistic effect of immune system activation and metabolic disorders on cognitive function in diabetic states.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting diabetic cognitive impairment based on metabolomics analysis, comprising: S1, obtaining metabolomics and immunomics data from peripheral blood samples of diabetic patients, extracting time-series sequence data related to glucose metabolism, processing the sequence data using a time-series analysis algorithm to obtain time-series change characteristics; S2, constructing a cross-omics interaction network based on the time-series change characteristics, integrating the correlation between metabolite concentration and immune factor expression, and defining a dynamic interaction pattern; S3, extracting regulatory factors from the dynamic interaction pattern, grouping the factors using a clustering method, and marking the factors within a group as candidate regulatory factors if the correlation strength between the factors and metabolic disorders exceeds a preset threshold, thereby obtaining a set of candidate regulatory factors; S4, analyzing the candidate regulatory factor set... S5. For signaling pathways, pathway activation levels are obtained by matching a pre-established pathway database with pathway enrichment analysis. Partial correlation analysis and mutual information calculation are used to evaluate the association between pathways and cognitive function indicators. If the significance test shows that the association coefficient is higher than the association strength threshold, the interaction driving mechanism is confirmed, and a description of the driving mechanism is obtained. S6. Based on the description of the driving mechanism, the factor extraction accuracy is optimized. Noise interference factors are removed by iterative threshold adjustment and noise identification criteria. Residual analysis and time alignment methods are used to process the remaining factors. If the clustering stability and matching degree measure reach the preset standard after filtering iterations, a list of regulatory factors is output. S7. From the list of regulatory factors, the interaction network details are integrated to generate a comprehensive interaction model to describe specific immune factors in the diabetic state, and the model parameters are obtained.

[0008] Preferably, step S1 includes extracting metabolomics and immunomics data from peripheral blood samples using liquid chromatography-mass spectrometry to obtain raw metabolic data and raw immune data; cleaning the raw metabolic and immune data to remove noise and outliers, resulting in cleaned metabolic and immune data; extracting glucose metabolism-related features using principal component analysis to obtain metabolic and immune feature sets; if the number of features in the metabolic and immune feature sets exceeds a preset threshold, screening features highly correlated with glucose metabolism using correlation analysis to obtain a screened feature set; processing the time-series data using an autoregressive moving average model to obtain a time-series model; extracting time-series variation features using Fourier transform from the time-series model to obtain a time-series variation feature set; and grouping the features using cluster analysis to obtain time-series variation feature groups.

[0009] Preferably, step S2 includes obtaining metabolomics and immunomics data from peripheral blood samples, separating metabolites and immune factors using liquid chromatography-mass spectrometry to obtain raw metabolic and immunological data; processing the raw metabolic and immunological data using data standardization methods to eliminate dimensional differences, obtaining standardized metabolic and immunological data; calculating the Pearson correlation coefficient and Spearman rank correlation coefficient between metabolite concentration and immune factor expression based on the standardized metabolic and immunological data to obtain a preliminary correlation matrix; if the absolute value of the correlation coefficient in the preliminary correlation matrix is ​​greater than a preset threshold, retaining the corresponding metabolite-immune factor pair to obtain a screened correlation matrix; constructing a cross-omics interaction network using a weighted gene co-expression network analysis algorithm for the screened correlation matrix to obtain the interaction network topology; dividing the network modules using a community detection algorithm based on the interaction network topology to determine the dynamic interaction mode; and analyzing the temporal correlation between metabolites and immune factors using time series analysis to obtain a temporal interaction feature set for the dynamic interaction mode.

[0010] Preferably, step S3 includes: acquiring potential regulatory factors from the dynamic interaction mode; extracting factor features through principal component analysis to obtain a factor feature set; grouping the potential regulatory factors according to the factor feature set using hierarchical clustering to obtain factor grouping results; calculating the Pearson correlation coefficient between factors and metabolic disorder features within each group to obtain a correlation strength matrix; if the absolute value of the correlation coefficient in the correlation strength matrix exceeds a preset threshold, marking the corresponding factor as a candidate regulatory factor to obtain a candidate regulatory factor set; constructing a prediction model of factors and metabolic disorders using a random forest algorithm based on the candidate regulatory factor set to obtain model output results; extracting the temporal correlation between candidate regulatory factors and metabolic disorders using time series analysis to obtain a temporal correlation feature set; and dividing the candidate regulatory factors into functional modules using a community detection algorithm based on the temporal correlation feature set to obtain a functional module set.

[0011] Preferably, step S4 includes extracting relevant signaling pathways from the candidate regulatory factor set, matching a preset pathway database using pathway enrichment analysis to obtain a pathway activation level set; calculating the correlation strength between each pathway and cognitive function indicators using partial correlation analysis based on the pathway activation level set to obtain a correlation strength set; if the absolute value of the correlation coefficient in the correlation strength set exceeds a preset threshold, verifying the interaction relationship between the pathway and cognitive function indicators using mutual information calculation to obtain a preliminary interaction relationship set; constructing an interaction network between pathways using network topology analysis for the preliminary interaction relationship set to obtain a network connection strength set; dividing the pathways into functional modules using a community detection algorithm based on the network connection strength set to obtain a functional module set; extracting the temporal correlation between each module and cognitive function indicators using time series analysis for the functional module set to obtain a temporal correlation feature set; and predicting the influence weight of the functional modules on cognitive function using a logistic regression model based on the temporal correlation feature set to obtain a description of the interaction driving mechanism.

[0012] Preferably, step S5 includes extracting an initial factor set from the candidate adjustment factor set, initially grouping the factors using a k-means clustering algorithm to obtain initial factor groups; if the intra-cluster variance of the initial factor groups exceeds a preset threshold, optimizing the grouping boundaries using an iterative threshold adjustment method to obtain an optimized factor set; calculating the signal-to-noise ratio (SNR) of the factor signals using a noise identification standard based on the optimized factor set; if the SNR is lower than a preset standard, removing noise interference factors to obtain a preliminary refined factor set; evaluating the clustering stability of the preliminary refined factor set using the DBSCAN algorithm to obtain a stability index set; if any index in the stability index set is lower than a preset standard, residual analysis is performed... The method calculates factor residuals, removes outliers, and obtains a modified factor set. Based on the modified factor set, dynamic time warping is used to align the factor time series, resulting in an aligned factor set. For the aligned factor set, the matching degree measure between the factors and a preset pathway database is calculated, resulting in a matching degree set. If any value in the matching degree set is lower than a preset standard, the modified factor set is returned to be re-aligned, resulting in an updated aligned factor set. Based on the updated aligned factor set, principal component analysis is used to extract principal component features, resulting in a feature factor set. For the feature factor set, the correlation strength with cognitive function indicators is verified. If the correlation strength reaches a preset standard, a refined moderating factor list is output.

[0013] Preferably, step S6 includes extracting interaction network data from the refined regulatory factor list, constructing an initial network structure using a graph neural network algorithm to obtain a network node set; calculating the interaction strength between nodes based on the network node set, evaluating the correlation between nodes using the Pearson correlation coefficient method to obtain an association matrix; if the correlation value in the association matrix is ​​lower than a preset threshold, removing the corresponding node and updating the network node set to obtain a refined node set; for the refined node set, using a community detection algorithm to divide the related sub-networks of immune factors and glucose transporters to obtain a sub-network set; based on the sub-network set, extracting the interaction path between cytokine storm and glucose transporter dysregulation, calculating the path interaction weights to obtain a weight vector; for the weight vector, using principal component analysis to extract the main interaction features to obtain a feature vector; if the correlation strength between the feature vector and the diabetic state reaches a preset standard, outputting a comprehensive interaction model to obtain model parameters.

[0014] Preferably, the method further includes S7: Simulating the cognitive impairment process based on model parameters, tracking the trend of change through parameter adjustment, verifying the deviation between the trend and actual patient data using heatmap visualization and multivariate adjustment; if the deviation is lower than a preset threshold, confirming the model's effectiveness and obtaining the verified interaction mechanism. Specifically, this includes obtaining dynamic change data of cognitive function from the model parameters, constructing an initial trend model using time series analysis to obtain a preliminary trend vector; extracting key change points in the time dimension based on the preliminary trend vector, grouping change features using cluster analysis to obtain a categorized change set; generating heatmap visualization data for the categorized change set, mapping change intensity using color coding to obtain a visualization matrix; extracting change areas from the visualization matrix, calculating the deviation value from patient data using multivariate regression to obtain a deviation statistic; if the deviation statistic is lower than a preset threshold, verifying model consistency using logistic regression to obtain a verification result; adjusting the trend model parameters based on the verification result, updating the interaction mechanism using iterative optimization to obtain an optimized interaction mechanism; extracting core interaction paths from the optimized interaction mechanism, quantifying path weights using path analysis to obtain a final weight vector.

[0015] Preferably, the method further includes S8: extracting remote micro-regulatory sub-networks from the verified interaction mechanism, refining the sub-network nodes using clustering stability analysis, and determining the fine-grained driving path for diabetes-induced memory decline if the matching degree between the sub-network nodes and temporal changes passes the preset standard verification. Specifically, this includes: extracting the regulatory sub-network from the optimized interaction mechanism, separating the sub-network nodes using network decomposition methods, obtaining the temporal correlation strength between nodes, and obtaining a node association set; extracting strong correlation pairs from the node association set, grouping node features using k-means clustering, and obtaining node feature groups; if the temporal change matching degree of the node feature groups passes the preset standard verification, refining the sub-network structure using path clustering methods to obtain a preliminary driving path; extracting key path nodes from the preliminary driving path, calculating path weights using weighted network analysis methods, and obtaining a path weight vector.

[0016] Preferably, step S8 further includes: if the stability of the path weight vector is verified by a preset threshold, then adjusting the path structure by the correlation strength between nodes to obtain an optimized driving path; extracting core driving nodes from the optimized driving path, quantifying the dynamic interaction between nodes using a time series analysis method to obtain a dynamic interaction matrix; extracting interaction regions from the dynamic interaction matrix, verifying the matching degree between the regions and diabetes-induced memory decline using clustering stability analysis to obtain the driving path.

[0017] As can be seen from the above technical solution, the present invention has the following beneficial effects:

[0018] This method for predicting diabetic cognitive impairment based on metabolomics analysis aims to address the complex mechanism underlying memory decline caused by metabolic and immune interactions in diabetic patients. By acquiring metabolomics and immunomics data from peripheral blood samples, glucose metabolism-related time-series sequences are extracted. Time-series analysis algorithms reveal dynamic changes, and a cross-omics interaction network is constructed. The correlation between metabolite concentration and immune factor expression is integrated, and Pearson and Spearman correlation analyses are used to determine dynamic interaction patterns. Furthermore, candidate regulatory factors are extracted using clustering methods, and the association between factors and cognitive function indicators is evaluated based on pathway enrichment analysis and partial correlation analysis to confirm the driving mechanism. Iterative threshold adjustment and residual analysis refine the factor list, generating a comprehensive interaction model. The cognitive impairment process is simulated, and the model's effectiveness is verified through heatmap visualization. Finally, a micro-regulatory sub-network is extracted to refine the driving pathway of diabetes-induced memory decline. This invention significantly improves the accuracy and reliability of elucidating the mechanism of diabetic cognitive impairment through multi-omics data integration and dynamic network analysis, providing a theoretical basis for precise intervention. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, the present invention provides a technical solution: a method for predicting diabetic cognitive impairment based on metabolomics analysis, including S1, obtaining metabolomics data and immunomics data from peripheral blood samples of diabetic patients, extracting time-series sequence data related to glucose metabolism, processing the sequence data using a time-series analysis algorithm, and obtaining time-series change characteristics;

[0022] S2. Based on the temporal variation characteristics, construct a cross-omics interaction network to integrate the correlation between metabolite concentration and immune factor expression, and define a dynamic interaction pattern. S3. Extract regulatory factors from the dynamic interaction pattern, group factors using clustering methods. If the correlation strength between a factor within a group and metabolic disorder exceeds a preset threshold, it is marked as a candidate regulatory factor, resulting in a candidate regulatory factor set. S4. For the candidate regulatory factor set, analyze related signaling pathways, match them with a pre-established pathway database through pathway enrichment analysis to obtain pathway activation levels, and evaluate the association between pathways and cognitive function indicators using partial correlation analysis and mutual information calculation. If the significance test shows that the correlation coefficient is higher than the correlation strength threshold, the interaction driving mechanism is confirmed, and a description of the driving mechanism is obtained. S5. Based on the driving mechanism description, optimize the factor extraction accuracy, remove noise-interfering factors through iterative threshold adjustment and noise identification criteria, and process the remaining factors using residual analysis and temporal alignment methods. If the clustering stability and matching degree measure reach the preset standard after filtering iterations, output a list of regulatory factors. S6. From the list of regulatory factors, integrate the interaction network details to generate a comprehensive interaction model to describe specific immune factors in the diabetic state, obtaining model parameters.

[0023] In this embodiment, peripheral blood samples from diabetic patients are first collected to obtain metabolomics and immunomics data, constructing time-series data containing glucose metabolism characteristics. Time-series analysis algorithms (such as Dynamic Time Warping (DTW) and multidimensional autoregressive models) are used to process the data, extracting time-series features reflecting dynamic changes in metabolic and immune states. Based on these features, a cross-omics interaction network is constructed to correlate changes in metabolite concentrations with immune factor expression levels, uncovering potential time-series dynamic interaction patterns. Further, potential regulatory factors are extracted from these interaction patterns and clustered based on their association strength with metabolic disorders to construct a candidate regulatory factor set. Subsequently, enrichment analysis and partial correlation analysis are used to identify statistical associations between relevant signaling pathways and cognitive function indicators, thereby identifying driving mechanisms. Based on these driving mechanisms, the extraction of regulatory factors is optimized, and residual analysis and time alignment are used to remove noise interference, forming a stable and reliable list of regulatory factors. Finally, combining this list and its interaction relationships, a comprehensive interaction model is generated to depict the dynamic mechanism of immune factor action in diabetic states, thereby enabling the prediction of cognitive impairment in diabetes.

[0024] The aforementioned methods comprehensively consider the temporal changes in metabolomics and immunomics, effectively establishing dynamic interaction models and significantly improving the accuracy and individualization of diabetes cognitive impairment prediction. By introducing clustering and pathway enrichment analysis, the systemic regulatory relationships of regulatory factors can be deeply explored, clarifying regulatory pathways and biological mechanisms. Residual analysis and temporal alignment further enhance data denoising capabilities and model stability, making the prediction results more clinically interpretable and valuable. The overall approach integrates multi-dimensional data, dynamic analysis, and mechanism modeling, exhibiting significant advantages over traditional biomarker screening methods in terms of accuracy, efficiency, and reliability.

[0025] S1 includes extracting metabolomics and immunomics data from peripheral blood samples using liquid chromatography-mass spectrometry (LC-MS / MS) to obtain raw metabolic and immunological data; cleaning the raw metabolic and immunological data to remove noise and outliers, resulting in cleaned metabolic and immunological data; extracting glucose metabolism-related features using principal component analysis (PCA) to obtain metabolic and immunological feature sets; if the number of features in the metabolic and immunological feature sets exceeds a preset threshold, filtering for highly correlated features using correlation analysis to obtain a filtered feature set; processing the time-series data using an autoregressive moving average model to obtain a time-series model; extracting time-series variation features using Fourier transform from the time-series model to obtain a time-series variation feature set; and grouping the features using cluster analysis to obtain time-series variation feature groups.

[0026] In implementing this method, peripheral blood samples were first collected from diabetic patients, with at least 5 ml of venous blood collected from each patient. After anticoagulation, the blood was centrifuged, and plasma was extracted for metabolomics and immunomics analysis. Data extraction was performed using liquid chromatography-mass spectrometry (LC-MS). In metabolomics analysis, samples were sequentially separated in the LC column. Different metabolites were separated according to their polarity and mass characteristics before entering the mass spectrometry section for ionization and detection. Approximately 800 metabolic features were output for each sample. In immunomics analysis, based on the different antigen-antibody binding sites of immune factors, the same mass spectrometry platform was used to identify at least 100 immune factor levels. The above detection results were compiled into metabolic data matrices and immunomics data matrices. The raw data is cleaned by first calculating the upper and lower quartiles of each feature based on its distribution range, and then determining the range of outliers by multiplying the difference between the upper and lower quartiles by 1.5. All values ​​below the lower limit or above the upper limit are considered outliers and are removed. Next, missing data is filled in by using the median of the same feature in all samples to fill in missing items, ensuring the integrity of the matrix and generating cleaned metabolic and immune data.

[0027] Principal component analysis (PCA) was performed on the cleaned metabolic and immune data. First, each feature data was standardized by subtracting its mean and dividing by its standard deviation to achieve a distribution with a mean of zero and a variance of one. Then, the covariance matrix between features was calculated, and eigenvalue decomposition was performed. The top principal components were selected based on their eigenvalues, with the criterion being that they collectively explained at least 85% of the variance. These principal components, corresponding to the original features, constituted the initial metabolic and immune feature sets. The number of features in each set was counted. If any set contained more than 30 features, the feature selection process began. During selection, the correlation between each feature and the glucose metabolism level of diabetic patients was calculated using the Pearson correlation coefficient. The selection criterion was an absolute correlation coefficient greater than or equal to 0.8, derived from prior clinical experience and data-driven assessment, serving as a threshold to ensure a close correlation between features and core metabolic indicators of diabetes. Only features meeting this correlation condition were retained, forming the selected feature set, with an average of approximately 20 to 25 features retained per sample.

[0028] Next, a time series was constructed for each selected feature, with a length of no less than 5 time points, generated based on continuous patient monitoring data or historical records. Each series was modeled using an autoregressive moving average model, with two key parameters: the autoregressive order and the moving average order. These were selected by iterating through all possible combinations within an integer range of 1 to 5, calculating the Akaike Information Criterion value for each combination, and choosing the combination with the smallest criterion value as the final model parameter configuration. After modeling, residuals and fitted values ​​were extracted to verify model stability and store the predicted output. Subsequently, each time series was subjected to a Fourier transform, converting the time series into a frequency series. Three types of frequency domain features—dominant frequency, amplitude, and phase—were extracted to characterize the main periodic trends and intensities of the features over time. All frequency domain features were grouped into a time-series variation feature set.

[0029] Within the selected feature set, each metabolic or immune-related feature includes its measurements at multiple time points. These time points represent actual test data obtained from regular patient follow-ups or continuous sampling, and consist of at least five time points, forming a complete time-series data sequence. For each feature's time-series data sequence, it is first input into a pre-established autoregressive moving average model. The values ​​at each time point are then fitted, outputting a fitted time-series curve. The purpose of this fitting process is to smooth fluctuations and compensate for trend changes, enabling a more accurate modeling of local fluctuations and overall trends in the original time-series sequence.

[0030] After obtaining the smoothed time series output by the model, a Fourier transform is performed on each sequence. The specific steps are as follows: The time-ordered values ​​from each sequence are input into the Fourier transform calculation module, which transforms the sequence from the time domain to the frequency domain. After the transformation, each sequence is represented as a superposition of multiple frequency components, specifically a set of frequency values ​​and their corresponding amplitude and phase information. To further simplify and extract key information, only the most important frequency components are retained in the transformation result; typically, the three frequency components with the largest amplitudes are selected as the core frequency domain features of the sequence. For the frequency domain features of each sequence, its dominant frequency (i.e., the frequency value with the largest amplitude) is extracted and recorded to reflect the main period of change of this feature; the amplitude corresponding to this frequency is recorded to represent the intensity of the periodic change; and the phase value is recorded to represent the offset position of the periodic signal on the time axis. These three types of frequency domain features are stored in the feature vector, corresponding to the original feature numbers. The above steps are repeated until all features have undergone Fourier transform and frequency domain feature extraction. Finally, the dominant frequency, amplitude, phase, and other information corresponding to all features are combined to form a complete "time-series variation feature set," which is used for cluster analysis and feature grouping in the next step. This process ensures that each feature is transformed into a structured feature representation with periodic expressive capabilities through a unified process, making it comparable and providing a basis for computational analysis.

[0031] For the aforementioned time-series variation feature set, cluster analysis was performed using the K-means clustering method, with the number of clusters selected between 3 and 8. Under each cluster number setting, the overall silhouette coefficient was calculated, and the cluster with the largest silhouette coefficient was selected as the number of groups. After determining the final number of clusters, the clustering process was executed, assigning all frequency domain features to different clusters based on Euclidean distance. The cluster centers were initially uniformly distributed and iteratively adjusted until convergence. After clustering, each frequency domain feature was assigned to its corresponding feature group, resulting in the final time-series variation feature groupings, which serve as the input basis for constructing the downstream interactive network.

[0032] S2 involves acquiring metabolomics and immunomics data from peripheral blood samples, separating metabolites and immune factors using liquid chromatography-mass spectrometry to obtain raw metabolic and immunological data. For these raw metabolic and immunological data, data standardization methods are used to eliminate dimensional differences, resulting in standardized metabolic and immunological data. Based on the standardized metabolic and immunological data, Pearson and Spearman rank correlation coefficients between metabolite concentrations and immune factor expression are calculated to obtain a preliminary correlation matrix. If the absolute value of the correlation coefficient in the preliminary correlation matrix is ​​greater than a preset threshold, the corresponding metabolite-immune factor pair is retained, resulting in a filtered correlation matrix. For the filtered correlation matrix, a weighted gene co-expression network analysis algorithm is used to construct a cross-omics interaction network, obtaining the interaction network topology. Based on the interaction network topology, a community detection algorithm is used to divide the network into modules to determine dynamic interaction patterns. For the dynamic interaction patterns, time series analysis is used to analyze the temporal correlation between metabolites and immune factors, obtaining a temporal interaction feature set.

[0033] In implementing this method, peripheral blood samples were first collected from diabetic patients. At least 5 ml of venous blood was collected from each patient, and after anticoagulation, plasma was extracted by centrifugation. The obtained plasma samples were then sent to a liquid chromatography-mass spectrometry (LC-MS) platform for the separation and detection of metabolites and immune factors. LC used a combination of stationary and mobile phases of different polarities to sequentially separate metabolites and immune factors in the plasma according to their retention time before entering the mass spectrometry system. The mass spectrometry portion employed electrospray ionization for ion detection. Quantitative and qualitative analysis of each component was performed based on its mass-charge ratio, generating raw metabolic and immunological data. Approximately 800 metabolite features and 100 immune factor features were obtained from each sample. All samples were arranged into a two-dimensional data matrix, with the horizontal axis representing the sample, the vertical axis representing the substance number, and each cell representing the detection intensity value. The raw data were then standardized by subtracting the mean from the value of each feature across all samples and then dividing by the standard deviation, converting all feature values ​​to data with a mean of 0 and a standard deviation of 1. This method can eliminate the deviations caused by differences in detection units, dimensions, or measurement scales between different features, ensuring that subsequent correlation calculations are comparable and stable.

[0034] After standardization, standardized metabolic and immune data were obtained, both in a two-dimensional matrix structure. The horizontal axis represents the patient sample number, and the vertical axes represent the metabolite and immune factor numbers, respectively. The value of each cell is the standardized intensity value of the corresponding sample under that feature. The calculation process is as follows: For each metabolite, its standardized concentration values ​​in all samples are iterated sequentially, and then paired with the standardized expression values ​​of each immune factor in all samples. For each pair, the Pearson correlation coefficient is calculated first. Specifically, the covariance of the two variables in the sample dimension is calculated, then the standard deviation of each variable is calculated separately, and finally the covariance is divided by the product of the two standard deviations to obtain the Pearson correlation coefficient value, which ranges from -1 to +1, representing the strength and direction of the linear correlation between the two variables. Subsequently, Spearman's rank correlation coefficient was calculated for the same variable pair. This process first sorts the values ​​of the two variables across all samples, assigning each value a rank number. Then, the covariance and standard deviation are calculated using the ranks to obtain the final rank correlation coefficient, which also ranges from -1 to +1. This coefficient is mainly used to identify non-linear but monotonic correlations. For a total of 800 metabolites and 100 immune factors, a total of 800 x 100 = 80,000 correlation calculations are required. The results are then entered into two preliminary correlation matrices: a Pearson correlation coefficient matrix and a Spearman rank correlation coefficient matrix. Rows represent metabolites, columns represent immune factors, and the corresponding cells record the correlation coefficient values ​​between the feature pairs. These two matrices are the preliminary correlation matrices, which will be used later to set thresholds for screening and network construction.

[0035] After standardization, the Pearson correlation coefficient and Spearman rank correlation coefficient were calculated between the standardized metabolic data and the standardized immune data. The former is suitable for identifying linear correlations between variables, calculated by dividing the covariance of the two variables by the product of their respective standard deviations, with the result ranging from -1 to +1. The latter is suitable for identifying nonlinear but monotonic ordered relationships between variables, by first ranking the variables and then calculating the correlation between their ordered values. For each pair of metabolite and immune factor combinations, both correlation coefficients were calculated and filled into a preliminary correlation matrix, forming a two-dimensional matrix with metabolites on the horizontal axis, immune factors on the vertical axis, and correlation coefficients in each cell. A correlation coefficient screening threshold of 0.7 was then set; if the absolute value of the Pearson or Spearman correlation coefficient is greater than or equal to 0.7, a significant statistical association is considered to exist between the metabolite and the immune factor. This threshold was determined by constructing networks under multiple thresholds and comparing the performance of each network in terms of module density, connection strength, and prediction stability. Finally, the threshold value of 0.7, which yielded the highest prediction consistency without sacrificing network sparsity, was selected as the standard. After screening, a correlation matrix is ​​constructed, and all metabolite-immune factor pairs that meet the criteria are retained.

[0036] Based on the selected association matrix, a weighted gene co-expression network analysis algorithm was used to construct a cross-omics interaction network. The specific steps included: First, constructing an adjacency matrix, where each pair of selected metabolites and immune factors was treated as a network node pair, with the weight of the edges between nodes being the absolute value of their correlation coefficients; Second, transforming the adjacency matrix into a topological overlap matrix, which comprehensively considers whether two nodes have common neighbors and their connection strength, used to measure the similarity between nodes; Third, drawing a complete interaction network graph based on the topological overlap matrix, where nodes represent metabolites or immune factors, edges represent the correlation between them, and the thickness of the edges represents their weight values. Subsequently, a community detection algorithm was used to partition the network graph into modules, selecting the partitioning method based on the maximum density principle. In the initial stage, all nodes were divided into several modules, and then through iterative optimization, the intra-module connections were gradually strengthened while the inter-module connections were weakened until convergence to a state with an optimal module structure. After partitioning, each module represents a dynamic interaction mode.

[0037] After completing the screening of the association matrix and constructing the interaction network topology, a weighted undirected network containing metabolite nodes and immune factor nodes has been formed. Nodes represent variables, edges represent significant statistical correlations between variables, and the weight of each edge is the absolute value of the correlation coefficient from the previous step. At this point, the network needs to be modularized to identify substructures with dense internal connections and sparse external connections, i.e., dynamic interaction patterns. The specific operation process is as follows: First, initialize the network structure by importing all nodes into the graph structure processing framework, retaining the category information (metabolite or immune factor) and edge weight information for each node. Next, select a community detection algorithm, preferably a partitioning method based on the principle of maximizing module density. The basic idea of ​​this algorithm is to group the network node set in different ways while maintaining the integrity of the network structure, making the connections between nodes within the same group more compact, while making the connections between different groups as sparse as possible.

[0038] In the specific execution process, all nodes are first treated as independent modules. Then, an iterative merging operation is performed. In each iteration, the module density index of the current partition state is evaluated. The density index is defined as the total strength of internal edges in all modules divided by the total number of possible internal edges in that module. Each merging operation attempts to move a node to the module with the strongest adjacent connection. It calculates whether the overall module density increases after the move. If it does, the merging action is retained; otherwise, the original state is restored, and the optimal assignment of the next node is evaluated. This process is repeated until the assignment relationships of all nodes remain unchanged in two consecutive iterations, or the module density no longer increases. This is considered convergence, and the final module partitioning is completed.

[0039] After partitioning, each module represents a dynamic interaction pattern. Nodes within a module represent metabolites and immune factors with close interactions within that module, and the internal structure of the module constitutes its dynamic interaction subnetwork. This module partitioning result can be used for subsequent time-series analysis, regulatory factor extraction, and mechanism identification, demonstrating good structural interpretability and biological visualization value. Each module contains at least 3 nodes, and the number of modules is automatically determined based on network complexity, generally between 3 and 10.

[0040] In each identified dynamic interaction pattern, its temporal interaction features are further extracted. The specific steps are as follows: extract the metabolites and immune factors belonging to each interaction module, and read their detection values ​​at at least five different time points to form time series data; then, construct time series for each pair of metabolites and immune factors at the same time points, and calculate their temporal correlation. Specifically, set the lag step size to a range of 0 to 3, calculate the Pearson correlation coefficient for each pair under each lag condition, and take the one with the largest absolute value as the temporal correlation strength of that variable pair. After processing all combinations of metabolites and immune factors in the network in this way, a temporal interaction feature set is finally formed, which is used to fully characterize the collaborative change patterns between variables in the dynamic interaction module over time, providing structured input for subsequent mechanism identification and network parameter optimization.

[0041] S3 includes: acquiring potential regulatory factors from dynamic interaction patterns; extracting factor features through principal component analysis to obtain a factor feature set; grouping the potential regulatory factors according to the factor feature set using hierarchical clustering to obtain factor grouping results; calculating the Pearson correlation coefficient between factors and metabolic disorder features within each group to obtain a correlation strength matrix; if the absolute value of the correlation coefficient in the correlation strength matrix exceeds a preset threshold, the corresponding factor is marked as a candidate regulatory factor to obtain a candidate regulatory factor set; constructing a prediction model of factors and metabolic disorders using a random forest algorithm based on the candidate regulatory factor set to obtain model output results; extracting the temporal correlation between candidate regulatory factors and metabolic disorders using time series analysis to obtain a temporal correlation feature set; and dividing the candidate regulatory factors into functional modules using a community detection algorithm based on the temporal correlation feature set to obtain a functional module set.

[0042] In this implementation, potential regulatory factors are first obtained from the dynamic interaction patterns identified in the previous stage. Specifically, the connectivity of nodes in each interaction module is calculated, and the top 20% of nodes by connectivity are selected as potential regulatory factors. These factors include metabolites and immune factors, indicating their strong influence or regulatory potential in the network. For these potential regulatory factors, standardized expression data at at least five time points are collected to form a factor sample data matrix. Principal component analysis is performed on this matrix. First, the covariance matrix is ​​calculated and eigenvalue decomposition is performed to extract all principal components and their contribution rates. The components are then sorted from highest to lowest contribution rate, and the top few principal components are cumulatively selected until their total contribution rate is greater than or equal to 85%. This proportion serves as the principal component selection criterion and is the optimal retention lower limit determined in pre-experiments by observing the impact of the number of principal components on model stability and dimensionality compression. Then, the principal component score of each factor on the retained principal components is calculated as the principal component feature of that factor, ultimately forming a factor feature set.

[0043] Subsequently, cluster analysis was performed on the factor feature set using a hierarchical clustering algorithm. First, the Euclidean distance between all factors in the principal component score space was calculated, constructing a distance matrix as the clustering input. The minimum distance method was used as the clustering criterion, progressively merging the two factors or cluster units with the smallest distance to construct a clustering tree structure. A grouping truncation height was set on this clustering tree, set to 40% of the total height of the clustering tree. This proportion was selected as the optimal boundary value after progressively testing clustering resolution and intra-group consistency in simulation experiments. The clustering tree was truncated according to this height, dividing it into multiple factor groups, forming the factor grouping results.

[0044] Next, correlation analysis was performed on the regulatory factors within each factor group to extract the Pearson correlation coefficient between each factor and key indicators representing metabolic disorders. The selected metabolic disorder indicators included glucose concentration, insulin resistance index, and glycated hemoglobin level, all using standardized expression values. Correlation calculations were performed on each factor individually. For each group of calculation results, a correlation strength matrix was established, with rows representing factors, columns representing metabolic indicators, and corresponding cells displaying the Pearson correlation coefficient values. A screening threshold of 0.75 was set; if the absolute value of the Pearson correlation coefficient between a factor and any metabolic indicator was greater than or equal to 0.75, the factor was considered to have a strong association with metabolic disorders and was included in the candidate regulatory factor set. This threshold was the optimal value selected after testing the biological consistency and network stability of the screening results under multiple candidate values, ensuring that the selected factors are both statistically significant and biologically plausible.

[0045] After obtaining the set of candidate regulatory factors, a predictive model was further constructed to verify the mapping relationship between these factors and metabolic disorder states. The model was constructed using a random forest algorithm, with standardized expression data of all candidate factors as input and metabolic disorder state labels as output. First, the number of decision trees in the random forest was set to 100, representing the optimal parameter corresponding to the lowest model error in the 5-fold cross-validation test. The training samples for each tree were randomly selected from the original sample using a bootstrap sampling method, with repeated sampling allowed. At each node of each tree, a total of the square root of the factors were randomly selected as candidate split variables from all input factors. For example, if the number of input variables was 36, 6 variables were selected for evaluation at each split point. After model training, the prediction results were output, and the feature importance of each factor was ranked, retaining the top 20% of the most important factors for time-series analysis.

[0046] Time-series analysis was performed on the selected factors to extract their temporal correlations with metabolic disorder indicators. This process involved extracting the expression values ​​of each factor and metabolic indicator at each time point from the sample time series, constructing a time-series data sequence. For each pair of sequences, Pearson correlation coefficients were calculated with lag steps of 0, 1, 2, and 3. The result with the largest absolute value was taken as the temporal correlation strength, and its corresponding lag period and correlation direction were recorded to obtain the temporal expression characteristics of the factor. The temporal correlation results of all factors were summarized to form a temporal association feature set.

[0047] Finally, a factor network is constructed using this feature set, where nodes are candidate factors and edges represent the similarity between temporal expression trends. Nodes are considered to have similar temporal patterns and are connected as a pair if they are completely consistent in lag time and correlation direction, and the correlation difference is less than 0.1. After construction, a community detection algorithm based on the principle of maximizing module density is used to divide functional modules. Finally, candidate factors with highly similar expression patterns are aggregated into the same functional module, forming a functional module set.

[0048] S4 includes extracting relevant signaling pathways from a candidate regulatory factor set, matching them with a pre-defined pathway database using pathway enrichment analysis to obtain a pathway activation level set; calculating the association strength between each pathway and cognitive function indicators using partial correlation analysis based on the pathway activation level set to obtain an association strength set; if the absolute value of the association coefficient in the association strength set exceeds a pre-defined threshold, verifying the interaction relationship between pathways and cognitive function indicators using mutual information calculation to obtain a preliminary interaction relationship set; constructing an interaction network between pathways using network topology analysis for the preliminary interaction relationship set to obtain a network connection strength set; dividing the pathways into functional modules using a community detection algorithm based on the network connection strength set to obtain a functional module set; extracting the temporal correlation between each module and cognitive function indicators using time series analysis for the functional module set to obtain a temporal correlation feature set; and predicting the influence weight of functional modules on cognitive function using a logistic regression model based on the temporal correlation feature set to obtain a description of the interaction driving mechanism.

[0049] In this embodiment, firstly, all factors with clear gene identifiers or metabolite codes are extracted from the screened set of candidate regulatory factors, and these factors are mapped and matched with each signaling pathway recorded in a pre-defined pathway database. The database used is a structured pathway database containing no fewer than 1000 signaling pathways related to human metabolism, immunity, and neurological function. Each pathway entry records a list of its constituent factor names, and the matching process involves counting the number of factors matched by candidate factors for each pathway. To determine whether a pathway is enriched among candidate factors, a hypergeometric test is used for pathway enrichment analysis, i.e., the probability of each pathway appearing is calculated based on the number of pathway factors, the number of candidate factors, and the number of intersection factors in the entire background database. Multiple hypothesis correction is applied to the p-values ​​corresponding to all pathways using the Belfani method, with a corrected p-value threshold of 0.05. This threshold is determined by controlling the false positive rate to no more than 5% in a simulation experiment. All pathways that meet the above conditions are retained as effective enriched pathways. For each enriched pathway, the expression data of all the hit regulatory factors are extracted, and their mean is calculated according to the sample dimension. This means the activation level of the pathway in that sample. The pathway activation levels of all samples are summed to form the pathway activation level set.

[0050] Subsequently, the cognitive function indicators measured in each sample were processed. The main indicators included memory test scores, attention test scores, and overall cognitive scores. These data were standardized to have a mean of 0 and a standard deviation of 1. To eliminate the influence of confounding variables, partial correlation analysis was used to calculate the true association between pathway activation levels and cognitive function indicators. Confounding variables included the patient's age, gender, body mass index, and baseline blood glucose level. The procedure was as follows: First, for each pair of pathway activation levels and cognitive function indicators, a multiple linear regression model was constructed with the confounding variables as independent variables, and the residual series were calculated. Then, the Pearson correlation coefficient was calculated for the two residual series; this coefficient is the net correlation coefficient after removing the confounding influence. The net correlation coefficients calculated for all pathways and each cognitive indicator were summarized to form a set of association strengths. A screening threshold of 0.6 was set. This value was used in the previous modeling experiment as an evaluation standard based on the pathway significance recognition rate. By comparing the goodness of fit between the recognition results and cognitive change trends under different thresholds, the minimum acceptable value was determined under the condition of ensuring a balance between recognition ability and biological rationality.

[0051] For all pathways with an absolute correlation coefficient greater than or equal to 0.6, mutual information calculation is performed to verify the existence of nonlinear interactions. Mutual information calculation uses a discretization method, dividing the expression value of each variable into five equidistant intervals, constructing a joint frequency table, calculating the marginal probability and joint probability of each variable, and substituting them into the information entropy formula to calculate the mutual information value. To determine the significance of the mutual information value, 1000 permutation tests are performed. Each time, the cognitive indicators are randomly shuffled and rearranged among the samples, and the mutual information value is recalculated to form a background distribution. The 95th percentile of this background distribution is used as the threshold; if the original mutual information value is greater than this value, a significant information interaction relationship is considered to exist between the pathway and the cognitive indicator. All pathway pairs that meet this condition are retained, forming a preliminary interaction set.

[0052] Based on the initial set of interaction relationships, an inter-pathway interaction network is constructed. Network nodes represent pathways, and edges indicate shared moderating factors or overlapping cognitive indicator interaction information between two pathways. Edge weights are set as the sum of the weighted average of the number of shared factors and their corresponding mutual information values. After calculation, a weighted undirected graph structure is constructed to form the interaction network topology. The weight value of each edge is calculated, and all edge weights are aggregated to form a network connection strength set. Subsequently, a topological analysis is performed on the network, including calculations of node degree, edge density, and module compactness, for subsequent community partitioning optimization.

[0053] The network is input into a community detection algorithm, and functional modules are partitioned using the principle of maximizing module density. Initially, each node is treated as a separate module. In each iteration, it is calculated whether merging two modules increases the overall network module density. Module density is calculated as the total weight of edges within a module divided by the maximum possible number of connections for that module. Merging operations that increase module density are retained, and this process is repeated until the density of all modules no longer increases, thus determining the set of functional modules.

[0054] For each set of functional modules, the activation levels of all pathways within each module at different time points were extracted, along with the corresponding cognitive index values, to construct time series data. Using time lag analysis steps of 0, 1, 2, and 3, Pearson correlation analysis was performed on each module and the cognitive index. The module with the largest absolute correlation coefficient under the lag condition was taken as the final time-series correlation strength, while also recording its lag point location and correlation direction. The time-series results for all modules were summarized into a time-series association feature set.

[0055] Finally, this feature set was used as input to the logistic regression model, with the module activation level and its lag information as independent variables, and standardized cognitive scores as the dependent variable. Maximum likelihood estimation was used for regression model training, with a significance level set at 0.05. If the p-value of the regression coefficient corresponding to a module was less than this value and the regression coefficient was not zero, it was considered to have a significant impact on cognitive function. The sign of the regression coefficient output by the model indicates the direction of modulation, and the absolute value indicates the intensity of modulation. This ultimately forms a description of the interaction driving mechanism between functional modules and cognitive function, used to support intervention design and target recommendation.

[0056] S5 includes extracting an initial factor set from the candidate adjustment factor set, using the k-means clustering algorithm for preliminary grouping to obtain initial factor groups; if the intra-cluster variance of the initial factor groups exceeds a preset threshold, the grouping boundary is optimized using an iterative threshold adjustment method to obtain an optimized factor set; based on the optimized factor set, the signal-to-noise ratio (SNR) of the factor signals is calculated using a noise identification standard; if the SNR is lower than a preset standard, noise interference factors are removed to obtain a preliminary refined factor set; for the preliminary refined factor set, the DBSCAN algorithm is used to evaluate the clustering stability to obtain a stability index set; if any index in the stability index set is lower than a preset standard, residual analysis is used to calculate... Calculate factor residuals, remove outliers, and obtain a modified factor set. Based on the modified factor set, use dynamic time warping to align the factor time series, resulting in an aligned factor set. For the aligned factor set, calculate the matching degree measure between the factors and a preset pathway database, resulting in a matching degree set. If any value in the matching degree set is lower than a preset standard, return to the modified factor set and re-align the time series, resulting in an updated aligned factor set. Based on the updated aligned factor set, use principal component analysis to extract principal component features, resulting in a feature factor set. For the feature factor set, verify its correlation strength with cognitive function indicators. If the correlation strength reaches a preset standard, output a refined moderating factor list.

[0057] In this implementation, firstly, all factors with complete data records across all sample time points are extracted from the candidate modulating factor set to form an initial factor set. To ensure data quality, the expression value of each factor is standardized across all time points by subtracting its mean and dividing by its standard deviation, ensuring it follows a distribution with a mean of 0 and a standard deviation of 1. The standardized factor expression data is used as input, and k-means clustering is used for initial grouping. The initial number of clusters K is set to 5, and Euclidean distance is used to calculate the similarity between factors during the clustering process. To determine if the K value is reasonable, the total sum of squares within each cluster is calculated for each K value, and a curve is plotted. The elbow rule is applied to identify the inflection point as the optimal K value. After clustering, the intra-cluster variance is calculated for each factor within each cluster. The intra-cluster variance is defined as the average of the sum of squared distances from all factors in the cluster to the cluster centroid, representing the degree of consistency within the cluster. The intra-cluster variance threshold is set to 1.5. This value is an empirical threshold derived by repeatedly testing and comparing the stability and accuracy of the clustered model on multiple training datasets, ensuring good internal structural consistency in the clustering. If the variance within any cluster exceeds the threshold, an iterative optimization process is executed. By adjusting the K value, resetting the initial centroid, and repeating the clustering operation, the variance within all clusters is less than or equal to 1.5, resulting in the optimized factor set.

[0058] Next, the signal-to-noise ratio (SNR) of each factor in the optimized factor set was calculated based on its expression data at different time points. The SNR was defined as the mean of each factor divided by its standard deviation, where the mean represents signal strength and the standard deviation represents the degree of fluctuation. A SNR threshold of 1.2 was set, derived by analyzing the relationship between SNR and prediction error, ensuring the removal of interfering factors with excessive fluctuations that could not stably reflect biological processes. All factors with an SNR below 1.2 were removed, and the remaining factors constituted the initial refined factor set. To evaluate the clustering stability of these factors, density clustering was performed using the DBSCAN clustering algorithm, with parameters set to a minimum neighborhood radius of 0.5 and a minimum number of points of 5. This setting was determined based on the actual sample size and factor dimensionality, enabling the identification of high-density regions while avoiding overly fragmented clustering. After clustering, the silhouette coefficient of each factor was calculated. This value reflects the distance difference between the factor and its current cluster and its nearest cluster, and is used to evaluate the clustering effect. Factor clusters with a silhouette coefficient below 0.4 were considered to have poor clustering stability. To further optimize factor quality, residual analysis was used to sum the squared residuals of each factor’s expression value in different samples and the value of its cluster center. After sorting the sums of squared residuals of all factors, the top 10% of factors with the largest residuals were removed to obtain the corrected factor set.

[0059] The modified factor set is input into a dynamic time warping algorithm for temporal alignment. The method involves calculating the distance matrix between each pair of samples in the time dimension and finding the alignment path with the minimum cost, ensuring optimal temporal alignment of the expression patterns of all factors across different samples. The output aligned factor set is used for further functional matching analysis. The temporal expression trend of each aligned factor is compared with the expression patterns of key pathways in a pre-defined pathway database. Both Euclidean distance and Pearson correlation are used for calculation, mapping the results to a matching degree value ranging from 0 to 1. A higher matching degree indicates a greater consistency between the factor and pathway expression behavior. A matching degree threshold of 0.7 is set, based on the consistency analysis between pathway prediction accuracy and biological validation experiments, ensuring that only factors with a high degree of consistency with actual pathway mechanisms are retained. If the matching degree of any factor is lower than 0.7, the algorithm reverts to the modified factor set, re-executes the dynamic time warping operation, and generates an updated aligned factor set to guarantee a high degree of matching between factor expression behavior and pathway characteristics.

[0060] Principal component analysis (PCA) was performed on the updated aligned factor set. The steps included calculating the covariance matrix among all factor expression values, performing eigenvalue decomposition, sorting the eigenvalues ​​from largest to smallest, and selecting the top principal components with a cumulative contribution rate of 85% to form the principal component space. The coordinates of each factor in the principal component space were used as the principal component eigenvectors, forming a feature factor set. Next, correlation analysis was performed between each factor in the feature factor set and cognitive function indicators, calculating their Pearson correlation coefficients to assess their responsiveness to changes in cognitive function. A correlation threshold of 0.65 was set, derived from a trade-off analysis between model accuracy and biological interpretability. Factors with an absolute correlation coefficient greater than or equal to 0.65 were considered to have a significant statistical correlation with cognitive function indicators, ultimately forming a refined list of moderating factors.

[0061] S6 involves extracting interaction network data from a refined list of regulatory factors, constructing an initial network structure using a graph neural network algorithm to obtain a network node set; calculating the interaction strength between nodes based on the network node set, evaluating the correlation between nodes using the Pearson correlation coefficient method to obtain an association matrix; if the correlation value in the association matrix is ​​lower than a preset threshold, the corresponding node is removed, the network node set is updated, and a refined node set is obtained; for the refined node set, a community detection algorithm is used to divide the network into related sub-networks of immune factors and glucose transporters, obtaining a sub-network set; based on the sub-network set, the interaction path between cytokine storm and glucose transporter dysregulation is extracted, the path action weights are calculated, and a weight vector is obtained; for the weight vector, principal component analysis is used to extract the main interaction features, obtaining a feature vector; if the correlation strength between the feature vector and the diabetic state reaches a preset standard, a comprehensive interaction model is output, and the model parameters are obtained.

[0062] In this embodiment, the expression time series data of all factors and their connections identified in omics association analysis are first extracted from the aforementioned refined list of regulatory factors, serving as the data source for the interactive network. Based on this data, a graph neural network model is constructed to capture the complex nonlinear relationships between factors. Specifically, each regulatory factor is used as a network node, and the input features of the node include its standardized expression values ​​at different time points. The initial connections of the network are formed by adjacency matrices composed of significant associated edges retained during the data mining phase. The graph neural network adopts a full-graph update mode. During each layer propagation, each node weighted and aggregates its features with the features of its neighboring nodes. This process achieves information transformation through a weight matrix and an activation function. The number of network propagation layers is set to 2. This parameter is the optimal value selected after cross-validating the model training accuracy at different propagation depths, ensuring sufficient information fusion while avoiding excessive smoothing. During training, an unsupervised structure is used to maintain the objective function. After 100 training epochs and error convergence, the embedding vector of each node is output, forming a network node set.

[0063] Subsequently, Pearson correlation coefficients were calculated for each pair of network nodes. This coefficient measures the consistency of expression behavior among nodes, ranging from -1 to +1, and the results form a symmetric correlation matrix. To remove nodes with significantly different expression patterns and no interaction, the correlation coefficients of each node in the correlation matrix with all other nodes were averaged. If the average value was below 0.4, the node was considered a marginal node. This 0.4 threshold was derived from a comparison of model recognition rate, stability, and generalization ability under different correlation standards in model tuning experiments. The results showed that the model structure was clearest and the recognition efficiency was optimal at a correlation coefficient of 0.4. Therefore, all nodes with an average correlation coefficient below 0.4 were removed, and the adjacency matrix was updated to form a refined node set.

[0064] Next, a community detection algorithm is applied to the refined node set for subnetwork partitioning. This algorithm performs network clustering based on the principle of maximizing module density, automatically grouping closely connected nodes into the same module. The module resolution parameter is set to 1.0, a recommended value for maintaining the balance of the subnetwork structure, which can identify the core structure while avoiding over-subdivision. After partitioning, subnetworks containing both immune factor and glucose transporter nodes are selected to form a subnetwork set. For each subnetwork, a path search operation is performed, using depth-first search to identify all direct and indirect paths from immune factor nodes to glucose transporter nodes. The weight of each edge in the path is calculated by fusing two pieces of information: first, the edge attention value learned during graph neural network propagation, reflecting the local network propagation strength; second, the correlation coefficient between the expression values ​​of the two nodes connected by the edge, reflecting the expression coupling strength. The weights are weighted and averaged to obtain the edge weight value, and the average of all edge weights in the path is calculated to form the influence strength of each path. The influence strengths of all paths form a weight vector.

[0065] The weight vectors are used as feature inputs, and principal component analysis (PCA) is employed for dimensionality compression to extract key interaction features. First, the covariance matrix of the weight vectors is calculated, followed by eigenvalue decomposition. The eigenvalues ​​are sorted, and the top principal components with a cumulative contribution rate of 85% are retained as feature vectors. This contribution rate threshold is determined based on experiments balancing the explanatory power and compression efficiency of multiple models. A one-to-one correspondence analysis is performed between the obtained feature vectors and diabetes status indicators. The Pearson correlation coefficient between each feature component and diabetes status is calculated, and a two-tailed significance test is performed to determine its statistical validity. A correlation threshold of 0.65 and a p-value significance threshold of 0.05 are set; this combination represents an empirical standard obtained when predictive power and statistical confidence are optimal at different correlation levels. If the correlation coefficient between any principal component feature and diabetes status is greater than or equal to 0.65, and the p-value is less than 0.05, then the feature is considered significantly correlated and further incorporated into the final output integrated interaction model, with its regression coefficient in logistic regression analysis used as a model parameter. This integrated interaction model can be used to predict whether an individual is at high risk of diabetes-related cognitive impairment. It can also quantify the regulatory pathways and interaction weights between cytokine storms and abnormal glucose transport, thereby enabling mechanism analysis and the generation of personalized intervention recommendations.

[0066] S7 specifically includes: obtaining dynamic change data of cognitive function from model parameters; constructing an initial trend model using time series analysis to obtain a preliminary trend vector; extracting key change points in the time dimension based on the preliminary trend vector; grouping change features using cluster analysis to obtain categorical change sets; generating heatmap visualization data for the categorical change sets; mapping change intensity using color coding to obtain a visualization matrix; extracting change regions from the visualization matrix; calculating the deviation value from patient data using multivariate regression to obtain a deviation statistic; verifying model consistency using logistic regression if the deviation statistic is below a preset threshold to obtain verification results; adjusting trend model parameters based on verification results; updating the interaction mechanism using iterative optimization to obtain an optimized interaction mechanism; extracting core interaction paths from the optimized interaction mechanism; quantifying path weights using path analysis to obtain a final weight vector.

[0067] In this implementation, the most critical model parameter set for predicting cognitive function is first extracted from the aforementioned integrated interaction model, including principal component scores, pathway weight coefficients, and core pathway expression trends. These parameters are then combined into a dynamic time series based on time points. To simulate the evolution of diabetes-related cognitive function, a time series modeling method is employed. First, a sliding window technique is used to analyze time-span behavior, with a window length set to five consecutive time points. This length was determined to be optimal in preliminary experiments by comparing the impact of the sliding window length on trend smoothness and response sensitivity. Within each window segment, the data sequence is first mean-filtered to reduce noise, and then the rate of change within that window is calculated as the average of the differences between the values ​​at subsequent and previous time points. The rates of change for all windows are arranged chronologically to form a preliminary trend vector. Subsequently, positions within this vector where the absolute value of the rate of change exceeds the vector's mean plus twice the standard deviation are identified as significant change points. This thresholding method is based on statistical analysis of the distribution of change amplitudes in the initial trend vector to ensure that truly significant fluctuation regions are screened out. A change feature set is constructed by selecting the time points and change values ​​corresponding to the selected change points. A density clustering algorithm is then applied to this feature set to group similar change patterns into groups. The cluster radius is set to 0.3 and the minimum cluster size to 3. These parameters are determined based on an assessment of the density of the sample change patterns to ensure that the clustering results are neither diluted nor overly subdivided. The clustering results constitute a categorized change set, representing different pattern categories in the evolution of cognitive function.

[0068] Next, a heatmap-style visualization matrix is ​​generated for the categorical change set. Rows correspond to different change categories, columns to time points, and cell values ​​are normalized change intensity values. Normalization linearly maps the maximum and minimum values ​​of the change feature set to a range of 0 to 1, with the mapping color transitioning from light blue to dark red to indicate progressively increasing change intensity. This visualization matrix is ​​used to identify significant hotspots. Further extraction of the time points and category combinations corresponding to these hotspots is used to locate the alignment period between the simulated trend and real patient data. After alignment, a multivariate regression model is constructed, using the model's simulated output value and principal component scores as independent variables, and the real patient cognitive assessment scores as the dependent variable. Stepwise regression analysis is used to estimate the coefficients of each variable. The regression residuals are first used to calculate the prediction error at each time point, squared, and then averaged as the bias statistic. A bias threshold of 0.15 is set. This value is derived from comparing the simulation results of the training set with real data; that is, when the average error is below 0.15, the deviation between the model prediction and the actual data is within a physiologically acceptable range. If the bias statistic meets this standard, logistic regression validation is performed. The model's predicted classification trend and the actual cognitive state labels are input into the logistic regression, and the model accuracy is calculated. Here, the accuracy threshold is set at 85%, based on the judgment standard that most trained models still maintain high recall and high precision under this standard. If it is met, the trend model is considered to have statistical consistency and is acceptable.

[0069] After successful model validation, the iterative optimization phase begins. Based on error feedback, the path weights and principal component coefficients in the initial trend model are cyclically adjusted. The optimization method involves incrementally adjusting the direction of the path weights. After each round of optimization, the bias statistic and logistic regression accuracy are recalculated, and the results are compared to determine whether further adjustments should continue. Optimization stops when neither significantly improves. The resulting structure is the optimized interaction mechanism. Subsequently, core interaction paths are selected from the optimized mechanism. The selection criteria are that the path edge weight must be higher than the 75th percentile of all path weights in the mechanism, and the path must be frequently activated in all samples, with a frequency standard of appearing in at least 80% of the sample paths. Path analysis is performed on the selected paths to evaluate their explanatory power for cognitive function scores. Specific steps include fixing other paths as control variables in the model, adding a single path predictor variable, and comparing the change in model fit to quantify the independent contribution of the path. The change in fit values ​​for each path constitutes the final path weight vector. This weight vector summarizes how immune factors and glucose transport mechanisms synergistically influence cognitive changes in a diabetic state. It is an important finding of the finally validated interaction mechanism and has practical value in guiding individualized intervention strategies.

[0070] This also includes S8, extracting obscure micro-regulatory subnetworks from the validated interaction mechanism, refining subnetwork nodes using clustering stability analysis, and determining the fine-grained driving path for diabetes-induced memory decline if the matching degree between subnetwork nodes and temporal changes passes the preset standard. Specifically, this involves extracting regulatory subnetworks from the optimized interaction mechanism, separating subnetwork nodes using network decomposition methods, obtaining the temporal correlation strength between nodes to obtain a node association set; extracting strong association pairs from the node association set, and grouping node features using k-means clustering to obtain node feature groups; if the temporal change matching degree of the node feature groups passes the preset standard... For preliminary verification, the sub-network structure is refined using path clustering to obtain the initial driving path. Key path nodes are extracted from the initial driving path, and path weights are calculated using weighted network analysis to obtain the path weight vector. If the stability of the path weight vector passes the preset threshold verification, the path structure is adjusted by the correlation strength between nodes to obtain the optimized driving path. Core driving nodes are extracted from the optimized driving path, and the dynamic interaction between nodes is quantified using time series analysis to obtain the dynamic interaction matrix. Interaction regions are extracted from the dynamic interaction matrix, and cluster stability analysis is used to verify the matching degree between the regions and diabetes-induced memory decline to obtain the driving path.

[0071] In one possible implementation, firstly, the information of the regulatory sub-networks in all regulatory paths is extracted from the aforementioned optimized interaction mechanism, and the factor nodes existing on the path are used as the initial node set of the regulatory sub-networks. Then, for each pair of nodes in the regulatory sub-network, the standardized expression value sequence at all time points is extracted, Pearson correlation coefficients at multiple time lags are calculated, and the maximum value is taken as the temporal association strength of the node pair. The maximum correlation coefficients of all node pairs form the node association set. Next, the node pairs with the strongest association strength are selected from the node association set as strong association pairs. The selection criterion is that the absolute value of the association coefficient is higher than the median of all coefficients plus twice the absolute deviation of the median; this threshold setting is based on the statistical distribution of node correlation to ensure that node pairs with significant temporal synchronization relationships are selected. After obtaining the strong association pairs, the nodes are grouped using the k-means clustering method. The clustering input is the expression value of the node at different time points. The initial number of clusters is set to 4, and the elbow rule is used to determine the final number of groups to ensure the consistency of the node expression patterns within each group. After obtaining the node feature groups, the time series of each group of nodes is matched one by one with the aforementioned classification change set, and the average maximum Pearson correlation coefficient is calculated. If the value is higher than 0.7, it is considered to be highly consistent with the cognitive function change pattern. This matching threshold is a boundary value determined by comparing the correlation distribution between the classification change set and cognitive indicators on multiple training samples. Node groups that meet this matching criterion are used for path clustering operations, that is, based on the original adjustment path, nodes in the same group are integrated to construct a new fine sub-path structure. All node and edge information is extracted from the newly generated preliminary driving path, and the weight of each path is calculated. The path weight is a weighted sum of the average expression of nodes in the path, the path length, and the edge connection strength. All path weight values ​​form a path weight vector. Further, the intra-sample repeatability stability of each path in the path weight vector is compared with other paths. If the weight of a certain path is above the 75th percentile of all path weights in multiple samples, and the standard deviation does not exceed 10% of the average weight, then the path is considered stable. This 75th percentile as the stable path threshold is determined by the quantitative evaluation results of the impact of paths on the model prediction accuracy in experiments. The stable path will then proceed to the structure optimization step. This involves analyzing the residuals between each pair of nodes, fine-tuning the directionality and strength of the connecting edges, removing redundant connections, and retaining the path structure that contributes most to the model. After optimization, key nodes are extracted from the final driving path—nodes with out-degree and in-degree numbers higher than the average. The time series of these nodes are then subjected to correlation analysis under multiple time lags to generate a dynamic interaction matrix. Each element in this matrix represents the interaction strength between two nodes at a given lag time point. Multiple consecutive highly correlated regions under different time lags are then identified from the matrix as candidate interaction regions.Subsequently, cluster stability analysis was performed on the nodes in this region, comparing the degree of overlap between these regions and real-world memory decline data induced by diabetes in terms of temporal patterns. A successful match was considered achieved if the average overlap rate was greater than 80%. This 80% matching threshold was derived from correlation analysis experiments between different matching regions and prediction accuracy during the earlier modeling phase. Ultimately, the path that met the four criteria of stability, synchronicity, path importance, and biological consistency was identified as the fine-tuning driving path for identifying diabetes-induced cognitive impairment. This path serves as a highly reliable biomarker network in subsequent prediction and intervention mechanisms.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting cognitive dysfunction in diabetes based on metabolomics analysis, characterized by, The method comprises the following steps: S1, obtaining metabolomics data and immunomics data from peripheral blood samples of diabetes patients, extracting time series data related to glucose metabolism, processing the time series data by using a time series analysis algorithm to obtain time series variation characteristics; S2, constructing a cross-omics interaction network according to the time series variation characteristics, integrating the correlation between metabolite concentration and immune factor expression, and determining the dynamic interaction mode; S3, extracting the regulatory factors from the dynamic interaction mode, grouping the factors by using a clustering method, and if the correlation strength between the factors in the group and the metabolic disorder exceeds a preset threshold, marking the factors as candidate regulatory factors to obtain a candidate regulatory factor set; S4, analyzing the related signal pathways for the candidate regulatory factor set, matching the pre-established pathway database through pathway enrichment analysis, obtaining the pathway activation level, evaluating the correlation between the pathway and the cognitive function index by using partial correlation analysis and mutual information calculation, and if the significance test shows that the correlation coefficient is higher than the correlation strength threshold, confirming the interaction driving mechanism, and obtaining the driving mechanism description; S5, optimizing the factor extraction accuracy according to the driving mechanism description, removing the factors disturbed by noise through iterative threshold adjustment and noise recognition standard, processing the remaining factors by using residual analysis and time series alignment method, and if the clustering stability and matching degree measurement reach the preset standard verification after the filtering cycle, outputting the regulatory factor list; S6, integrating the details of the interaction network from the regulatory factor list to generate a comprehensive interaction model for depicting specific immune factors in the state of diabetes, and obtaining model parameters.

2. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that: The S1 comprises: extracting metabolomics data and immunomics data from peripheral blood samples by using liquid chromatography mass spectrometry technology to obtain original metabolic data and original immune data; According to the original metabolic data and the original immune data, removing noise and outliers by data cleaning to obtain cleaned metabolic data and cleaned immune data; For the cleaned metabolic data and the cleaned immune data, extracting features related to glucose metabolism by using principal component analysis to obtain a metabolic feature set and an immune feature set; If the number of features in the metabolic feature set and the immune feature set is greater than a preset threshold, filtering the features highly related to glucose metabolism by correlation analysis to obtain a filtered feature set; According to the filtered feature set, processing the time series data by using an autoregressive moving average model to obtain a time series model; By using Fourier transform, extracting time series variation characteristics from the time series model to obtain a time series variation characteristic set; For the time series variation characteristic set, grouping the features by using clustering analysis method to obtain time series variation characteristic grouping.

3. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that: The S2 comprises: obtaining metabolomics data and immunomics data from peripheral blood samples, separating metabolites and immune factors by liquid chromatography mass spectrometry technology to obtain original metabolic data and original immune data; For the original metabolic data and the original immune data, processing by using a data standardization method to eliminate dimension differences to obtain standardized metabolic data and standardized immune data; According to the standardized metabolic data and the standardized immune data, calculating the Pearson correlation coefficient and the Spearman rank correlation coefficient between metabolite concentration and immune factor expression to obtain a preliminary correlation matrix; If the absolute value of the correlation coefficient in the preliminary correlation matrix is greater than the preset threshold value, the corresponding metabolite and immune factor pair is retained to obtain a screened correlation matrix; For the screened correlation matrix, a weighted gene co-expression network analysis algorithm is used to construct a cross-omics interaction network to obtain an interaction network topology; According to the interaction network topology, a network module is divided by a community detection algorithm to determine a dynamic interaction mode; For the dynamic interaction mode, time series analysis is used to extract the time sequence correlation of metabolites and immune factors to obtain a time sequence interaction feature set.

4. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that: The S3 includes: From the dynamic interaction mode, a potential regulatory factor is obtained, a principal component analysis is used to extract a factor feature to obtain a factor feature set; According to the factor feature set, a hierarchical clustering method is used to group the potential regulatory factors to obtain a factor grouping result; For the factor grouping result, the Pearson correlation coefficient of the factors in each group and the metabolic disorder characteristics is calculated to obtain a correlation strength matrix; If the absolute value of the correlation coefficient in the correlation strength matrix exceeds the preset threshold value, the corresponding factor is marked as a candidate regulatory factor to obtain a candidate regulatory factor set; According to the candidate regulatory factor set, a random forest algorithm is used to construct a factor and metabolic disorder prediction model to obtain a model output result; For the model output result, time series analysis is used to extract the time sequence correlation of the candidate regulatory factors and the metabolic disorder to obtain a time sequence correlation feature set; According to the time sequence correlation feature set, the functional modules of the candidate regulatory factors are divided by a community detection algorithm to obtain a functional module set.

5. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that: The S4 includes: From the candidate regulatory factor set, relevant signaling pathways are extracted, and a pathway enrichment analysis is used to match a preset pathway database to obtain a pathway activation level set; According to the pathway activation level set, a partial correlation analysis is used to calculate the correlation strength of each pathway and the cognitive function index to obtain a correlation strength set; If the correlation coefficient in the correlation strength set exceeds the preset threshold value, the interaction relationship between the pathway and the cognitive function index is verified by mutual information calculation to obtain a preliminary interaction relationship set; For the preliminary interaction relationship set, an interaction network between the pathways is constructed by network topology analysis to obtain a network connection strength set; According to the network connection strength set, the functional modules of the pathways are divided by a community detection algorithm to obtain a functional module set; For the functional module set, the time sequence correlation of each module and the cognitive function index is extracted by time series analysis to obtain a time sequence correlation feature set; According to the time sequence correlation feature set, a logistic regression model is used to predict the influence weight of the functional module on the cognitive function to obtain an interaction driving mechanism description.

6. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that: The S5 includes: From the candidate regulatory factor set, an initial factor set is extracted, and a k-means clustering algorithm is used for preliminary grouping to obtain an initial factor grouping; If the intra-cluster variance of the initial factor grouping exceeds the preset threshold value, the grouping boundary is optimized by an iterative threshold adjustment method to obtain an optimized factor set; According to the optimized factor set, a noise recognition standard is used to calculate the signal-to-noise ratio of the factor signal, and if the signal-to-noise ratio is lower than the preset standard, noise interference factors are removed to obtain a preliminary refined factor set; For the preliminary refined factor set, a DBSCAN algorithm is used to evaluate the clustering stability to obtain a stability index set; If any index in the stability index set is lower than the preset standard, calculate the factor residual by residual analysis method, eliminate abnormal factors, and obtain the corrected factor set; According to the corrected factor set, use dynamic time warping method to align the time sequence of the factors, and obtain the aligned factor set; According to the aligned factor set, calculate the matching degree of the factors and the preset pathway database, and obtain the matching degree set; If any value in the matching degree set is lower than the preset standard, return the corrected factor set to perform time sequence alignment again, and obtain the updated aligned factor set; According to the updated aligned factor set, extract the principal component features by principal component analysis, and obtain the feature factor set; According to the feature factor set, verify the correlation strength between the feature factor set and the cognitive function index, and if the correlation strength reaches the preset standard, output the refined regulation factor list.

7. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that: The S6 comprises: Extract the interaction network data from the refined regulation factor list, construct an initial network structure using a graph neural network algorithm, and obtain a network node set; According to the network node set, calculate the interaction strength between the nodes, evaluate the correlation between the nodes using the Pearson correlation coefficient method, and obtain a correlation matrix; If the correlation value in the correlation matrix is lower than the preset threshold, eliminate the corresponding node, update the network node set, and obtain a refined node set; According to the refined node set, divide the immune factors and glucose transporter related subnetworks using a community detection algorithm, and obtain a subnetwork set; According to the subnetwork set, extract the interaction path between the cytokine storm and the glucose transporter disorder, calculate the path weight, and obtain a weight vector; According to the weight vector, extract the main interaction features using the principal component analysis method, and obtain a feature vector; If the correlation strength between the feature vector and the diabetes state reaches the preset standard, output the comprehensive interaction model, and obtain the model parameters.

8. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that, Further comprising S7, simulating the cognitive dysfunction process according to the model parameters, tracking the change trend by parameter adjustment, visualizing the trend using a heat map and verifying the deviation between the trend and the actual patient data using multivariate adjustment, if the deviation is lower than the preset threshold, confirming the effectiveness of the model, obtaining the verified interaction mechanism, specifically comprising: Obtain the dynamic change data of cognitive function from the model parameters, construct an initial trend model using time series analysis method, and obtain a preliminary trend vector; According to the preliminary trend vector, extract the key change points in the time dimension, group the change characteristics using clustering analysis method, and obtain a classified change set; According to the classified change set, generate heat map visualization data, map the change strength using color coding method, and obtain a visualization matrix; Extract the change region from the visualization matrix, calculate the deviation value from the patient data using multivariate regression method, and obtain a deviation statistic; If the deviation statistic is lower than the preset threshold, verify the consistency of the model by logistic regression method, and obtain a verification result; According to the verification result, adjust the trend model parameters, update the interaction mechanism using iterative optimization method, and obtain an optimized interaction mechanism; Extract the core interaction path from the optimized interaction mechanism, quantify the path weight using path analysis method, and obtain the final weight vector.

9. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 1, characterized in that, S8. Extracting the isolated micro-regulatory sub-network from the verified interaction mechanism, refining the sub-network nodes by clustering stability analysis, and determining the fine driving path for diabetes-induced memory decline if the matching degree of the sub-network nodes and the time sequence changes passes the preset standard verification, specifically including: Extracting the regulatory sub-network from the optimized interaction mechanism, separating the sub-network nodes by network decomposition method, obtaining the time sequence correlation strength between nodes, and obtaining the node correlation set; Extracting the strength correlation pairs from the node correlation set, grouping the node features by k-means clustering method, and obtaining the node feature grouping; If the time sequence change matching degree of the node feature grouping passes the preset standard verification, refining the sub-network structure by path clustering method, and obtaining the preliminary driving path; Extracting the key path nodes from the preliminary driving path, calculating the path weight by weighted network analysis method, and obtaining the path weight vector.

10. The method for predicting diabetes cognitive dysfunction based on metabolomics analysis according to claim 9, characterized in that: The S8 further includes: If the stability of the path weight vector passes the preset threshold verification, adjusting the path structure by the correlation strength between nodes, and obtaining the optimized driving path; Extracting the core driving nodes from the optimized driving path, quantifying the dynamic interaction between nodes by time sequence analysis method, and obtaining the dynamic interaction matrix; Extracting the interaction area from the dynamic interaction matrix, verifying the matching degree of the area and diabetes-induced memory decline by clustering stability analysis, and obtaining the driving path.

Images