A method and system for predicting type 2 diabetes based on microbial co-abundance features

Abstract

The present application relates to a kind of based on microbial co-abundance characteristics of type 2 diabetes prediction method and system. By extracting microbial co-abundance characteristics to sample data, type 2 diabetes is predicted based on prediction model. The type 2 diabetes prediction method is based on prediction model to predict type 2 diabetes;The construction method of the prediction model includes the following steps:S1: co-abundance feature mining module: identify core microbial abundance data from sample data;S2: prediction module: construct prediction model according to core microbial abundance data;S3: consensus feature optimization module: after the prediction model is optimized, the prediction model is adjusted accordingly.A kind of based on microbial co-abundance characteristics of type 2 diabetes prediction system, the above-mentioned type 2 diabetes prediction method based on microbial co-abundance characteristics is applied, including data acquisition equipment and prediction processing equipment electrically connected with each other.The present application improves the stability of feature;Robustness of prediction model is improved.

Description

Technical Field

[0001] This invention relates to the field of disease prediction, and in particular to a method and system for predicting type 2 diabetes based on microbial co-abundance characteristics. Background Technology

[0002] Type 2 diabetes (T2D) is a complex metabolic disease with a continuously rising global prevalence. Early prediction and intervention are crucial for delaying disease progression. Recent studies have shown that the structure and function of the gut microbiota are closely related to the occurrence and development of T2D. Analyzing the composition of the microbial community using high-throughput omics technologies such as 16S rRNA sequencing and metagenomic sequencing, and constructing predictive models using machine learning methods, has become a cutting-edge direction in disease risk assessment. Common technical processes typically include sample processing, sequence analysis, feature engineering, model building, and validation. However, existing technologies still have the following significant limitations in practical applications:

[0003] 1. Feature instability and poor cross-cohort generalization: The composition of gut or oral microbiota is highly heterogeneous due to multiple factors such as region and diet. Most existing prediction models are trained on single cohort data, and the selected features often contain population-specific noise, resulting in a significant performance drop and insufficient generalization ability when the model is validated across regions and populations.

[0004] 2. Poor model interpretability and limitations of single algorithms: Most existing models rely on a single machine learning algorithm (such as random forest, support vector machine, etc.), making it difficult to simultaneously capture the complex linear and nonlinear relationships in microbial data. In addition, these models are often "black boxes," lacking reasonable interpretations of feature contributions, resulting in low clinical credibility and acceptability.

[0005] Therefore, there is an urgent need to develop a predictive method and system for type 2 diabetes based on microbial co-abundance features that has high feature stability, robustness, and interpretability. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting type 2 diabetes based on microbial co-abundance characteristics. This method identifies sample data, improves feature stability, and enhances the robustness and interpretability of the prediction model.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for predicting type 2 diabetes based on microbial co-abundance characteristics, wherein the method predicts type 2 diabetes based on a prediction model; the method for building the prediction model includes the following steps:

[0009] S1: Co-abundance feature mining module: Identifies core microbial abundance data from sample data;

[0010] S2: Prediction module: Constructs a prediction model based on core microbial abundance data;

[0011] S3: Consensus Feature Optimization Module: After optimizing the consensus features of the prediction model, the prediction model is adjusted accordingly.

[0012] Furthermore, step S1 includes the following sub-steps:

[0013] S101: After performing the first preprocessing operation on the sample data, an aggregation operation is performed to obtain genus-level abundance data;

[0014] S102: Perform a second preprocessing operation on the genus-level abundance data and divide the genus-level abundance data into multiple disease groups and multiple healthy groups according to the differences in the sample data and whether or not the patient has type 2 diabetes. The multiple disease groups are divided based on the differences in the sample data, and the multiple healthy groups are divided based on the differences in the sample data.

[0015] S103: After performing correlation analysis on multiple disease groups and multiple healthy groups, perform co-abundance feature mining.

[0016] Furthermore, the correlation analysis operation involves calculating the correlation between multiple disease groups and multiple healthy groups to construct a microbial co-abundance network.

[0017] The co-abundance feature mining operation includes the following sub-steps:

[0018] (1) The co-abundance networks of microorganisms in multiple disease groups and multiple healthy groups were compared to obtain stable disease correlation pairs and stable health correlation pairs;

[0019] (2) Core microbial abundance data were obtained by performing functional module division operations on disease stable correlation pairs and health stable correlation pairs;

[0020] Furthermore, the functional module division operation includes the following sub-steps:

[0021] 1) Perform connectivity analysis on disease-stable correlation pairs and health-stable correlation pairs respectively to obtain the largest connected clusters;

[0022] 2) Weighted correlation network analysis was performed on the maximum connected clusters of disease-stable correlation pairs and health-stable correlation pairs to obtain multiple sub-modules with competitive or cooperative relationships;

[0023] 3) Integrate species from multiple sub-modules with competing or synergistic relationships to obtain core microbial abundance data.

[0024] Furthermore, step S2 includes the following sub-steps:

[0025] S201: Perform data preprocessing on the core microbial abundance data;

[0026] S202: The optimal hyperparameter combination is obtained by performing hyperparameter optimization based on the core microbial abundance data after data preprocessing.

[0027] S203: Construct a prediction model using the optimal combination of hyperparameters;

[0028] Furthermore, the hyperparameter optimization operation is a Bayesian hyperparameter optimization operation;

[0029] The Bayesian hyperparameter optimization operation is a Bayesian hyperparameter optimization operation based on cross-validation.

[0030] The cross-validation mechanism for the Bayesian hyperparameter optimization operation based on cross-validation is as follows:

[0031] For each set of hyperparameter combinations proposed by the hyperparameter optimization operation, perform the following verification steps:

[0032] ① Divide the training set evenly into multiple mutually exclusive subsets;

[0033] ②Use each subset as the validation set and the remaining subsets as the training set. Use the training set to train the prediction model and use the validation set to validate it to obtain the validation results.

[0034] ③ Summarize the multiple verification results from step ② and evaluate the performance of the hyperparameter combination;

[0035] Furthermore, the prediction model selects the best-performing candidate base model as the final classifier by comparing the performance of multiple candidate base models laterally.

[0036] The candidate base models include, but are not limited to, LightGBM, XGBoost, Random Forest, Support Vector Machine, and Logistic Regression;

[0037] The performance evaluation metrics used in the horizontal comparison of the prediction models include ROC-AUC value, accuracy, recall, and F1 score.

[0038] Furthermore, step S3 includes the following sub-steps:

[0039] S301: Perform SHAP analysis on each feature of the prediction model to obtain SHAP analysis data;

[0040] S302: Stable core biomarker data are obtained by screening based on SHAP analysis data;

[0041] S303: Adjust the prediction model based on stable core biomarker data.

[0042] Furthermore, the SHAP analysis operation calculates global and local SHAP values ​​for each feature of the prediction model;

[0043] The filtering mechanism of the filtering operation includes, but is not limited to, filtering based on Top-K voting, filtering based on average ranking, and filtering based on directional stability.

[0044] The screening operation integrates the screening results of multiple screening mechanisms to obtain stable core biomarker data;

[0045] A type 2 diabetes prediction system based on microbial co-abundance characteristics, applying the aforementioned type 2 diabetes prediction method based on microbial co-abundance characteristics, includes a data acquisition device and a prediction processing device electrically connected to each other. The prediction processing device predicts type 2 diabetes based on a prediction model using patient health data collected by the data acquisition device.

[0046] The beneficial effects of this invention are:

[0047] The co-abundance feature mining module identifies sample data, improving feature stability; the consensus feature optimization module interprets feature contribution and adjusts the prediction model, enhancing the robustness and interpretability of the prediction model. Attached Figure Description

[0048] Figure 1 This is the overall flowchart;

[0049] Figure 2 Here is a flowchart of the co-abundance feature mining module;

[0050] Figure 3 Core microbial co-abundance network Figure 1 ;

[0051] Figure 4 Core microbial co-abundance network Figure 2 ;

[0052] Figure 5 For stable correlation pairs;

[0053] Figure 6 To stabilize the correlation between cluster partitioning;

[0054] Figure 7 This is a structural diagram of two competing submodules;

[0055] Figure 8 This is a structural diagram of the adaptive machine learning module;

[0056] Figure 9Optimize the flowchart for training;

[0057] Figure 10 Optimize the module structure diagram for consensus features;

[0058] Figure 11 A summary chart of the PR curves for each candidate base model;

[0059] Figure 12 A summary chart of ROC curves for each candidate base model;

[0060] Figure 13 A summary chart of ROC curves for the LightGBM model;

[0061] Figure 14 A summary chart of PR curves for the LightGBM model. Detailed Implementation

[0062] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0063] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0064] Example 1:

[0065] like Figures 1 to 14 As shown, a method for predicting type 2 diabetes based on microbial co-abundance characteristics is presented. This method predicts type 2 diabetes based on a prediction model; the method for building the prediction model includes the following steps:

[0066] S1: Co-abundance feature mining module: Identifies core microbial abundance data from sample data;

[0067] S2: Prediction module: Constructs a prediction model based on core microbial abundance data;

[0068] S3: Consensus Feature Optimization Module: After optimizing the consensus features of the prediction model, the prediction model is adjusted accordingly.

[0069] The sample data refers to gut microbiota or oral microbiota samples from the subjects.

[0070] The co-abundance feature mining module is a stable co-abundance feature mining module;

[0071] The prediction module is an adaptive machine learning prediction module;

[0072] The consensus feature selection module is a consensus feature selection module based on SHAP.

[0073] In this embodiment, a method for predicting type 2 diabetes based on microbial co-abundance features mainly includes three core modules: a stable co-abundance feature mining module (M1), an adaptive machine learning prediction module (M2), and a consensus feature optimization module based on SHAP (M3).

[0074] The training and validation data acquisition method in this embodiment is as follows: 16S rRNA sequencing raw data with the numbers PRJNA910326 (Shanghai), PRJNA661673 (Xinjiang), PRJNA670300 (Xiamen), PRJNA891875 (Sichuan), and PRJNA787412 (Gansu) were downloaded from the NCBI database. Among them, the Shanghai and Xinjiang cohorts are the main targets of the stable co-abundance feature mining module.

[0075] The stable co-abundance feature mining module aims to identify core microbial community structures from sample data that are not affected by geographical location or are only slightly affected by geographical location.

[0076] Step S1 includes the following sub-steps:

[0077] S101: After performing the first preprocessing operation on the sample data, an aggregation operation is performed to obtain genus-level abundance data;

[0078] The sample data in this embodiment are the Shanghai and Xinjiang queues downloaded from the NCBI database.

[0079] The differences in the sample data include the different geographical locations from which the data was sourced.

[0080] The first preprocessing operation is the generation of an ASV table.

[0081] ASV table generation process: Use CutAdapt to remove primers from sample data, dynamically truncate sequences based on base quality distribution, and then generate an ASV (amplifier sequence variant) table using QIIME2.

[0082] Aggregation operation: Species annotation was performed using the Naïve Bayes classifier (based on the Silva database), and the ASV table was aggregated to the genus level to obtain genus-level abundance data (i.e., the genus-level abundance table, which refers to the table reflecting the relative or absolute abundance of each microbial genus in each sample by summing the counts of all ASVs belonging to the same genus in the microbial sequencing data).

[0083] S102: Perform a second preprocessing operation on the genus-level abundance data and divide the genus-level abundance data into multiple disease groups and multiple healthy groups according to the differences in the sample data and whether or not the patient has type 2 diabetes. The multiple disease groups are divided based on the differences in the sample data, and the multiple healthy groups are divided based on the differences in the sample data.

[0084] The second preprocessing step is a standardized operation.

[0085] Standardization process: Perform CSS (Cumulative Sum Scaling) standardization on the genus-level abundance data to remove rare genera with a sample coverage of less than 10%.

[0086] The genus-level abundance data were divided into multiple disease groups (T1 and T2) and multiple healthy groups (H1 and H2) based on differences in sample data (different data source regions) and whether or not the patient had type 2 diabetes. T1 represents the genus-level abundance data of the Shanghai cohort with type 2 diabetes, T2 represents the genus-level abundance data of the Xinjiang cohort with type 2 diabetes, H1 represents the genus-level abundance data of the Shanghai cohort without type 2 diabetes, and H2 represents the genus-level abundance data of the Xinjiang cohort without type 2 diabetes.

[0087] S103: After performing correlation analysis on multiple disease groups and multiple healthy groups, perform co-abundance feature mining.

[0088] Correlation analysis operation: Correlation calculations were performed on multiple disease groups and multiple healthy groups to construct microbial co-abundance networks respectively.

[0089] Correlation analysis was performed using FastSpar.

[0090] The purpose of correlation analysis is to screen for stable related species. Specifically, for the divided T1, T2, H1 and H2 cohorts, the FastSpar algorithm (for calculating the correlation of component data) is used to construct the microbial co-abundance network for each subset.

[0091] The co-abundance feature mining operation includes the following sub-steps:

[0092] (1) The co-abundance networks of microorganisms in multiple disease groups and multiple healthy groups were compared to obtain stable disease correlation pairs and stable health correlation pairs;

[0093] Disease-stable correlation pairs and health-stable correlation pairs refer to significant correlation pairs with consistent correlation directions and a correlation strength greater than 0.1, which are stable related species.

[0094] In other words, by comparing the microbial co-abundance networks of the disease group (T1 vs T2), significant correlation pairs with consistent correlation directions and a correlation strength greater than 0.2 are retained in both the T1 and T2 microbial co-abundance networks; these are the stable disease correlation pairs. The same operation is performed on the network of the healthy group to obtain stable health correlation pairs.

[0095] (2) Core microbial abundance data were obtained by performing functional module division operations on disease stable correlation pairs and health stable correlation pairs;

[0096] In other words, step S103 adopts the following process:

[0097] Stability-related selection and network construction: First, the partitioned Shanghai queue ( Figure 3 Graph 1) and Xinjiang queue ( Figure 3 In Graph 2), an initial microbial co-abundance network was constructed using FastSpar. Through cross-cohort comparisons, significant (P < 0.01) correlation terms with consistent correlation direction and strength greater than 0.2 were retained to obtain stable disease / health correlation pairs (e.g., ...). Figure 4 As shown, the intersection of the Venn diagram reveals stable correlations shared across queues, thus eliminating region-specific noise.

[0098] The specific implementation process of the functional module division operation is as follows: Figures 5 to 6 As shown.

[0099] The functional module division operation includes the following sub-steps:

[0100] 1) Perform connectivity analysis on disease-stable correlation pairs and health-stable correlation pairs respectively to obtain the largest connected clusters;

[0101] Maximum Connected Cluster Extraction: New networks are constructed based on preserved stable correlation pairs, and connected component analysis (CCA) is used to identify the network topology. The maximum connected clusters (such as those with the most nodes and the tightest connections) are extracted from complex microbial networks. Figure 5 (C1 cluster in the network). This step aims to filter out isolated nodes and tiny branches in the network, focusing on the microbial community that plays a core synergistic role in the occurrence and development of disease.

[0102] 2) Weighted correlation network analysis was performed on the maximum connected clusters of disease-stable correlation pairs and health-stable correlation pairs to obtain multiple sub-modules with competitive or cooperative relationships;

[0103] Sub-module (Guilds) partitioning: The extracted largest connected clusters are further analyzed using Weighted Correlation Network Analysis (WGCNA) to partition them into different sub-modules (i.e., "guilds") based on the cooperative or competitive relationships between species. For example... Figure 6 As shown, this algorithm identifies sub-modules with significant biological characteristics, such as C1A (containing 19 genomes, such as Enterococcus and Veillonella) and C1B (containing 16 genomes, such as Lachnospiraceae). The H1 and H2 networks are then compared with... Figures 3 to 6 The same steps are used to obtain submodules D1A and D1B that have a competitive or cooperative relationship.

[0104] 3) Integrate species from multiple sub-modules with competing or synergistic relationships to obtain core microbial abundance data.

[0105] Core Feature Integration: The species in these sub-modules are deduplicated and integrated to form the final core stable microbial abundance table, which serves as high-quality input features for subsequent machine learning prediction models.

[0106] Core microbial abundance data is the core stable microbial abundance table.

[0107] By integrating the species of modules C1A, C1B, D1A, and D1B using a union approach, a core stable microbial abundance table can be obtained, which serves as the core feature set for subsequent machine learning.

[0108] The adaptive machine learning prediction module adopts an end-to-end design concept, achieving fully automated processing from the input of raw microbial feature data to the final clinical diagnostic label output. Based on the selected stable microbial features, multi-level technology fusion ensures prediction accuracy and robustness.

[0109] Step S2 includes the following sub-steps:

[0110] S201: Perform data preprocessing on the core microbial abundance data;

[0111] The adaptive machine learning prediction module uses the core stable microbial abundance table (i.e., the core microbial feature abundance matrix) obtained by the stable co-abundance feature mining module as its raw input. The raw data undergoes four processing stages to lay a solid foundation for subsequent modeling.

[0112] Data preprocessing operations include, but are not limited to, missing value imputation, distribution standardization, robust standardization, and class imbalance handling.

[0113] Missing value imputation: A zero-value constant imputation strategy is adopted to effectively handle the common data sparsity problem in microbial detection, while preserving the biological characteristics of the data.

[0114] Distribution standardization: The Yeo-Johnson power transform is applied to perform a nonlinear transformation on the skewed abundance data to make it closer to a normal distribution, thereby improving the performance of subsequent linear models.

[0115] Robust standardization: The interquartile range-based standardization method effectively suppresses the influence of outliers and enhances the stability of the model in complex real-world scenarios.

[0116] To address the class imbalance problem in the dataset, a hybrid sampling strategy was adopted.

[0117] Class imbalance handling includes SMOTE oversampling, random undersampling, and dynamic parameter adjustment.

[0118] SMOTE oversampling: intelligently synthesizes minority class samples in the feature space, increasing the diversity of T2D samples;

[0119] Random undersampling: Reduces the majority class samples by a preset ratio to balance the overall data distribution;

[0120] Dynamic parameter adjustment: The optimal sampling parameters are automatically determined based on the data distribution to avoid overfitting.

[0121] S202: The optimal hyperparameter combination is obtained by performing hyperparameter optimization based on the core microbial abundance data after data preprocessing.

[0122] The hyperparameter optimization operation is a Bayesian hyperparameter optimization operation.

[0123] The Bayesian hyperparameter optimization operation is a Bayesian hyperparameter optimization operation based on cross-validation.

[0124] The hyperparameter optimization operation uses the Optuna framework combined with the tree-structured Parzen estimator (TPE) for automated hyperparameter search. As a hyperparameter optimization framework, Optuna's built-in TPE sampler is based on the mature Bayesian optimization technique.

[0125] The cross-validation mechanism for the Bayesian hyperparameter optimization operation based on cross-validation is as follows:

[0126] For each set of hyperparameter combinations proposed by Optuna and optimized by hyperparameter operations, perform the following verification steps:

[0127] ① Divide the training set evenly into multiple mutually exclusive subsets;

[0128] In this embodiment, the training set includes the Shanghai, Xinjiang, Xiamen and Sichuan queues from the training data, and the generalization ability test of the prediction model after training is completed uses the Gansu queue.

[0129] The proportions of each category are kept consistent during the partitioning process; in this embodiment, the training set is evenly divided into 5 mutually exclusive subsets.

[0130] ②Use each subset as the validation set and the remaining subsets as the training set. Use the training set to train the prediction model and use the validation set to validate it to obtain the validation results.

[0131] In this embodiment, the verification result is the ROC-AUC value;

[0132] In this embodiment, four subsets are used as the training set.

[0133] In this way, the model is trained at each fold and the ROC-AUC value of the validation set is calculated;

[0134] ③ Summarize the multiple verification results from step ② and evaluate the performance of the hyperparameter combination;

[0135] In this embodiment, the average ROC-AUC is used as the final performance evaluation of the hyperparameter combination.

[0136] This example shows the results of 5 verifications.

[0137] The Bayesian hyperparameter optimization operation based on cross-validation embeds hierarchical 5-fold cross-validation into the optimization loop.

[0138] S203: Construct a prediction model using the optimal combination of hyperparameters;

[0139] The prediction model selects the candidate base model with the best performance on the validation set as the final classifier by comparing the performance of multiple candidate base models laterally.

[0140] The candidate base models include, but are not limited to, LightGBM, XGBoost, Random Forest, Support Vector Machine, and Logistic Regression.

[0141] LightGBM: An efficient gradient boosting framework based on histogram algorithm;

[0142] XGBoost: A gradient boosting model with regularization terms, which has strong generalization ability;

[0143] Random Forest: An ensemble decision tree method employing a Bagging strategy;

[0144] Support Vector Machine: Uses kernel functions to handle non-linear classification problems;

[0145] Logistic regression: provides linear decision boundaries and good interpretability.

[0146] The model selection mechanism is as follows: After the hyperparameters are optimized, each candidate base model obtains a stable performance evaluation value (such as average ROC-AUC) through hierarchical cross-validation under its own optimal hyperparameter combination.

[0147] The horizontal comparison is based on performance evaluation metrics on the validation set. The performance evaluation metrics used in the horizontal comparison of the prediction models include ROC-AUC value, accuracy, recall and F1 score.

[0148] By comparing performance evaluation metrics across different scenarios, we can ensure that the selected candidate base models perform optimally in the prediction task.

[0149] The predictive model provides two levels of decision output: firstly, it outputs continuous disease probability values ​​(0-1), providing clinicians with detailed risk assessment data; secondly, through a configurable decision threshold (default 0.5), it transforms the continuous probability into explicit binary classification labels ("healthy" or "T2D"), directly serving clinical diagnostic decisions. The ROC and PR curves of each candidate base model are shown below. Figure 9 and 10 As shown in the figure, the comparison of the other performance indicators is shown in Table 1.

[0150] Table 1 Performance Comparison of Candidate Base Models

[0151] Logistic 0.724 0.823 0.760 0.790 0.792 SVM 0.737 0.848 0.750 0.796 0.780 LightGBM 0.868 0.889 0.923 0.906 0.850 RandomForest 0.770 0.817 0.856 0.836 0.808 XGBoost 0.750 0.837 0.788 0.812 0.759

[0152] The consensus feature optimization module based on SHAP is used to solve the problem of "the model is accurate but we don't know why", and further refines high-confidence biomarkers.

[0153] The consensus feature optimization module based on SHAP takes the SHAP analysis results of each base module as input and combines three different strategies to screen core and stable microbial species.

[0154] Step S3 includes the following sub-steps:

[0155] S301: Perform SHAP analysis on each feature of the prediction model to obtain SHAP analysis data;

[0156] The SHAP analysis operation calculates global and local SHAP values ​​for each feature of the prediction model.

[0157] Global and local SHAP value calculation: For each base model in the prediction model, the SHAP value matrix of the feature is calculated using the SHAP interpreter; the global importance of the feature (average of the absolute values ​​of SHAP) and the direction of local influence (positive or negative SHAP value, positive value represents increased risk of disease, negative value represents decreased risk) are obtained.

[0158] SHAP stands for SHapley Additive exPlanations.

[0159] S302: Stable core biomarker data are obtained by screening based on SHAP analysis data;

[0160] Step S302 adopts a strategy consensus feature screening mechanism, which introduces three complementary screening strategies for "voting".

[0161] The filtering mechanism of the filtering operation includes, but is not limited to, filtering based on Top-K voting, filtering based on average ranking, and filtering based on directional stability.

[0162] Top-K Voting-Based Selection: Extract the top K features by SHAP importance from each base model (e.g., XGBoost, SVM, etc.). Count how many models selected each feature into the Top-K. Set a threshold (e.g., min_votes=3) to retain frequently occurring features.

[0163] Average Ranking Selection: Calculate the average ranking of each feature across all models and retain the features with the highest average ranking.

[0164] Stability-based selection: This method calculates the consistency of a feature's influence direction across different models. A stability score is defined as the proportion of features whose SHAP values ​​have the same sign (both positive or both negative) between any two models. A threshold (e.g., 0.6) is set to remove features with contradictory effects in different algorithms (i.e., unstable features that show pathogenicity in model A but protection in model B).

[0165] The screening operation integrates the screening results of multiple screening mechanisms to obtain stable core biomarker data;

[0166] Final biomarker determination: The intersection of the screening results from the three screening mechanisms mentioned above is taken as the stable core biomarker list, i.e., stable core biomarker data, thereby determining the final "stable core taxonomy" and outputting it for subsequent retraining of the prediction model. The final determined core biomarkers include: Intestinibacter, Lachnospiraceae_NK4A136_group, Lachnospiraceae_UCG-010, Family_XIII_UCG-001, Family_XIII_AD3011_group, Anaerostipes, Blautia, Bacteroides, Prevotella_9, Lachnoclostridium, Veillonella, [Eubacterium]_ruminantium_group.

[0167] S303: Adjust the prediction model based on stable core biomarker data.

[0168] In step S303, after feature selection, the selected features are used to perform secondary training on the candidate base model with the best performance, and the generalization ability of the prediction model is tested using the Gansu queue. The ROC curve and PR curve of the secondary-trained LightGBM model are shown below. Figure 12 and Figure 13 As shown.

[0169] Existing methods for predicting type 2 diabetes rely on simplistic and unstable feature selection strategies: traditional feature selection methods are usually based on a single indicator (such as information gain, Gini importance, etc.), ignoring the consistency of features across different models and the stability of their biological effects (promoting or inhibiting the disease). This can easily lead to a high false positive rate for the selected biomarkers, which is not conducive to subsequent mechanism research and clinical translation.

[0170] This problem can be solved by comparing multiple base models and performing a selection based on directional stability in step S3.

[0171] A type 2 diabetes prediction system based on microbial co-abundance characteristics, applying the aforementioned type 2 diabetes prediction method based on microbial co-abundance characteristics, includes a data acquisition device and a prediction processing device electrically connected to each other. The prediction processing device predicts type 2 diabetes based on a prediction model using patient health data collected by the data acquisition device.

[0172] The co-abundance feature mining module identifies sample data, improving feature stability; the consensus feature optimization module interprets feature contribution and adjusts the prediction model, enhancing the robustness and interpretability of the prediction model.

[0173] A method and system for predicting type 2 diabetes based on microbial co-abundance features is proposed. This method and system is based on SHAP analysis and multi-model consensus feature selection. A T2D prediction system integrating cross-cohort network analysis, adaptive machine learning and SHAP consensus feature selection is proposed.

[0174] A novel T2D prediction method and system based on microbial co-abundance features is a novel T2D prediction system that can mine stable microbial features from multi-center, sample data and integrate high-performance machine learning and interpretability analysis to improve the robustness, generalization ability and clinical applicability of the model.

[0175] A method and system for predicting type 2 diabetes based on microbial co-abundance characteristics has the following significant advantages:

[0176] 1. Improve model generalization and stability: By screening stable co-abundance features through cross-cohort network analysis, population-specific noise is reduced, significantly enhancing the model's generalization ability and robustness on heterogeneous datasets.

[0177] 2. Enhance prediction performance and reliability: Employ a multi-algorithm comparison and selection mechanism to improve classification accuracy and AUC value; combine Bayesian hyperparameter optimization and cross-validation to avoid overfitting and ensure optimal model performance.

[0178] 3. Improve interpretability and clinical acceptability: Introduce the SHAP multi-model consensus feature optimization mechanism, which not only provides the global importance of features, but also reveals the consistency of their direction of action, outputting a stable and interpretable set of biomarkers, thereby enhancing the credibility of the model in clinical decision-making.

[0179] 4. Achieve automated and systematic processes: From feature selection, model training, optimization to interpretation and output, the process scripts are automated, reducing manual intervention and improving system efficiency and practicality. It is suitable for multi-center, large-sample T2D risk screening scenarios.

[0180] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for predicting type 2 diabetes based on microbial co-abundance characteristics, characterized in that: The type 2 diabetes prediction method is based on a prediction model to predict type 2 diabetes; the method for building the prediction model includes the following steps: S1: Co-abundance feature mining module: Identifies core microbial abundance data from sample data; S2: Prediction module: Constructs a prediction model based on core microbial abundance data; S3: Consensus Feature Optimization Module: After optimizing the consensus features of the prediction model, the prediction model is adjusted accordingly; Step S3 includes the following sub-steps: S301: Perform SHAP analysis on each feature of the prediction model to obtain SHAP analysis data; S302: Stable core biomarker data are obtained by screening based on SHAP analysis data; S303: The prediction model is adjusted based on stable core biomarker data and then retrained. Step S1 includes the following sub-steps: S101: After performing the first preprocessing operation on the sample data, an aggregation operation is performed to obtain genus-level abundance data; S102: Perform a second preprocessing operation on the genus-level abundance data and divide the genus-level abundance data into multiple disease groups and multiple healthy groups based on the differences in the sample data and whether or not the patient has type 2 diabetes. S103: After performing correlation analysis on multiple disease groups and multiple healthy groups, perform co-abundance feature mining. The correlation analysis operation involves calculating the correlation between multiple disease groups and multiple healthy groups to construct a microbial co-abundance network. The co-abundance feature mining operation includes the following sub-steps: (1) The co-abundance networks of microorganisms in multiple disease groups and multiple healthy groups were compared to obtain stable disease correlation pairs and stable health correlation pairs; (2) Perform functional module division operations on disease stable correlation pairs and health stable correlation pairs to obtain core microbial abundance data.

2. The method for predicting type 2 diabetes based on microbial co-abundance characteristics according to claim 1, characterized in that: The multiple disease groups are divided based on differences in sample data, and the multiple healthy groups are divided based on differences in sample data.

3. The method for predicting type 2 diabetes based on microbial co-abundance characteristics according to claim 1, characterized in that: The functional module division operation includes the following sub-steps: 1) Perform connectivity analysis on disease-stable correlation pairs and health-stable correlation pairs respectively to obtain the largest connected clusters; 2) Weighted correlation network analysis was performed on the maximum connected clusters of disease-stable correlation pairs and health-stable correlation pairs to obtain multiple sub-modules with competitive or cooperative relationships; 3) Integrate species from multiple sub-modules with competing or synergistic relationships to obtain core microbial abundance data.

4. The method for predicting type 2 diabetes based on microbial co-abundance characteristics according to claim 1, characterized in that: Step S2 includes the following sub-steps: S201: Perform data preprocessing on the core microbial abundance data; S202: The optimal hyperparameter combination is obtained by performing hyperparameter optimization based on the core microbial abundance data after data preprocessing. S203: Construct a prediction model using the optimal combination of hyperparameters.

5. The method for predicting type 2 diabetes based on microbial co-abundance characteristics according to claim 4, characterized in that: The hyperparameter optimization operation is a Bayesian hyperparameter optimization operation. The Bayesian hyperparameter optimization operation is a Bayesian hyperparameter optimization operation based on cross-validation. The cross-validation mechanism for the Bayesian hyperparameter optimization operation based on cross-validation is as follows: For each set of hyperparameter combinations proposed by the hyperparameter optimization operation, perform the following verification steps: ① Divide the training set evenly into multiple mutually exclusive subsets; ②Use each subset as the validation set and the remaining subsets as the training set. Use the training set to train the prediction model and use the validation set to validate it to obtain the validation results. ③ Summarize the multiple verification results from step ② and evaluate the performance of the hyperparameter combination.

6. The method for predicting type 2 diabetes based on microbial co-abundance characteristics according to claim 1, characterized in that: The prediction model selects the best-performing candidate base model as the final classifier by comparing the performance of multiple candidate base models laterally. The candidate base models include LightGBM, XGBoost, Random Forest, Support Vector Machine, and Logistic Regression; The performance evaluation metrics used in the horizontal comparison of the prediction models include ROC-AUC value, accuracy, recall, and F1 score.

7. The method for predicting type 2 diabetes based on microbial co-abundance characteristics according to claim 1, characterized in that: The SHAP analysis operation calculates global and local SHAP values ​​for each feature of the prediction model. The filtering mechanism of the filtering operation includes filtering based on Top-K voting, filtering based on average ranking, and filtering based on directional stability. The screening operation integrates the screening results of multiple screening mechanisms to obtain stable core biomarker data.

8. A type 2 diabetes prediction system based on microbial co-abundance characteristics, employing the type 2 diabetes prediction method based on microbial co-abundance characteristics as described in any one of claims 1-7, characterized in that: It includes a data acquisition device and a prediction processing device that are electrically connected to each other. The prediction processing device makes a prediction of type 2 diabetes based on a prediction model using patient health data collected by the data acquisition device.

Images