Cvd death subpopulation identification method and device combining causal reasoning and consensus clustering

Abstract

The application discloses a CVD death subpopulation identification method and device combining causal reasoning and consensus clustering, and the method comprises the following steps: collecting original feature data and performing pretreatment; performing intermediate causal relationship reasoning on the pretreated feature data, and taking the features contained in the intermediate causal relationship reasoning result as an initial feature set; using a plurality of machine learning models to respectively perform death risk prediction according to input feature data, and training the initial feature set to obtain a plurality of death risk prediction models; using a SHAP algorithm to calculate SHAP values of feature variables of all death risk prediction models, and using the feature variable SHAP values to select an optimal model and an optimal feature set; performing consensus clustering on death risk prediction results output by the optimal feature set and the optimal model to obtain a plurality of death subpopulations; and performing statistical analysis on feature data and survival results of each death subpopulation, and identifying a pathway of each death subpopulation in combination with an intermediate causal relationship reasoning result.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology, and in particular to a method and apparatus for identifying CVD death subgroups by combining causal reasoning and consensus clustering. Background Technology

[0002] As an emerging artificial intelligence technology, machine learning (ML) has been widely applied in predicting CVD mortality risk in recent years, demonstrating significant advantages, especially in handling large-scale, complex, and multidimensional data. ML can help healthcare institutions identify individuals at potential CVD mortality risk early, enabling timely intervention to reduce CVD-related deaths. Currently, various features are used to build CVD mortality risk prediction models, including dietary data, lifestyle factors, environmental exposure data, retinal imaging, low-dose computed tomography (CT) scans, and traditional clinical variables such as demographic characteristics, medical history, and basic laboratory indicators.

[0003] However, existing CVD mortality risk prediction models primarily target specific disease populations. While these models have achieved good performance in predicting future CVD mortality risk, they still have certain limitations. The key features identified by these models are often closely related to CVD mortality in specific patient groups, but they lack mediating causal inferences and cannot reveal potential mediating causal relationships associated with CVD mortality. Understanding these causal relationships is crucial for fundamentally reducing CVD mortality and elucidating its underlying mechanisms. Furthermore, previous studies have rarely conducted stratified analyses of CVD mortality subgroups to explore the potential mechanisms of transitions between different survival statuses, which is particularly critical for healthcare institutions to prevent further disease progression and implement precise interventions. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art.

[0005] The technical solution adopted by this invention to solve its technical problem is: to provide a CVD death subgroup identification method combining causal reasoning and consensus clustering, including the following steps:

[0006] Raw feature data is collected and preprocessed to obtain preprocessed feature data, wherein the feature variables in the preprocessed feature data include exposure factors and laboratory indicators;

[0007] Mediation causal reasoning is performed on the preprocessed feature data, using the exposure factors and laboratory indicators contained in the mediation causal reasoning results as the initial feature set;

[0008] Several machine learning models are used to predict mortality risk based on the input feature data. Each model is then trained using an initial feature set to obtain several mortality risk prediction models. The SHAP algorithm is used to calculate the SHAP value of the feature variables of all mortality risk prediction models. The optimal model and the optimal feature set are selected using the SHAP value of the feature variables of each mortality risk prediction model.

[0009] Consensus clustering is performed on the optimal feature set and the mortality risk prediction results output by the optimal model to obtain several mortality subgroups;

[0010] Statistical analysis was performed on the characteristic data and survival outcomes of each mortality subgroup, and the pathways of each mortality subgroup were identified by combining the statistical analysis results with the results of mediation causal reasoning.

[0011] Preferably, the preprocessing includes:

[0012] Remove feature variables with more than 50% missing values;

[0013] Data from subjects under 20 years of age were excluded.

[0014] Feature data of subjects with more than 50% missing values ​​were removed;

[0015] Each subject was categorized based on their survival status at different time points.

[0016] Preferably, the classification of each subject based on their survival status at different time points includes:

[0017] At the three-year follow-up, the subjects were classified into the survival group or the death group based on whether they survived.

[0018] At the five-year follow-up, subjects were classified into the survival group or the death group based on whether they survived.

[0019] At the ten-year follow-up, subjects were categorized into the survival group or the death group based on whether they were alive.

[0020] Preferably, the mediation causal relationship inference on the preprocessed feature data includes the following steps:

[0021] The K-nearest neighbor algorithm is used to fill in the preprocessed feature data;

[0022] Exhaustive combinations of exposure factors and laboratory indicators are performed; for each pair of exposure factors and laboratory indicators, verification is carried out based on several conditionally independent tests to obtain several P-values, and the maximum value among them is taken as the overall P-value.

[0023] If the overall P-value is less than 0.05, it indicates that the exposure factor can influence CVD mortality through this laboratory indicator as a mediating variable.

[0024] The aforementioned conditionally independent tests include:

[0025] The exposure factor was significantly associated with mortality outcomes;

[0026] After adjusting for mortality outcomes, the exposure factor and the mediating variable remained significantly associated; the mediating variable refers to laboratory indicators.

[0027] After adjusting for exposure factors, the mediating variable remained significantly associated with mortality outcomes;

[0028] After adjusting for mediating variables, the exposure factor was not associated with mortality outcomes.

[0029] Preferably, the SHAP algorithm is used to calculate the SHAP value for the feature variables of all mortality risk prediction models, and the optimal model and optimal feature set are selected using the SHAP value of the feature variables of each mortality risk prediction model. Specifically:

[0030] The SHAP algorithm was used to calculate the SHAP values ​​for the feature variables of all mortality risk prediction models.

[0031] For each mortality risk prediction model, the number of input features is gradually reduced based on the SHAP value of the feature variables, and the prediction effect is evaluated. The optimal model is selected by combining the prediction effect and the number of feature variables; the set of input feature variables of the optimal model is the optimal feature set.

[0032] Preferably, the optimal feature set includes 4 exposure factors and 6 laboratory indicators; the 4 exposure factors are age, household income to poverty ratio, gender and marital status; the 6 laboratory indicators are lymphocyte percentage, neutrophil percentage, hemoglobin, hematocrit, mean corpuscular volume, and osmolarity; and the optimal model is a logistic regression model with the optimal feature set as input feature variables.

[0033] Preferably, the consensus clustering of the mortality risk prediction results output by the optimal feature set and the optimal model to obtain several mortality subgroups includes the following steps:

[0034] The feature values ​​of the final feature set are input into the optimal prediction model to obtain the mortality risk prediction result;

[0035] Consensus clustering is performed on the eigenvalues ​​and the prediction results to obtain three death subgroups.

[0036] Preferably, the statistical analysis of the characteristic data and survival outcomes of each mortality subgroup includes:

[0037] Survival analysis was performed, using Kaplan-Meier survival curves to estimate the survival probability of each mortality subgroup, and the Log-rank test was used to compare the differences in survival curves between groups.

[0038] Descriptive analysis uses median and quartiles for continuous variables and frequency and percentage for categorical variables;

[0039] For comparisons between groups, the Mann-Whitney U test is used for continuous variables and the chi-square test or Fisher's exact test is used for categorical variables. For multiple comparisons between three groups, Bonferroni correction is used to adjust the p-value to control for Type I error.

[0040] The continuous variables include age, household income to poverty ratio, percentage of lymphocytes, percentage of neutrophils, hemoglobin, hematocrit, mean corpuscular volume, and osmolarity; the categorical variables include gender and marital status.

[0041] Preferably, the pathway for identifying each mortality subgroup by combining statistical analysis results and mediating causal reasoning results is specifically as follows:

[0042] Age affects the transformation between different subsets by influencing hemoglobin, hematocrit, lymphocyte percentage, neutrophil percentage, and osmotic pressure.

[0043] Sex affects the transformation between different subsets by influencing lymphocyte percentage, osmotic pressure, and mean erythrocyte volume.

[0044] Marital status influences the transition between different subsets by affecting changes in lymphocyte percentage, neutrophil percentage, osmolarity, and mean erythrocyte volume.

[0045] Household income and poverty ratios influence the transition between different subgroups by affecting changes in hemoglobin and hematocrit.

[0046] The present invention also provides a CVD death subgroup identification device that combines causal reasoning and consensus clustering, comprising:

[0047] The data collection module collects raw feature data and preprocesses it to obtain preprocessed feature data. The feature variables in the preprocessed feature data include exposure factors and laboratory indicators.

[0048] The causal reasoning module performs mediating causal reasoning on the preprocessed feature data, using the exposure factors and laboratory indicators contained in the mediating causal reasoning results as the initial feature set.

[0049] The model building module uses several machine learning models to predict mortality risk based on the input feature data, and then trains each model using an initial feature set to obtain several mortality risk prediction models. The SHAP algorithm is used to calculate the SHAP value of the feature variables of all mortality risk prediction models, and the optimal model and optimal feature set are selected using the SHAP value of the feature variables of each mortality risk prediction model.

[0050] The subgroup acquisition module performs consensus clustering on the optimal feature set and the mortality risk prediction results output by the optimal model to obtain several mortality subgroups.

[0051] The subpopulation identification module performs statistical analysis on the characteristic data and survival outcomes of each mortality subpopulation, and identifies the pathways of each mortality subpopulation by combining the statistical analysis results with the results of mediation causal relationship inference.

[0052] This invention has the following beneficial effects: First, it reveals the potential pathways of CVD mortality by using mediating causal inference, providing an important reference for medical institutions to reduce CVD mortality and optimize prevention strategies; second, it identifies the 10 most critical characteristic variables in the CVD mortality pathway, all of which are common clinical indicators, effectively reducing model complexity while ensuring model performance, and laying a good foundation for the model's promotion in the real world; finally, these 10 characteristic variables can be further used for risk stratification of CVD mortality populations and reveal the potential causal mechanisms of mutual transformation between different CVD mortality subgroups.

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the steps of a CVD death subgroup identification method combining causal reasoning and consensus clustering, according to an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the data preprocessing process for a CVD death subpopulation identification method combining causal reasoning and consensus clustering, according to an embodiment of the present invention.

[0056] Figure 3 This is a visual diagram illustrating the mediating causal reasoning results of a CVD death subgroup identification method combining causal reasoning and consensus clustering, according to an embodiment of the present invention.

[0057] Figure 4 This invention presents the prediction results of four mortality risk prediction models for a CVD mortality subgroup identification method that combines causal reasoning and consensus clustering for a 3-year follow-up period.

[0058] Figure 5The graph shows the AUC of a mortality risk prediction model based on a CVD mortality subgroup identification method combining causal reasoning and consensus clustering, as the number of features decreases, according to an embodiment of the present invention.

[0059] Figure 6 This is a Kaplan-Meier survival curve diagram of different subgroups in a CVD death subgroup identification method that combines causal reasoning and consensus clustering, according to an embodiment of the present invention.

[0060] Figure 7 This is a distribution diagram of different features in different subgroups of a CVD death subgroup identification method that combines causal reasoning and consensus clustering, according to an embodiment of the present invention.

[0061] Figure 8 This is a visual diagram illustrating the transformation network between different death subgroups in a CVD death subgroup identification method that combines causal reasoning and consensus clustering, according to an embodiment of the present invention.

[0062] Figure 9 This is a schematic diagram of a CVD death subgroup identification device that combines causal reasoning and consensus clustering according to an embodiment of the present invention. Detailed Implementation

[0063] See Figure 1 The diagram illustrates the steps of a CVD death subgroup identification method combining causal reasoning and consensus clustering according to an embodiment of the present invention, including the following steps:

[0064] S101, Collect raw feature data and preprocess it to obtain preprocessed feature data. The feature variables in the preprocessed feature data include exposure factors and laboratory indicators.

[0065] S102, perform mediation causal reasoning on the preprocessed feature data to obtain the exposure factors and laboratory indicators contained in the mediation causal reasoning results as the initial feature set;

[0066] S103: Several machine learning models are used to predict mortality risk based on the input feature data. Then, each model is trained using an initial feature set to obtain several mortality risk prediction models. The SHAP algorithm is used to calculate the SHAP value of the feature variables of all mortality risk prediction models. The optimal model and the optimal feature set are selected using the SHAP value of the feature variables of each mortality risk prediction model.

[0067] S104, consensus clustering is performed on the optimal feature set and the mortality risk prediction results output by the optimal model to obtain several mortality subgroups;

[0068] S105, statistical analysis is performed on the characteristic data and survival outcomes of each mortality subgroup, and the pathways of each mortality subgroup are identified by combining the statistical analysis results and the results of mediation causal reasoning.

[0069] Specifically, in S101, this embodiment of the invention uses the NHANES dataset, which includes 40,617 participants. The characteristic variables include basic demographic information, anthropometric data, laboratory indicators, and questionnaire data. The data is then preprocessed. This preprocessing first removes characteristic variables with more than 50% missing values, selecting 20 exposure factors and 47 laboratory indicators for further analysis; then... (See...) Figure 2 As shown, individuals under 20 years of age and participants with more than 50% missing features were excluded to further optimize the study population and advance the next phase of analysis. Mortality status data were matched with NHANES exposure data to classify participants as dead or alive at different time points. Due to variations in follow-up duration, follow-up intervals were standardized, with three key time points of 3, 5, and 10 years pre-defined. At these three time points, all participants were categorized as either CVD-related deaths or survivors. Furthermore, to ensure the study focused on CVD-related deaths, all individuals who died from non-CVD causes were excluded.

[0070] Specifically, the 20 exposure factors include: Age; Gender; Education Level - Adults 20+; Marital Status; Race; Body Mass Index (kg / m²); Waist Circumference (cm); Minutes of sedentary activity; Ratio of family income to poverty; Ever told you had high blood pressure; Doctor told you have diabetes; Ever told you had coronary heart disease; Ever told you had a stroke; Had at least 12 alcohol drinks / 1 year; Vigorous work activity; Moderate work activity; Walk or bicycle; Vigorous recreational activities; Moderate Recreational activities of moderate intensity; Smoked at least 100 cigarettes in a lifetime. The 47 laboratory indicators include: Albuminurine (mg / L), urinary albumin; Creatinineurine (umol / L), urinary creatinine; Direct HDL-Cholesterol (mmol / L), high-density lipoprotein cholesterol (HDL); Whiteblood cell count (1000 cells / uL), Lymphocyte percent (%), Monocyte percent (%), Segmented neutrophils percent (%), and Eosinophils percent (%).Basophils percent (%), absolute number of lymphocytes (1000 cells / uL), absolute number of monocytes (1000 cells / uL), absolute number of segmented neutrophils (1000 cells / uL), absolute number of neutrophils (1000 cells / uL), absolute number of eosinophils (1000 cells / uL), absolute number of basophils (1000 cells / uL), absolute number of red blood cells (million cells / uL), red blood cell count (million cells / uL), hemoglobin (g / dL), hematocrit (%), mean corpuscular volume (fL), mean corpuscular volume (MCV), mean corpuscular hemoglobin (pg), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (g / dL), mean corpuscular hemoglobin concentration (MCHC), red cell distribution. Red blood cell distribution width (%), RDW; Platelet count (1000 cells / uL), PLT; Mean platelet volume (fL), MPV; Glycohemoglobin (%), HbA1c; Albumin (g / L), serum albumin; Alanine Aminotransferase (ALT) (U / L), alanine transaminase; Aspartate Aminotransferase (AST) (U / L), aspartate transaminase; Alkaline Phosphatase (ALP) (IU / L), alkaline phosphatase; Blood urea nitrogen (mmol / L), blood urea nitrogen; Total calcium (mmol / L), total calcium; Cholesterol (mmol / L), total cholesterol; Bicarbonate (mmol / L), bicarbonate (HCO3⁻); Creatinine (umol / L), serum creatinine; Gamma Glutamyl Transferase (GGT) (U / L), γ-glutamyl transferase;Glucose serum (mmol / L); Iron refigerated (umol / L); Lactate dehydrogenase (LDH) (U / L); Phosphorus (mmol / L); Total bilirubin (umol / L); Total protein (g / L); Triglycerides (mmol / L); Uric acid (umol / L); Sodium (mmol / L); Potassium (mmol / L); Chloride (mmol / L); Osmolality (mmol / Kg); Globulin (g / L).

[0071] Specifically, in S102, for subject feature data with missing values, the K-Nearest Neighbors (KNN) algorithm is used for imputation. KNN is a nonparametric statistical learning method that can estimate missing data using feature values ​​of similar individuals in the dataset. After the missing values ​​are imputed, the R software package 'cit' is used to explore the mediating causal relationship between exposure factors, laboratory indicators, and CVD mortality. This software package uses a likelihood-based hypothesis testing method to assess the causal mediating effect. Specifically, it examines whether exposure factors affect CVD mortality through laboratory indicators as mediating variables and verifies this based on the following four conditional independence tests: (1) exposure factors are significantly associated with mortality; (2) after adjusting for mortality, exposure factors are still significantly associated with the mediating variable (laboratory indicators); (3) after adjusting for exposure factors, the mediating variable is still significantly associated with mortality; (4) after adjusting for the mediating variable, exposure factors are not associated with mortality. If the p-value of the above four independence tests is less than 0.05, it means that the hypothesis is met. For example, if exposure factors are significantly associated with mortality, the output p-value should be less than 0.05. The p-values ​​of the four tests were comprehensively evaluated using the intersection-union test. The significance level of the overall causal inference test was set at p < 0.05, meaning that the maximum p-value output by the four tests was the final overall p-value. If the overall p-value was less than 0.05, it indicated that a certain exposure factor could influence CVD mortality through a certain laboratory indicator as a mediating variable.

[0072] In this embodiment of the invention, subjects were divided into a survival group and a CVD death group based on different follow-up time points. Significant differences existed between the two groups in terms of exposure factors and laboratory indicators (e.g., the CVD death group was significantly older and had a higher proportion of males compared to the survival group). Mediation causal inference revealed that during the 3-year follow-up period, a total of 18 exposure factors mediated CVD death through 36 laboratory indicators; while during the 5-year and 10-year follow-up periods, 19 exposure factors mediated CVD death through 40 and 42 laboratory indicators, respectively. These exposure factors and laboratory indicators with mediating causal relationships will be used in the next step to construct the predictive model. Finally, the mediation causal network was visualized using Cytoscape (version 3.10.2), and the visualization effect can be seen in the diagram below. Figure 3 As shown, the left column represents exposure factors, the middle column represents mediating variables (using laboratory indicators as mediating variables), and the right column represents mortality outcomes; the causal relationship between items in different columns is indicated by lines.

[0073] Specifically, in S103, the ML models used include Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost). The input features are divided into categorical and continuous variables. All continuous variables are standardized before model training to eliminate bias caused by different measurement scales. Categorical variables are converted to numerical format using one-hot encoding to prevent the introduction of order relations or implicit distance assumptions into the dataset. After feature preprocessing, the dataset is randomly divided into training and validation sets in a 7:3 ratio, and model training is performed only on the training set. To determine the optimal hyperparameters and prevent overfitting, this embodiment of the invention combines grid search cross-validation with manual tuning for parameter optimization. The adjusted hyperparameters are as follows:

[0074] Logistic Regression (LR): C, max_iter, penalty, solver;

[0075] Random Forest (RF): max_depth, min_samples_leaf, n_estimators;

[0076] Support Vector Machine (SVM): C, gamma, kernel;

[0077] Extreme Gradient Boosting (XGBoost): colsample_bytree, gamma, learning_rate, max_depth, n_estimators, subsample.

[0078] All models employed 5-fold cross-validation, with the area under the receiver operating characteristic (ROC) curve (AUC) used as the primary evaluation metric to select the optimal model. To address class imbalance, this invention adjusted the `class_weight` parameter in LR, RF, and SVM, and the `scale_pos_weight` parameter in XGBoost. All models were trained using the optimal estimator and validated on the validation set. Model performance evaluation metrics included sensitivity (Sn), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), F1 score, Matthews correlation coefficient (MCC), and accuracy (Acc). AUC was also used for comprehensive model performance evaluation. Furthermore, to further assess model robustness, bootstrapping was used to calculate all performance evaluation metrics on the validation set, and their 95% confidence intervals (CIs) were determined.

[0079] This invention evaluates the performance of exposure factors and laboratory indicators that mediate causality in CVD mortality pathways in constructing CVD mortality prediction models. Exposure factors and laboratory indicators were used individually or in combination, and four traditional machine learning (ML) methods were employed for modeling. The results show that the model constructed using both exposure factors and laboratory indicators as feature variables outperformed models using only exposure factors or laboratory indicators. The logistic regression (LR) model showed the best prediction performance for 3-year, 5-year, and 10-year CVD mortality, with AUC values ​​of 0.914 (95% CI: 0.879–0.941), 0.923 (95% CI: 0.895–0.946), and 0.908 (95% CI: 0.886–0.929), respectively. For example, the mortality risk prediction results of the four models for 3-year CVD mortality are shown below. Figure 4 As shown.

[0080] Due to the "black box" nature of machine learning (ML), it is difficult to explain the contribution of each feature to the model's predictions. Therefore, the SHapley Additive exPlanations (SHAP) algorithm was introduced to interpret the output of machine learning models. The SHAP algorithm assigns a SHAP value to each feature to measure its impact on the predictive model. Based on the SHAP values, the importance of all features is ranked across different machine learning models. Then, models with a reduced number of features are built based on these rankings to identify the optimal machine learning model and the appropriate number of features. AUC is used as the primary metric to evaluate model performance. Furthermore, model complexity, training cost, and interpretability are also considered to ultimately select the best model architecture.

[0081] Specifically, the SHAP algorithm was used to provide a unified interpretation for the four ML models to ensure comparability of the results. Comparison of the feature SHAP values ​​of the four ML models revealed that although there were slight differences in feature ranking among the different algorithms, there was considerable overlap among the top 20 features, and these features maintained a highly consistent ranking trend across the 3-year, 5-year, and 10-year CVD mortality prediction models. Notably, age ranked first in all models, consistent with the traditional understanding of CVD mortality risk factors. Subsequently, the number of features in the model was gradually reduced based on the SHAP value ranking to leverage the advantages of ML in feature selection, identify the most critical CVD mortality drivers, and simultaneously maintain good model performance while reducing model complexity. The trend of the model's AUC with feature reduction is shown in [reference needed]. Figure 5 As shown in the figure, the top 61, 50, 40, 30, 20, 10, and 5 features with the highest SHAP values ​​were selected to evaluate the prediction performance. As can be seen from the figure, in CVD mortality prediction at different follow-up time points, the LR and XGBoost models showed the smallest decrease in AUC when the number of features decreased. When determining the final number of features, it was found that when the number of features in the LR model was reduced from 20 to 5, the AUC decreased significantly. To balance model performance and complexity, a 10-feature LR model was selected as the optimal model. These 10 features included age, household income and poverty ratio, lymphocyte percentage, neutrophil percentage, hemoglobin, hematocrit, mean corpuscular volume, osmolarity, sex, and marital status. The AUCs for predicting CVD mortality over 3, 5, and 10 years were 0.884 (95% CI: 0.847–0.917), 0.896 (95% CI: 0.865–0.924), and 0.879 (95% CI: 0.855–0.902), respectively. Other performance metrics also showed good results, as shown in Table 1. Considering factors such as low model complexity, low training cost, and strong interpretability, the LR model was ultimately chosen as the optimal model.

[0082] Table 1 - Performance evaluation results of the LR model at different follow-up time points:

[0083]

[0084] Specifically, in S104, consensus clustering is used to identify CVD mortality subgroups. The input data for consensus clustering includes the features of the optimal model and the predicted CVD mortality probability of individuals. This embodiment of the invention filters out all individuals who died from CVD during a 10-year follow-up period and performs consensus clustering analysis. The input data for clustering includes all features of the optimal model and the CVD mortality prediction probability generated by the model. This embodiment of the invention uses the 'ConsensusClusterPlus' R package for consensus clustering, with the following parameter settings: maxK=6, reps=1000, pItem=0.8, pFeature=1, clusterAlg="km", distance="euclidean". The number of clusters is determined based on two factors: the average pairwise consensus matrix within the consensus clusters and the area under the cumulative distribution function (CDF) curve (deltaplot). Ultimately, the optimal number of subgroups was determined to be 3. This decision was based on the following two points: (1) the separation effect of the consensus matrix is ​​clearest when K=3 and K=4; (2) the increase in the area under the CDF curve is smaller when K=4 than when K=3, so K=3 is the most suitable choice. Based on the clustering results, the CVD-related deaths were divided into three subgroups.

[0085] Specifically, in step S105, survival analysis was performed on subgroups. The purpose of the survival analysis was to compare the differences in CVD mortality risk among different groups during the follow-up period. In this embodiment of the invention, Kaplan-Meier survival curves were used to estimate the survival probability of each group; the Log-rank test was used to compare the differences in survival curves between groups. For example, the CVD mortality risk of a certain subgroup was significantly higher than that of other groups (Log-rank test, P < 0.05), showing a survival difference. Continuous variables were expressed as medians (25th and 75th percentiles), and categorical variables were expressed as frequencies (percentages).

[0086] To further explore the differences in characteristics among the three subgroups, statistical tests were performed on all 10 characteristics across the different subgroups, including descriptive analysis and intergroup comparisons. The aim was to summarize and describe the basic characteristics of the study sample and compare differences in variables between different groups, including: continuous variables expressed as median and quartiles, and categorical variables expressed as frequency and percentage; comparisons between two groups were performed using the Mann-Whitney U test for continuous variables and the chi-square test (χ²) or Fisher's exact test for categorical variables; multiple comparisons among the three groups used Bonferroni correction to adjust p-values ​​to control for Type I errors. For example, some experimental groups had significantly higher medians for variable X than the control group (P < 0.05), indicating a statistically significant difference.

[0087] All statistical tests were two-tailed, and p < 0.05 was considered statistically significant. Survival analysis was performed using R (version 4.4.2), and other statistical analyses were performed using Python (version 3.11.5). The Kaplan-Meier survival curves for different CVD mortality subgroups were obtained in this embodiment of the invention; see [link to relevant documentation]. Figure 6 As shown, subgroup 1 had the best survival rate, followed by subgroup 2, while subgroup 3 had the worst survival rate. For the statistical distribution results of different characteristics in different mortality subgroups, see [link to relevant documentation]. Figure 7 As shown in Table 2, Figure 7 The statistical graphs in the graphs are represented by box plots or stacked bar plots. ns indicates no significance, ** indicates p < 0.01, and *** indicates p < 0.001. The horizontal axis of the statistical graphs represents subgroups, and the vertical axis represents different characteristic variables. It can be seen that, except for the ratio of family income to poverty, the other characteristics show different degrees of difference among the three groups.

[0088] Table 2 - Statistical analysis results of the characteristics of participants in different CVD mortality subgroups:

[0089]

[0090] Finally, these statistical analysis results were integrated with the previous mediating causal inference results to construct a transformation network among different CVD mortality subgroups, which was then visualized to enhance the understanding of the potential mediating causal relationships driving subgroup transformation; the integration process involved... Figure 3 The red nodes are replaced with different subgroups, and the features are simplified to the 10 key features selected above. The trends of the red and blue arrows are drawn based on the results of the differential features. For example, if subgroup 2 (Subgroup 1) is older than subgroup 1 (Subgroup 2), a red arrow is added next to the age; if subgroup 3 (Subgroup 3) is older than subgroup 2, two red arrows are added; if the age decreases, a blue arrow is added, and so on. For the visualization effect of this embodiment, see [link to visualization]. Figure 8As shown, the network includes four exposure factors, six laboratory indicators, and three identified CVD mortality subgroups. A single red arrow in the figure indicates an increase in value, two arrows indicate a further increase, and a blue arrow indicates a decrease. Results show that age mediates the transition between different subgroups by affecting hemoglobin, hematocrit, lymphocyte percentage, neutrophil percentage, and osmolarity; sex affects subgroup transition by affecting lymphocyte percentage, osmolarity, and mean corpuscular volume; marital status affects subgroup transition by mediating changes in lymphocyte percentage, neutrophil percentage, osmolarity, and mean corpuscular volume; and household income and poverty ratio primarily affect subgroup transition by mediating changes in hemoglobin and hematocrit.

[0091] For details, see Figure 9 The diagram shown is a structural schematic of a CVD death subgroup identification device combining causal reasoning and consensus clustering according to an embodiment of the present invention, comprising:

[0092] The data collection module 901 collects raw feature data and preprocesses it to obtain preprocessed feature data. The feature variables in the preprocessed feature data include exposure factors and laboratory indicators.

[0093] The causal reasoning module 902 performs mediating causal reasoning on the preprocessed feature data to obtain the exposure factors and laboratory indicators contained in the mediating causal reasoning results as the initial feature set;

[0094] The model building module 903 uses several machine learning models to predict mortality risk based on the input feature data, and then trains each model using an initial feature set to obtain several mortality risk prediction models. The SHAP algorithm is used to calculate the SHAP value of the feature variables of all mortality risk prediction models, and the optimal model and optimal feature set are selected using the SHAP value of the feature variables of each mortality risk prediction model.

[0095] The subgroup acquisition module 904 performs consensus clustering on the optimal feature set and the mortality risk prediction results output by the optimal model to obtain several mortality subgroups.

[0096] The subpopulation identification module 905 performs statistical analysis on the characteristic data and survival outcomes of each mortality subpopulation, and identifies the pathways of each mortality subpopulation by combining the statistical analysis results and the results of mediation causal relationship inference.

[0097] In this embodiment, firstly, mediating causal inference reveals potential pathways to CVD mortality, providing important reference for medical institutions to reduce CVD mortality and optimize prevention strategies. Secondly, by combining machine learning (ML) technology, the 10 most critical characteristic variables in the CVD mortality pathway are identified. These characteristics are all common clinical indicators, which effectively reduce model complexity while ensuring model performance, laying a good foundation for the model's promotion in the real world. Finally, these 10 clinical variables can be further used for risk stratification of the CVD mortality population, and reveal the potential causal mechanism of mutual transformation between different CVD mortality subgroups.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying CVD death subgroups combining causal reasoning and consensus clustering, characterized in that, Includes the following steps: Collect raw feature data and preprocess it to obtain preprocessed feature data; The original feature data includes basic demographic information, anthropometric data, laboratory indicators, and questionnaire survey data; the feature variables in the preprocessed feature data include exposure factors and laboratory indicators. Mediation causal reasoning is performed on the preprocessed feature data, using the exposure factors and laboratory indicators contained in the mediation causal reasoning results as the initial feature set; Several machine learning models are used to predict mortality risk based on the input feature data. Each model is then trained using an initial feature set to obtain several mortality risk prediction models. The SHAP algorithm is used to calculate the SHAP value of the feature variables of all mortality risk prediction models. The optimal model and the optimal feature set are selected using the SHAP value of the feature variables of each mortality risk prediction model. Consensus clustering is performed on the optimal feature set and the mortality risk prediction results output by the optimal model to obtain several mortality subgroups; Statistical analysis was performed on the characteristic data and survival outcomes of each mortality subgroup, and the pathways of each mortality subgroup were identified by combining the statistical analysis results with the results of mediation causal relationship inference. The method of inferring mediated causal relationships from preprocessed feature data includes the following steps: The K-nearest neighbor algorithm is used to fill in the preprocessed feature data; Exhaustive combinations of exposure factors and laboratory indicators are performed; for each pair of exposure factors and laboratory indicators, verification is carried out based on several conditionally independent tests to obtain several P-values, and the maximum value among them is taken as the overall P-value. If the overall P-value is less than 0.05, it indicates that the exposure factor can influence CVD mortality through this laboratory indicator as a mediating variable. The aforementioned conditionally independent tests include: The exposure factor was significantly associated with mortality outcomes; After adjusting for mortality outcomes, the exposure factor and the mediating variable remained significantly associated; the mediating variable refers to laboratory indicators. After adjusting for exposure factors, the mediating variable remained significantly associated with mortality outcomes; After adjusting for mediating variables, the exposure factor was not associated with mortality outcomes; The SHAP algorithm is used to calculate SHAP values ​​for the feature variables of all mortality risk prediction models. The optimal model and optimal feature set are then selected using the SHAP values ​​of the feature variables for each mortality risk prediction model. Specifically: The SHAP algorithm was used to calculate the SHAP values ​​for the feature variables of all mortality risk prediction models. For each mortality risk prediction model, the number of input features is gradually reduced based on the SHAP value of the feature variables, and the prediction effect is evaluated. The optimal model is selected by combining the prediction effect and the number of feature variables; the set of input feature variables of the optimal model is the optimal feature set.

2. The CVD death subgroup identification method combining causal reasoning and consensus clustering according to claim 1, characterized in that, The preprocessing includes: Remove feature variables with more than 50% missing values; Data from subjects under 20 years of age were excluded. Feature data of subjects with more than 50% missing values ​​were removed; Each subject was categorized based on their survival status at different time points.

3. The CVD death subgroup identification method combining causal reasoning and consensus clustering according to claim 2, characterized in that, The classification of each subject based on their survival status at different time points includes: At the three-year follow-up, the subjects were classified into the survival group or the death group based on whether they survived. At the five-year follow-up, subjects were classified into the survival group or the death group based on whether they survived. At the ten-year follow-up, subjects were categorized into the survival group or the death group based on whether they were alive.

4. The CVD death subgroup identification method combining causal reasoning and consensus clustering according to claim 1, characterized in that, The optimal feature set includes four exposure factors and six laboratory indicators; the four exposure factors are age, household income and poverty ratio, gender and marital status; the six laboratory indicators are lymphocyte percentage, neutrophil percentage, hemoglobin, hematocrit, mean corpuscular volume, and osmolarity; the optimal model is a logistic regression model with the optimal feature set as input feature variables.

5. The CVD death subgroup identification method combining causal reasoning and consensus clustering according to claim 1, characterized in that, The consensus clustering of the mortality risk prediction results output by the optimal feature set and the optimal model to obtain several mortality subgroups includes the following steps: Input the feature values ​​of the optimal feature set into the optimal prediction model to obtain the mortality risk prediction result; Consensus clustering is performed on the eigenvalues ​​and the prediction results to obtain three death subgroups.

6. The CVD death subgroup identification method combining causal reasoning and consensus clustering according to claim 1, characterized in that, The statistical analysis of the characteristic data and survival outcomes of each mortality subgroup includes: Survival analysis was performed, using Kaplan-Meier survival curves to estimate the survival probability of each mortality subgroup, and the Log-rank test was used to compare the differences in survival curves between groups. Descriptive analysis uses median and quartiles for continuous variables and frequency and percentage for categorical variables; For comparisons between groups, the Mann-Whitney U test is used for continuous variables and the chi-square test or Fisher's exact test is used for categorical variables. For multiple comparisons between three groups, Bonferroni correction is used to adjust the p-value to control for Type I error. The continuous variables include age, household income to poverty ratio, percentage of lymphocytes, percentage of neutrophils, hemoglobin, hematocrit, mean corpuscular volume, and osmolarity; the categorical variables include gender and marital status.

7. The CVD death subgroup identification method combining causal reasoning and consensus clustering according to claim 1, characterized in that, The pathways for identifying each mortality subgroup are determined by combining statistical analysis results and mediating causal reasoning results. The specific pathways for each mortality subgroup are as follows: Age affects the transformation between different subsets by influencing hemoglobin, hematocrit, lymphocyte percentage, neutrophil percentage, and osmotic pressure. Sex affects the transformation between different subsets by influencing lymphocyte percentage, osmotic pressure, and mean erythrocyte volume. Marital status influences the transition between different subsets by affecting changes in lymphocyte percentage, neutrophil percentage, osmolarity, and mean erythrocyte volume. Household income and poverty ratios influence the transition between different subgroups by affecting changes in hemoglobin and hematocrit.

8. A CVD death subgroup identification device combining causal reasoning and consensus clustering, characterized in that, include: The data collection module collects raw feature data and preprocesses it to obtain preprocessed feature data; The original feature data includes basic demographic information, anthropometric data, laboratory indicators, and questionnaire survey data; the feature variables in the preprocessed feature data include exposure factors and laboratory indicators. The causal reasoning module performs mediating causal reasoning on the preprocessed feature data, using the exposure factors and laboratory indicators contained in the mediating causal reasoning results as the initial feature set. The model building module uses several machine learning models to predict mortality risk based on the input feature data, and then trains each model using an initial feature set to obtain several mortality risk prediction models. The SHAP algorithm is used to calculate the SHAP value of the feature variables of all mortality risk prediction models, and the optimal model and optimal feature set are selected using the SHAP value of the feature variables of each mortality risk prediction model. The subgroup acquisition module performs consensus clustering on the optimal feature set and the mortality risk prediction results output by the optimal model to obtain several mortality subgroups. The subpopulation identification module performs statistical analysis on the characteristic data and survival outcomes of each mortality subpopulation, and identifies the pathways of each mortality subpopulation by combining the statistical analysis results and the results of mediation causal relationship inference. The method of inferring mediated causal relationships from preprocessed feature data includes the following steps: The K-nearest neighbor algorithm is used to fill in the preprocessed feature data; Exhaustive combinations of exposure factors and laboratory indicators are performed; for each pair of exposure factors and laboratory indicators, verification is carried out based on several conditionally independent tests to obtain several P-values, and the maximum value among them is taken as the overall P-value. If the overall P-value is less than 0.05, it indicates that the exposure factor can influence CVD mortality through this laboratory indicator as a mediating variable. The aforementioned conditionally independent tests include: The exposure factor was significantly associated with mortality outcomes; After adjusting for mortality outcomes, the exposure factor and the mediating variable remained significantly associated; the mediating variable refers to laboratory indicators. After adjusting for exposure factors, the mediating variable remained significantly associated with mortality outcomes; After adjusting for mediating variables, the exposure factor was not associated with mortality outcomes; The SHAP algorithm is used to calculate SHAP values ​​for the feature variables of all mortality risk prediction models. The optimal model and optimal feature set are then selected using the SHAP values ​​of the feature variables for each mortality risk prediction model. Specifically: The SHAP algorithm was used to calculate the SHAP values ​​for the feature variables of all mortality risk prediction models. For each mortality risk prediction model, the number of input features is gradually reduced based on the SHAP value of the feature variables, and the prediction effect is evaluated. The optimal model is selected by combining the prediction effect and the number of feature variables; the set of input feature variables of the optimal model is the optimal feature set.

Images