Acute myocardial infarction and atrial fibrillation coexistence risk prediction model construction method and application

Abstract

The invention provides a prognosis risk prediction model construction method based on coexistence of acute myocardial infarction and atrial fibrillation and application. The method comprises the following steps: S1, acquiring historical patient clinical information of a target patient and death time after coexistence of diseases; s2, preprocessing clinical information of historical patients; s3, using an LASSO method and a ten-fold cross validation method to identify related clinical features related to death time after coexistence of diseases; s4, dividing a training set and a verification set; s5, performing training optimization on the four machine models by using the training set, evaluating the prediction performance of the models by using the verification set, and determining a Glmnet model with the optimal prediction performance; and S6, performing model interpretation on the Glmnet model by using an SHAP value, and determining an optimal prediction factor of which the model prediction contribution ranks the first three. The prognosis risk of coexistence of acute myocardial infarction and atrial fibrillation can be accurately predicted, and reference is provided for clinical diagnosis and treatment of patients with coexistence of acute myocardial infarction and atrial fibrillation.

Description

Technical Field

[0001] The present invention belongs to the technical field of prediction model construction, and specifically relates to a method for constructing a prognosis risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation and its application. Background Art

[0002] Acute myocardial infarction (AMI), as a serious complication of coronary artery disease, has become a major health threat worldwide. Atrial fibrillation (AF) is one of the most common arrhythmias and is often associated with a variety of cardiovascular diseases such as heart failure, stroke and AMI. Research data show that about 4% to 12% of AMI patients will have AF. The two diseases have a clear tendency to coexist and affect each other's disease progression. Clinical observations have found that compared with patients with simple AMI, AMI patients with AF have more common lesions at the base of the right coronary artery (RCA), a higher risk of in-hospital death, and worse short-term and long-term prognosis. Myocardial infarction (MI) and atrial fibrillation (AF) each lead to poor prognosis, and when the two diseases coexist, the patient's risk will be further aggravated. Therefore, for this special patient group with AMI and AF, it is particularly important to accurately predict the prognostic risk, and accurate prediction of the prognostic risk can provide a reference for clinical diagnosis and treatment. However, there is still a lack of research on prognosis prediction for this type of patients. Summary of the invention

[0003] The present invention provides a method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation and its application to solve the above problems.

[0004] In a first aspect, the present invention provides a method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation, the method comprising the following steps: S1. Based on the preset data extraction rules, the historical clinical information of the target patients and the time of death after the coexistence of the disease are obtained from the online hospital comprehensive management system. The target patients are patients with coexistence of acute myocardial infarction and atrial fibrillation; S2. Preprocessing the historical clinical information of patients, classifying the preprocessed historical clinical information of patients into basic patient information, laboratory test information, AMI and AF comorbidity information, and clinical treatment measure information; S3. Use the LASSO method and the ten-fold cross-validation method to identify the associated clinical characteristics related to the time of death after the coexistence of the disease from the patient's basic information, laboratory test information, AMI and AF comorbidity information, and clinical treatment measures information. The associated clinical characteristics include risk factors and protective factors. The risk factors are pulse, serum creatinine, and neutrophils. The protective factors are hemoglobin, ACEI / ARB, aspirin, and reperfusion therapy. S4. Divide the preprocessed historical patient clinical information into a training set and a validation set; S5. Use the training set to train and optimize four machine learning models, namely, the logistic regression model, the naive Bayes model, the AdaBoost model, and the Glmnet model. Use the validation set to evaluate the prediction performance of the four machine learning models and determine the Glmnet model with the best prediction performance. S6. The SHAP value was used to interpret the Glmnet model and determine the top three optimal predictors that contributed to the model prediction among the associated clinical characteristics. The optimal predictors were pulse, hemoglobin, and neutrophils.

[0005] Optionally, the data extraction rules in step S1 include: an upper and lower limit of the data extraction quantity; the coexistence of acute myocardial infarction and atrial fibrillation; the proportion of data whose death time exceeds one year after the coexistence of the symptoms; and the proportion of data whose death time does not exceed one year after the coexistence of the symptoms.

[0006] Optionally, in step S1, the online hospital comprehensive management system integrates a HIS system, a LIS system and an EMR system.

[0007] Optionally, the basic information of the patient in step S2 includes age, gender, heart rate, pulse, systolic blood pressure, diastolic blood pressure and infarction type, the laboratory test information includes white blood cells, lymphocytes, monocytes, hemoglobin, platelets, neutrophils, ALT, AST, Glu, TG, TC, LDL-C, HDL-C, blood creatinine, plasma protein and globulin, the AMI and AF comorbidity information includes hypertension, type 2 diabetes, hyperlipidemia, renal dysfunction, stroke, paroxysmal atrial fibrillation and left atrial enlargement, and the clinical treatment measures information includes reperfusion therapy, aspirin, clopidogrel, statins, ACEI / ARB, beta-blockers and diuretics.

[0008] Optionally, in step S4, 70% of the preprocessed historical patient clinical information is divided into a training set, and 30% of the preprocessed historical patient clinical information is divided into a validation set.

[0009] Optionally, in step S5, when the training set is used for training optimization, the logistic regression C parameter of the logistic regression model is 2.07, the random parameter is 42; the variable smoothing parameter of the Gaussian naive Bayes classifier in the naive Bayes model is 3.85×10 -10 ; The learning rate of the AdaBoost classifier in the AdaBoost model is 0.22, the number of weak evaluators is 53, and the random parameter is 42; the alpha value of the cross-validation classifier in the Glmnet model is 98.69, and the random parameter is 42.

[0010] Optionally, in step S5, the prediction performance is evaluated by the Brier score, the AUC value and the 95% confidence interval of the AUC value, the calibration curve and the decision curve.

[0011] Optionally, in step S6, the prediction contributions of the optimal prediction factors are pulse, hemoglobin, and neutrophils from large to small.

[0012] In a second aspect, the present invention also provides a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation constructed according to the method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation described in the first aspect.

[0013] In a third aspect, the present invention further provides an application of the prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation described in the second aspect in evaluating the prognostic risk prediction of patients with coexistence of acute myocardial infarction and atrial fibrillation.

[0014] The beneficial effects of the present invention are: The present invention effectively solves the problem of insufficient research on prognosis prediction for patients with acute myocardial infarction and atrial fibrillation in the prior art through systematic data analysis and machine learning methods, and has significant innovation and practical value. The present invention obtains the clinical information of target patients from the hospital comprehensive management system through preset data extraction rules, ensuring the reliability and integrity of the data source. Through the scientific classification and preprocessing of historical patient clinical information, complex medical data is converted into structured feature information, laying the foundation for subsequent analysis. In particular, the combination of the LASSO method and ten-fold cross validation not only achieves accurate identification of associated clinical features, but also clearly distinguishes risk factors and protective factors. This classification has important guiding significance for clinical practice.

[0015] On the other hand, the present invention constructs multiple machine learning models including logistic regression, naive Bayes, AdaBoost and Glmnet. Through a rigorous training and verification process, the Glmnet model with the best prediction performance is finally determined, thereby improving the accuracy and reliability of the prediction. More importantly, the scheme uses SHAP values ​​to interpret the model and identifies the three best predictors of pulse, hemoglobin and neutrophils, which not only improves the interpretability of the model, but also provides reference indicators for clinicians. These advantages enable the present invention to provide clinicians with a more accurate prognosis risk assessment tool, which helps to formulate personalized treatment plans and improve the efficiency of patient prognosis management. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a LASSO regularized path diagram of all historical patient clinical information in one implementation of the present application.

[0017] Figure 2 This is a LASSO coefficient path diagram of all historical patient clinical information in one implementation method of the present application.

[0018] Figure 3 This is a schematic diagram of the correlation coefficients between 7 related clinical characteristics in one embodiment of the present application.

[0019] Figure 4 This is a schematic diagram of the calculation results of the AUC value and its 95% confidence interval of each model in one embodiment of the present application.

[0020] Figure 5 It is a schematic diagram of the calibration curves of each model in one embodiment of the present application.

[0021] Figure 6 It is a schematic diagram of the decision curves of each model in one embodiment of the present application.

[0022] Figure 7 This is a heat map of SHAP values ​​of 7 related clinical characteristics in one embodiment of the present application.

[0023] Figure 8 This is a schematic diagram of the analysis results of importance analysis of 7 related clinical characteristics based on the SHAP summary diagram in one embodiment of the present application.

[0024] Fig. 9 This is a diagram showing the front-end display and application process of a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation in one of the embodiments of the present application. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0026] The present invention discloses a method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation, which specifically comprises the following steps: S1. Obtain the historical patient clinical information and death time after coexistence of symptoms of the target patient from the online hospital comprehensive management system based on the preset data extraction rules.

[0027] Among them, the target patients are those with coexisting acute myocardial infarction and atrial fibrillation. The HIS system provides the patient's medical information and basic medical records, the LIS system provides laboratory test data, and the EMR system provides complete electronic medical records. Data extraction must follow strict rules. In terms of patient screening criteria, the diagnostic criteria for acute myocardial infarction (AMI) and atrial fibrillation (AF) are strictly defined. The diagnosis of AMI is based on the following conditions: typical chest pain symptoms, electrocardiogram ST segment changes (such as ST segment elevation or depression), and significant increases in myocardial markers (such as troponin, CK-MB, etc.). The diagnosis of AF is based on electrocardiogram manifestations: atrial electrical activity disorder, irregular RR interval, lack of regular P waves and other characteristics. The determination of coexistence of symptoms requires ensuring that the two diseases exist in the same patient at the same time and have been clearly diagnosed.

[0028] In terms of the distribution of prognosis time, a stratified sampling method was used, and the time of death after the coexistence of symptoms was used as the key stratification indicator. Specifically, the data of death time exceeding one year after the coexistence of symptoms and the data of death time less than one year each accounted for a specific proportion. This stratified design ensured the representativeness of the sample and the balance of the prediction model, so that the prediction model can accurately predict the mortality rate of the target patients within one year. In the process of data extraction, ICD-10 disease classification codes (AMI is I21 series, AF is I48 series) were used for precise screening, and cross-validation was performed in combination with clinical diagnosis records.

[0029] The extracted data covers the complete clinical information chain, including but not limited to: Extract from the HIS system: basic patient information (hospitalization number, consultation time, etc.), doctor's advice information, cost information, etc.; Extract from the LIS system: various test results to ensure the timeliness and integrity of the data; Extract from the EMR system: detailed medical records, treatment process records, surgical records, etc.

[0030] S2. Preprocess the historical clinical information of patients and classify the preprocessed historical clinical information of patients into basic patient information, laboratory test information, AMI and AF comorbidity information, and clinical treatment measure information.

[0031] The extracted historical patient clinical information is systematically preprocessed and accurately classified according to the preset classification standards. The preprocessing process includes three main links: data cleaning, standardization, and structured processing. Each link has its own specific processing methods and quality control standards. First, the data cleaning process includes the following specific steps: 1. Data integrity check: The missing value statistics of each variable are calculated, and the missing rate is calculated. For numerical variables with a missing rate of less than 20%, multiple imputation is used, specifically using the MICE (Multiple Imputation by Chained Equations) algorithm, and the interpolation is performed considering the correlation between variables. The mode is used to fill the categorical variables, but missing marker variables are added to retain the missing information. Variables with a missing rate of more than 20% are specially treated, including deleting or creating missing categories.

[0032] 2. Identification and processing of outliers: The box plot method was used to mark values ​​that exceeded Q3+1.5IQR or were less than Q1-1.5IQR as potential outliers. Combined with the opinions of clinical experts, a reasonable value range for each test indicator was established. For confirmed outliers, truncation or interpolation was used according to the specific circumstances. All outlier processing operations were recorded to ensure the traceability of data processing.

[0033] 3. Data standardization processing: Perform z-score standardization on continuous variables: (x-μ) / σ, consider logarithmic transformation for variables with skewed distribution, encode categorical variables, select one-hot encoding or label encoding according to the characteristics of the variables, unify the time format, and convert all time information into a standard format.

[0034] The preprocessed data is strictly classified and organized according to the following four categories: 1. The specific contents of the patient's basic information include: 1. Demographic characteristics: age (accurate to years), gender (male / female); 2. Vital signs: heart rate (beats / min), pulse (beats / min), systolic blood pressure (mmHg), diastolic blood pressure (mmHg); 3. Disease characteristics: infarction type (such as anterior wall, inferior wall, etc.).

[0035] 2. Laboratory testing information includes: 1. Blood cell count: white blood cells, lymphocytes, monocytes, neutrophils; 2. Hemoglobin and platelets; 3. Liver function: ALT, AST; 4. Blood sugar: Glu; 5. Blood lipids: TG, TC, LDL-C, HDL-C; 6. Renal function: serum creatinine; 7. Protein levels: plasma protein, globulin.

[0036] 3. Information on complications of AMI and AF includes: 1. Underlying diseases: hypertension, type 2 diabetes, hyperlipidemia; 2. Organ function: renal dysfunction; 3. Complications: stroke; 4. Heart condition: paroxysmal atrial fibrillation, left atrial enlargement.

[0037] 4. Clinical treatment measures information includes: 1. Interventional therapy: reperfusion therapy; 2. Drug treatment: 1) Antiplatelet drugs: aspirin, clopidogrel; 2) Lipid-lowering drugs: statins; 3) Cardiovascular drugs: ACEI / ARB, beta-receptor blockers; 4) Others: diuretics.

[0038] This study included data from 249 patients with coexisting AMI and AF, of whom 179 died more than one year after the coexistence of the disease, and 70 died within one year after the coexistence of the disease. For each characteristic data, the median, interquartile range (IQR) and P value were calculated, and four data lists were finally integrated, as shown in Tables 1-4:

[0039] Table 1 Distribution table of basic information characteristics of patients

[0040] Table 2 Laboratory test information characteristic data distribution table

[0041] Table 3 Distribution of information characteristics of AMI and AF comorbidities

[0042] Table 4. Distribution table of clinical treatment measures information characteristics S3. The LASSO method and ten-fold cross-validation method were used to identify the associated clinical characteristics related to the time of death after the coexistence of the disease from the patient's basic information, laboratory test information, AMI and AF comorbidity information, and clinical treatment measures information.

[0043] In the feature recognition stage, the LASSO (Least Absolute Shrinkage and Selection Operator) method combined with ten-fold cross validation was used to screen the associated clinical features. The LASSO method unifies feature selection and parameter estimation by introducing the L1 regularization term, and its objective function is defined as:

[0044] Where n represents the number of samples, p represents the number of features, y is the target variable (i.e., time of death), x is the feature variable, β is the regression coefficient, and λ is the penalty parameter. The main reason for choosing the LASSO method is that it can automatically perform variable selection and parameter shrinkage, while dealing with multicollinearity problems. This method has significant advantages for high-dimensional data analysis. In actual operation, four types of clinical information (basic patient information, laboratory test information, AMI and AF comorbidity information, and clinical treatment measures information) are screened for features, and the optimal penalty parameter λ is determined by ten-fold cross validation. Specifically, the data is randomly divided into 10 subsets of equal size, 9 subsets are used for training each time, and the remaining 1 subset is used for validation. This is repeated 10 times to ensure that each subset is used as a validation set once. During the cross-validation process, the LASSO path of each feature under different λ values ​​is calculated:

[0045] At the same time, the stability index of feature selection is calculated:

[0046] Where I is the indicator function, represents the coefficient of feature j in the k-fold cross validation.

[0047] Calculation results refer to Figure 1 and Figure 2 , Figure 1 The figure shows how the model performance changes with the penalty parameter λ in the LASSO regression process. The horizontal axis represents the logarithm of λ, and the vertical axis represents the average error of the model. The solid red line represents the average error, and the dashed blue line represents the position of the optimal λ value. As can be seen from the figure, as the λ value increases (i.e., moves to the left), the average error of the model first decreases and then increases, showing a typical "U"-shaped curve. This trend reflects the trade-off between model complexity and goodness of fit in LASSO regression. When λ is small, the model may be too complex and prone to overfitting; when λ is too large, the model may be too simple and cannot capture important features in the data. The blue dashed line in the figure marks the optimal λ value, that is, the λ value that produces the minimum average error during the cross-validation process. The selection of this optimal λ value is achieved through ten-fold cross-validation, which balances the complexity of the model and the accuracy of prediction. It is worth noting that when the λ value is large (left side of the figure), the average error is relatively stable, which indicates the performance of the model under a highly regularized state. As the λ value decreases, the average error begins to decrease significantly until it reaches the optimal point. After that, the error begins to rise again, which reflects that the model may begin to overfit.

[0048] Figure 2The trajectory of each feature coefficient in LASSO regression as the penalty parameter λ changes is presented. The horizontal axis also represents the logarithm of λ, and the vertical axis represents the size of the coefficient of each feature. Each colored line represents a feature, and the trajectory of the line shows the trend of the feature coefficient as λ changes. It can be observed from the figure that when the λ value is large (left side of the figure), the coefficients of most features are compressed to zero, which reflects the feature selection ability of LASSO. As the λ value decreases, the coefficients of different features gradually become non-zero, but their change speed and direction are different. The coefficients of some features increase rapidly, indicating that they may be important predictors; while the coefficients of other features remain close to zero, indicating that they are less important in the model. It is particularly noteworthy that the coefficients of some features change their signs as λ decreases, which may reflect the complex interactions between features or the existence of multicollinearity. It can also be observed in the figure that the coefficients of some features show nonlinear changes, which further confirms the advantage of LASSO in dealing with complex feature relationships.

[0049] In summary, clinical features with significant predictive value were identified through the calculation of the LASSO pathway, and these features can be divided into two categories: risk factors and protective factors. Risk factors include pulse, serum creatinine, and neutrophils, and the increase of these indicators often indicates a poor prognosis. Specifically, an increased pulse reflects the compensatory state of cardiac function and may indicate an increased risk of heart failure or heart insufficiency; an increase in serum creatinine indicates renal impairment, which not only reflects the impaired function of important organs, but may also affect drug metabolism and clearance; an increase in neutrophil count indicates the presence of an inflammatory response, which may be related to the degree of myocardial damage and the risk of complications. Protective factors include hemoglobin, ACEI / ARB, aspirin, and reperfusion therapy, and the presence or reasonable use of these factors is associated with improved prognosis. Appropriate hemoglobin levels ensure sufficient oxygen transport capacity and help maintain oxygen supply to vital organs; ACEI / ARB drugs improve ventricular remodeling and reduce the risk of heart failure by inhibiting the renin-angiotensin system; aspirin reduces the occurrence of thrombotic events through its antiplatelet effect; and timely reperfusion therapy can save ischemic myocardium and reduce the area of ​​infarction.

[0050] Special attention was paid to clinical interpretability during feature selection to ensure that the selected features were not only statistically significant but also had clear clinical significance. In order to verify the reliability of feature selection, a stability selection method was used, that is, the feature selection process was repeated on multiple resampled data sets, and only the features that were selected in most resamples were retained. At the same time, considering the possible interactions between features, the interaction terms of important features were also evaluated during the feature selection process. The evaluation results are shown in Figure 2. Figure 3As shown. For continuous features, their nonlinear relationships were examined, and spline functions or polynomial terms were introduced when necessary to capture nonlinear associations. Through this systematic feature identification process, the finally determined associated clinical features provide a reliable feature basis for the subsequent prediction model construction. These features not only have good predictive ability, but also facilitate monitoring and intervention in clinical practice. The results of feature selection have been reviewed by medical experts to ensure the rationality and feasibility of its clinical application. The final selected feature set showed good predictive efficiency in the subsequent prediction model, confirming the effectiveness of the feature selection method.

[0051] S4. Divide the preprocessed historical patient clinical information into a training set and a validation set.

[0052] The data set was divided into 70% training set and 30% validation set. This ratio is widely recognized in the field of machine learning and can provide sufficient test samples for model validation while ensuring the adequacy of model training. The specific division process uses stratified random sampling to ensure the consistency of the distribution of various samples in the training set and the validation set. Before the division, the entire data set was first randomized, and the data order was disrupted using a random number generator to eliminate possible timing effects or selection bias. Special attention was paid to the following key points during the division process: First, the consistency of the distribution of death time was maintained, ensuring that the proportion of cases with a death time of more than one year and less than one year in the training set and the validation set was similar; second, ensure that important prognostic factors (such as age group, gender, comorbidity type, etc.) were evenly distributed in the two data sets; third, considering the time evolution characteristics of clinical treatment measures, ensure that cases in different periods can be reasonably distributed in the two data sets.

[0053] To ensure the scientific nature of data partitioning, the following specific steps were adopted: First, the distribution of each key feature in the overall data set was calculated, including the mean and standard deviation of continuous variables, and the proportion of each category of categorical variables. Then, the stratified random sampling technique was used to stratify the samples according to key stratification variables (such as death time, age group, gender, etc.), and samples were randomly selected in each layer at a ratio of 7:3 to be assigned to the training set and the validation set. Statistical tests were performed on the partitioning results, including t-test or rank sum test for continuous variables and chi-square test for categorical variables to ensure that there were no statistically significant differences between the two data sets in various important features. In order to verify the rationality of the partitioning, the distribution differences between the training set and the validation set in each feature dimension were calculated: for continuous variables, the standardized mean difference (SMD) was calculated; for categorical variables, the proportion differences of each category were compared. This systematic evaluation ensured the scientific nature and reliability of data partitioning and provided a reliable data basis for subsequent model training and validation. At the same time, the random seed of the partitioning scheme was saved to ensure the reproducibility of the study.

[0054] In addition, considering the data imbalance problem that may exist in clinical practice, the distribution of minority samples was also paid attention to during the division process to ensure that they were sufficiently representative in both the training set and the validation set. For some rare but clinically significant cases, special sampling strategies were used to ensure that they were reasonably distributed in both data sets. After the division was completed, the basic statistical characteristics of the two data sets were recorded in detail, including sample size, distribution of various features, etc. This information is of great reference value for understanding the subsequent model training and validation results.

[0055] The final data set division results and the corresponding P values ​​are shown in Tables 5 to 8 below:

[0056] Table 5 Results of dividing the training set and validation set of patients’ basic information

[0057] Table 6 Division results of training set and validation set for laboratory test information

[0058] Table 7. Results of training set and validation set division for AMI and AF comorbidity information

[0059] Table 8 Division results of training set and validation set for clinical treatment information S5. Use the training set to train and optimize four machine learning models: logistic regression model, naive Bayes model, AdaBoost model and Glmnet model. Use the validation set to evaluate the prediction performance of the four machine learning models and determine the Glmnet model with the best prediction performance.

[0060] In the model training stage, the logistic regression model, naive Bayes model, AdaBoost model and Glmnet model were trained and optimized respectively. The parameter settings of each model have clear theoretical basis, and the optimal parameter combination is determined by repeated verification. The logistic regression model uses a regularization parameter of C=2.07 and a random parameter of 42, where the selection of the C value is based on the trade-off between model complexity and generalization ability; the naive Bayes model uses a Gaussian naive Bayes classifier, and sets the variable smoothing parameter to 3.85×10-10, which is obtained by grid search optimization; the AdaBoost model sets the learning rate to 0.22, the number of weak evaluators to 53, and the random parameter to 42. The selection of these parameters is intended to balance the learning speed and accuracy of the model; the Glmnet model uses a cross-validation classifier, sets the alpha value to 98.69, and the random parameter to 42, and determines the optimal regularization strength through cross-validation.

[0061] The model performance evaluation uses a multi-dimensional indicator system, including Brier score, AUC value and its 95% confidence interval, calibration curve and decision curve analysis. The Brier score is used to evaluate the accuracy of the predicted probability, and the calculation formula is:

[0062] Where N is the number of samples, is the predicted probability, is the actual observed value. The AUC value reflects the discriminatory ability of the model, and its 95% confidence interval is calculated by bootstrap, and the number of repeated sampling is set to 1000 times to ensure statistical stability. The calibration curve is used to evaluate the calibration of the predicted probability. The Hosmer-Lemeshow test is used to evaluate the calibration by comparing the relationship between the predicted probability and the actual observed frequency. Decision curve analysis (DCA) is used to evaluate the clinical net benefit of the model at different decision thresholds. The calculation formula is:

[0063] Among them, pt is the decision threshold and n is the total sample size.

[0064] During training, the logistic regression model uses the following optimization objective function:

[0065] The naive Bayes model is based on Bayes' theorem and assumes that the features are conditionally independent:

[0066] The AdaBoost model adjusts sample weights in an iterative manner, and the weight update formula for each iteration is:

[0067] in, is the weight of the i-th sample in the t-th iteration, is the prediction result of the weak classifier, is the normalization factor.

[0068] The Glmnet model uses elastic net regularization, and its objective function is:

[0069] During the model training process, K-fold cross validation (K=10) was used to evaluate the stability of the model performance, and the division of each fold maintained a consistent category distribution. For each model, the learning curve during the training process was recorded to monitor whether the model was overfitting or underfitting. To address the problem of unbalanced data, a category weight adjustment strategy was adopted to ensure that the model has sufficient learning attention to minority class samples. The model evaluation adopted a binary classification standard with a death time of one year, and calculated the sensitivity, specificity, precision, recall and other indicators of each model. By comparing and analyzing the performance of each model on the validation set, it was finally determined that the Glmnet model had the best predictive performance. This choice was based on a comprehensive consideration of the model's discrimination ability, calibration performance and clinical application value. For specific model evaluation results, please refer to Figure 4-Figure 6 , Figure 4 The overall performance of each model is shown, and the Glmnet model achieves an AUC of 0.87. Figure 5 and Figure 6 The calibration curve and decision curve of each model are shown separately.

[0070] S6. The SHAP value was used to interpret the Glmnet model and determine the top three optimal predictors that contributed to the model prediction among the associated clinical characteristics. The optimal predictors were pulse, hemoglobin, and neutrophils.

[0071] The Glmnet model is explained using the SHAP (SHapley Additive exPlanations) value, a game theory-based approach that can accurately quantify the contribution of each feature to the model's prediction results. The calculation of the SHAP value is based on the following mathematical principle: for feature i, its SHAP value The calculation formula is:

[0072] Among them, F is the set of all features, S is the feature subset that does not contain feature i, is the model prediction value when only the feature subset S is used. By calculating all possible feature combinations, the marginal contribution of each feature is obtained. The SHAP value has important mathematical properties: efficiency, symmetry, dummy, and additivity.

[0073] In this study, the SHAP value calculation results and feature importance analysis results are as follows: Figure 7 and Figure 8 As shown in the figure, three optimal predictors were determined through SHAP value analysis: pulse, hemoglobin and neutrophils, and their contribution to the model prediction results was ranked from large to small. For each predictor, its global and local SHAP values ​​were calculated: the global SHAP value reflects the average influence of the feature on the entire data set, and the local SHAP value shows the influence of the feature on the prediction results of a specific sample. The specific analysis process includes: First, calculate the mean absolute SHAP value for each feature:

[0074] Where n is the number of samples, is the SHAP value of feature i on the jth sample. Then, a SHAP dependency graph is generated to show the relationship between the change in feature value and its prediction contribution:

[0075] Specific analysis of the optimal predictor: 1. Pulse (main predictor): SHAP interaction values ​​were used to analyze its synergistic effect with other features, calculate the average prediction contribution of different pulse intervals, and identify key threshold points and their clinical significance.

[0076] 2. Hemoglobin (secondary predictor): Analyze its nonlinear relationship with patient prognosis, evaluate its predictive value under different disease severity, and determine the optimal reference range.

[0077] 3. Neutrophils (the third predictor): Study the association between inflammatory response and prognostic risk, evaluate the interaction with other inflammatory indicators, and determine the risk warning threshold.

[0078] To further understand the mechanism of action of these predictors, the following additional analyses were performed in this study: 1. Feature interaction analysis:

[0079] in, represents the interaction effect of features i and j.

[0080] 2. Conditional expectation analysis:

[0081] Study the conditional dependencies between features.

[0082] 3. Cumulative effect analysis:

[0083] Evaluate the cumulative impact of changes in eigenvalues.

[0084] The above analysis reveals the characteristics of the three optimal predictive factors: 1. Pulse reflects the state of cardiovascular function, and its variability is positively correlated with prognostic risk; 2. Hemoglobin levels are directly related to tissue oxygen supply capacity, showing a U-shaped risk relationship; 3. Neutrophil count reflects the inflammatory state and is positively correlated with the risk of poor prognosis.

[0085] The present invention also discloses a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation obtained by constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation, and an application of the prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation in evaluating the prognostic risk prediction of patients with coexistence of acute myocardial infarction and atrial fibrillation.

[0086] Reference Fig. 9 After building the model, select Vue.js as the front-end framework to build the front-end web interface. Design an intuitive user interface, including input forms, submit buttons, and result display areas. Implement client-side form validation to ensure the integrity and validity of input data. Use Axios or Fetch API to send HTTP requests to the backend server, transmit the user input parameters, and receive the prediction results.

[0087] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0088] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation, characterized in that: The steps include: S1. Based on the preset data extraction rules, the historical clinical information of the target patients and the time of death after the coexistence of the disease are obtained from the online hospital comprehensive management system. The target patients are patients with coexistence of acute myocardial infarction and atrial fibrillation; S2. Preprocessing the historical clinical information of patients, classifying the preprocessed historical clinical information of patients into basic patient information, laboratory test information, AMI and AF comorbidity information, and clinical treatment measure information; S3. Use the LASSO method and the ten-fold cross-validation method to identify the associated clinical characteristics related to the time of death after the coexistence of the disease from the patient's basic information, laboratory test information, AMI and AF comorbidity information, and clinical treatment measures information. The associated clinical characteristics include risk factors and protective factors. The risk factors are pulse, serum creatinine, and neutrophils. The protective factors are hemoglobin, ACEI / ARB, aspirin, and reperfusion therapy. S4. Divide the preprocessed historical patient clinical information into a training set and a validation set; S5. Use the training set to train and optimize four machine learning models, namely, the logistic regression model, the naive Bayes model, the AdaBoost model, and the Glmnet model. Use the validation set to evaluate the prediction performance of the four machine learning models and determine the Glmnet model with the best prediction performance. S6. The SHAP value was used to interpret the Glmnet model and determine the top three optimal predictors that contributed to the model prediction among the associated clinical characteristics. The optimal predictors were pulse, hemoglobin, and neutrophils.

2. The method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 1, characterized in that: The data extraction rules in step S1 include: upper and lower limits of data extraction quantity; coexistence of acute myocardial infarction and atrial fibrillation; the proportion of data with death time exceeding one year after the coexistence of the symptoms; and the proportion of data with death time within one year after the coexistence of the symptoms.

3. The method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 2, characterized in that: In step S1, the online hospital comprehensive management system integrates the HIS system, LIS system and EMR system.

4. The method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 3, characterized in that: In step S2, the basic information of the patient includes age, gender, heart rate, pulse, systolic blood pressure, diastolic blood pressure and infarction type; the laboratory test information includes white blood cells, lymphocytes, monocytes, hemoglobin, platelets, neutrophils, ALT, AST, Glu, TG, TC, LDL-C, HDL-C, blood creatinine, plasma protein and globulin; the comorbidity information of AMI and AF includes hypertension, type 2 diabetes, hyperlipidemia, renal dysfunction, stroke, paroxysmal atrial fibrillation and left atrial enlargement; the clinical treatment measures information includes reperfusion therapy, aspirin, clopidogrel, statins, ACEI / ARB, β-receptor blockers and diuretics.

5. The method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 4, characterized in that: In step S4, 70% of the preprocessed historical patient clinical information is divided into a training set, and 30% of the preprocessed historical patient clinical information is divided into a validation set.

6. The method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 5, characterized in that: In step S5, when the training set is used for training optimization, the logistic regression C parameter of the logistic regression model is 2.07 and the random parameter is 42; the variable smoothing parameter of the Gaussian naive Bayes classifier in the naive Bayes model is 3.85×10 -10 ; The learning rate of the AdaBoost classifier in the AdaBoost model is 0.22, the number of weak evaluators is 53, and the random parameter is 42; the alpha value of the cross-validation classifier in the Glmnet model is 98.69, and the random parameter is 42.

7. The method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 6, characterized in that: In step S5, the prediction performance is evaluated by the Brier score, AUC value and 95% confidence interval of the AUC value, calibration curve and decision curve.

8. The method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 7, characterized in that: The prediction contributions of the optimal prediction factors in step S6 are pulse, hemoglobin, and neutrophils from large to small.

9. A prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation constructed according to the method for constructing a prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to any one of claims 1 to 8.

10. An application of the prognostic risk prediction model based on the coexistence of acute myocardial infarction and atrial fibrillation according to claim 9 in evaluating the prognostic risk prediction of patients with coexistence of acute myocardial infarction and atrial fibrillation.

Images