Heart implantation equipment patient survival rate prediction method based on multi-layer sensor model, electronic equipment and program product

Abstract

The invention discloses a method for predicting the survival rate of a patient with heart implantation equipment, which is used for predicting the survival rate of the patient with the heart implantation equipment through a trained heart implantation prognosis model. The cardiac implantation prognosis model construction method comprises the following steps: obtaining clinical data of a patient with cardiac implantation equipment, and preprocessing the clinical equipment; selecting prediction features for predicting the survival rate of the patient with the heart implantation equipment from the existing clinical features; taking the patient data with the prediction features as a training sample set to train a prognosis model based on a multi-layer perceptron model so as to obtain the prognosis model; the prognosis model is optimized by adopting 5-fold cross validation.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical data processing, and in particular relates to a method for predicting the survival rate of patients with cardiac implant devices based on a multi-layer perceptron model, an electronic device and a program product. Background Art

[0002] Implantable Cardioverter Defibrillator (ICD) and Cardiac Resynchronization Therapy Defibrillator (CRT-D) are widely used to prevent cardiac arrest in patients with heart failure and arrhythmia. Although these devices effectively reduce the risk of sudden cardiac death, the survival rate of patients after implantation is affected by multiple factors. Traditional statistical prediction methods usually use a single variable or linear combination of patients for risk assessment, which has the problem of insufficient ability to integrate complex data, especially showing great limitations when dealing with multidimensional and nonlinear data. Summary of the invention

[0003] The disclosed embodiment discloses a method for predicting the survival rate of patients with cardiac implant devices based on a multi-layer perceptron (MLP) model, which is applicable to patients with implantable cardioverter defibrillators or cardiac resynchronization therapy defibrillators. The survival status prediction is achieved by integrating multi-dimensional physiological data and using a machine learning model, and the influence of key features is analyzed using SHAP values. The model helps doctors formulate personalized treatment plans through high and low risk stratification, and has clinical application value.

[0004] In the field of cardiovascular disease prevention, the present disclosure proposes a prediction model for the prognosis of cardiac device implantation based on machine learning, which significantly improves the ability to predict the long-term survival rate of patients with cardiac implantable devices (such as implantable cardioverter defibrillators (ICDs) and cardiac resynchronization therapy defibrillators (CRT-Ds)). BRIEF DESCRIPTION OF THE DRAWINGS

[0005] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:

[0006] Figure 1 Schematic diagram of the data preprocessing and machine learning model development process according to one embodiment of the present invention.

[0007] Figure 2 A schematic diagram of predictive model performance evaluation according to one embodiment of the present invention.

[0008] Figure 3Schematic diagram of model-based Youden index risk stratification and Kaplan-Meier survival analysis according to one embodiment of the present invention.

[0009] Figure 4 A schematic diagram of an explanatory analysis of a model according to one embodiment of the present invention. DETAILED DESCRIPTION

[0010] Cardiac implants (ICD or CRT-D) are widely used to prevent cardiac arrest. However, the survival rate and prognosis of different patients after device implantation vary greatly, and accurate prediction tools are needed to better guide personalized treatment plans. In recent years, machine learning methods have gradually been applied to prognosis prediction in the medical field, especially in terms of survival rate and risk assessment. By processing large-scale, high-dimensional medical data, machine learning models can capture the complex relationships between variables and effectively improve prediction accuracy. However, there are few existing studies on the prediction of survival rate of patients with ICD or CRT-D implants, and most of them are limited to traditional statistical models, which make it difficult to fully integrate the various physiological and clinical characteristics of patients.

[0011] According to one or more embodiments, a method for predicting survival of a patient with a cardiac implant device based on a multi-layer perceptron (MLP) model comprises the following steps:

[0012] S201, Data Collection:

[0013] Multidimensional data is collected from the patient's clinical information, including cardiac function parameters (such as left ventricular ejection fraction LVEF) and other physiological indicators (such as glomerular filtration rate GFR, uric acid UA, etc.), to ensure that the data dimensions are broad and the information is sufficient.

[0014] S202, data preprocessing:

[0015] The data were filled with missing values, normalized and encoded. Multiple imputation algorithms were used for missing value filling, Min-Max or Z-score methods were used for normalization, and one-hot encoding was performed on categorical variables to improve the training quality of the model.

[0016] S203, feature selection:

[0017] The key prediction features are selected based on the BorutaShap algorithm to ensure that the MLP model focuses on the variables that have the greatest impact on survival prediction, thereby improving the accuracy and stability of the model.

[0018] S204, model building and training:

[0019] An MLP model with at least three hidden layers is constructed. The number of hidden layer nodes and activation function of the model are optimized through cross-validation to ensure the best prediction performance, and regularization techniques such as Dropout are used to prevent overfitting.

[0020] S205, Survival prediction and risk stratification:

[0021] The trained MLP model was used to predict patient survival and to divide patients into high-risk and low-risk groups based on the threshold determined by the Youden index to guide the formulation of personalized treatment plans.

[0022] S206, Model Interpretation:

[0023] SHapley Additive exPlanations (SHAP) values ​​were used to analyze the impact of each feature on the model prediction results and provide explanatory results to facilitate clinical understanding and application.

[0024] The number of hidden layers in the MLP model is three or more, each layer contains at least 50 neuron nodes, and ReLU is used as the activation function to enhance the model's nonlinear expression ability and ability to capture complex relationships.

[0025] The model uses 5-fold cross validation to optimize hyperparameters and determines the optimal learning rate, batch size, and training rounds through grid search to balance the generalization ability and computational efficiency of the model.

[0026] The key features selected in the feature selection step of the model include seven variables: left ventricular ejection fraction (LVEF), glomerular filtration rate (GFR), N-terminal pro-brain natriuretic peptide (pro-BNP), uric acid (UA), left atrial diameter (LAD), serum creatinine (CR), and lymphocyte count (LYMPH) to achieve the optimal survival prediction effect.

[0027] Kaplan-Meier survival curve analysis was used to stratify patients' risks, and the optimal threshold determined by the Youden index was used to distinguish high-risk from low-risk patients, thereby obtaining a prediction of the patients' survival rate within three years.

[0028] According to one or more embodiments, a cardiac implant patient survival prediction system based on a multi-layer perceptron (MLP) model comprises:

[0029] Data collection module: used to receive and store multi-dimensional clinical data of patients, and supports automatic data import and real-time update;

[0030] Data preprocessing module: performs data missing filling, normalization, and one-hot encoding to ensure data consistency and model input quality;

[0031] Feature selection module: Automatically screen important variables based on the BorutaShap method to reduce the model calculation burden and improve the prediction effect;

[0032] Model training module: used to train and optimize the MLP model, including hyperparameter adjustment, model training and model validation;

[0033] Survival prediction module: Calculates the survival probability of patients based on the trained model and performs risk stratification to generate high-risk and low-risk patient groups;

[0034] Result interpretation module: interprets the impact of each feature through SHAP value and outputs the contribution of each feature to the prediction result to facilitate clinicians' understanding and assessment of patient risks.

[0035] Furthermore, the survival prediction module of the system includes a Kaplan-Meier survival curve generation tool, which can draw a survival curve based on the results of risk stratification and intuitively display the survival trends of high- and low-risk groups to facilitate clinical comparative analysis.

[0036] The result interpretation module of the system further supports the visualization of SHAP values ​​and generates feature impact charts to facilitate intuitive understanding of the positive and negative impact of each variable on the survival prediction results.

[0037] According to one or more embodiments, an innovative survival prediction method based on a multi-layer perceptron (MLP) model is specifically designed to predict the long-term survival rate of patients with cardiac implantable devices (such as ICD or CRT-D). The model is designed to predict the three-year survival rate with higher accuracy by integrating the patient's multi-dimensional physiological indicators and cardiac function data, thereby providing data support for doctors, assisting decision-making, and optimizing personalized treatment plans. The method has the following steps:

[0038] S301, data collection.

[0039] Collect and summarize multidimensional data sets containing patients' cardiac function data and other physiological parameters. The data include key variables such as patients' left ventricular ejection fraction (LVEF), glomerular filtration rate (GFR), N-terminal pro-brain natriuretic peptide (pro-BNP), uric acid (UA), left atrial diameter (LAD), serum creatinine (CR), lymphocyte count (LYMPH), etc. This step not only covers the direct physiological indicators of the patient's heart, but also focuses on non-cardiac factors that may affect survival rate to ensure the comprehensiveness of the data.

[0040] S302, data preprocessing.

[0041] In order to ensure the integrity and high quality of medical data and thus improve the prediction accuracy of the model, the acquired data needs to be preprocessed. Specifically:

[0042] S3021, filling in missing data.

[0043] In data preprocessing, multiple imputation methods were used to handle missing data. Missing values ​​in medical data may lead to information loss, affecting the learning and prediction performance of the model. Multiple imputation methods (such as the chain equation method MICE) generate multiple reasonable imputation values ​​based on other available information through multiple imputation to reduce the bias of missing data on the analysis results. For medical data, especially key physiological indicators, reasonable imputation of missing values ​​can maintain the integrity of the patient's physiological information, enable the model to understand the patient's condition more comprehensively, and avoid bias caused by missing data.

[0044] S3022, data normalization.

[0045] The numerical ranges of different medical indicators vary greatly, for example, serum creatinine (CR) has different ranges from left ventricular ejection fraction (LVEF) or uric acid (UA). Normalization is used to adjust different indicators to similar ranges to avoid the model giving higher weights to certain features because their values ​​are too large. Through normalization, the model can treat each feature more fairly during the training process, ensure that the relative relationship between the data is preserved, and help identify important features related to survival rate.

[0046] S3023, one-hot encoding.

[0047] Medical data often contains categorical variables (such as gender, underlying disease type, etc.), which need to be converted into numerical format through one-hot encoding, that is, One-Hot encoding, so that the model can process and analyze. By converting categorical variables into binary features, one-hot encoding avoids bias in numerical differences, allowing the model to more accurately capture the potential impact of different categories of variables on patient survival. For medical data, this step ensures that the model can effectively utilize various types of patient information to achieve more accurate risk assessment.

[0048] S303, feature selection.

[0049] S3031, using BorutaShap feature selection method.

[0050] BorutaShap combines the advantages of the Boruta algorithm and SHAP (Shapley Additive Explanations) values. The Boruta algorithm is a wrapper feature selection method based on random forests, which can stably filter out the most important features when there are many noise features. The SHAP value provides a high degree of interpretability by calculating the contribution of each feature to the model output. There are often a large number of nonlinear relationships and interactions in medical data. The BorutaShap method can extract key physiological parameters from a large number of variables, ensuring that the model is more concise and efficient when it is built. The Boruta algorithm is a feature selection algorithm based on the random forest algorithm, which is used to identify the truly important set of variables in the data set.

[0051] S3032, selection of key features.

[0052] Through the BorutaShap method, seven key features most relevant to patient survival were finally screened out from numerous candidate variables, including:

[0053] Left ventricular ejection fraction (LVEF): LVEF is an important indicator of cardiac systolic function. Low LVEF usually indicates an increased risk of heart failure, and patients may face a higher risk of death. Incorporating LVEF into the model helps reflect the effectiveness of cardiac systolic function and provides important cardiac function information for survival prediction. Glomerular filtration rate (GFR): GFR represents the filtration function of the kidneys. Low GFR indicates renal insufficiency, which is closely related to the risk of cardiovascular disease. The health of the kidneys will directly affect the patient's prognosis, so GFR as a predictive indicator helps the model more comprehensively consider the impact of heart-kidney interaction on survival.

[0054] N-terminal pro-brain natriuretic peptide (pro-BNP): pro-BNP is a biomarker for myocardial stress and heart failure. Elevated pro-BNP levels often indicate that the heart is under greater stress and are an important prognostic factor for patients with heart failure. Using it as a key feature helps the model identify patients with decompensated cardiac function.

[0055] Uric acid (UA): Uric acid is associated with a variety of diseases such as metabolic syndrome, heart failure and hypertension. High uric acid levels increase the risk of cardiovascular events. Incorporating UA into survival prediction helps the model identify the potential impact of metabolic abnormalities on patients' long-term survival.

[0056] Left atrial diameter (LAD): LAD reflects changes in cardiac structure and is associated with an increased risk of atrial fibrillation and heart failure. A larger left atrial diameter usually indicates that the heart has a volume burden, which may aggravate the progression of heart failure. Therefore, LAD is used as a cardiac structural indicator in the model to provide an additional assessment of the cardiac stress state.

[0057] Serum creatinine (CR): CR is an indicator for assessing renal function. Elevated creatinine levels indicate renal impairment. Given the presence of cardiorenal syndrome, changes in renal function will directly affect heart health, so the introduction of CR helps the model reflect the impact of renal health on survival.

[0058] Lymphocyte count (LYMPH): LYMPH reflects the state of immune function. Low lymphocyte count is common in chronic diseases.

[0059] The disease state may indicate that the patient is in an immunocompromised state. Immune system abnormalities are closely related to the occurrence and deterioration of cardiovascular diseases, so using LYMPH as a predictor variable helps to capture survival risk factors more comprehensively.

[0060] S304, model training and optimization

[0061] In the selection of the survival prediction model in the embodiment of the present disclosure, a variety of machine learning models were evaluated, including logistic regression (LR), random forest (RF), extreme gradient boosting (XGB), support vector machine (SVM) and K nearest neighbor (KNN). By comparing the performance of each model in terms of indicators such as the area under the ROC curve (AUC) and the precision-recall curve (PRC), it was found that the MLP model performed well in both the internal and external test sets. On the internal test set, the AUC of the MLP model reached 0.70 and the PRC was 0.34, while on the external test set, the AUC increased to 0.72 and the PRC reached 0.41. These results show that MLP has a significant predictive advantage in identifying the survival rate of ICD and CRT-D patients.

[0062] From a medical perspective, LR is suitable for processing variables with significant linear relationships, but it is usually difficult to capture nonlinear relationships in complex medical data, which limits its predictive accuracy. In contrast, RF has advantages in processing certain nonlinear relationships and data imbalance, but the computational cost is high on high-dimensional data, and it is difficult to explain the complex interactions between features. XGB performs well in many machine learning tasks, but is sensitive to parameter settings when processing high-dimensional medical data and is prone to overfitting. SVM performs well on small and medium-sized data sets, but is less efficient when there are more features and lacks interpretability in clinical applications. The KNN model is based on distance calculation and is suitable for low-dimensional data, but has high computational complexity in high-dimensional medical data and is difficult to capture nonlinear features.

[0063] As a feedforward neural network, the MLP model has a strong ability to capture nonlinearities and is particularly suitable for medical data containing multidimensional features and complex interactions. Through a multi-layer neural network structure, the MLP model can capture deep nonlinear relationships and excel in the correlation between complex variables such as cardiac function, renal function, and biomarkers. Medical data usually contains a large number of features. Through the connection and weighted processing of neurons, MLP can automatically identify important features and effectively deal with high-dimensional data. At the same time, the model's adaptability to complex data is further enhanced through activation functions (such as ReLU or tanh), thereby optimizing survival prediction.

[0064] In order to improve the accuracy and stability of the model, the present invention uses cross-validation and grid search to optimize the hyperparameters of MLP. Cross-validation can ensure the generalization ability of the model on different data subsets, and grid search explores the combination of hyperparameters such as learning rate, batch size, and number of hidden layer neurons, thereby reducing the risk of overfitting while ensuring the stable performance of the model on different data sets. In addition, MLP introduces regularization techniques such as Dropout during the training process, randomly ignoring the connections of some neurons to reduce overfitting and ensure the consistency and reliability of the model in different hospitals and patient groups.

[0065] S305, Risk stratification and survival prediction.

[0066] In the disclosed embodiment, the MLP model is used to predict the three-year survival probability of patients with ICD or CRT-D implants, and the optimal threshold is determined by the Youden index to divide the patients into high-risk and low-risk groups. Compared with other models, MLP performs well in accuracy and sensitivity, making this risk stratification method more practical in medical applications. The nonlinear capture ability of the MLP model enables it to identify the deep relationship between complex physiological and clinical variables, thereby achieving accurate prediction of patient survival probability.

[0067] The stratification effect of the MLP model was further verified by using the Kaplan-Meier survival curve. The results showed that the survival rate of patients in the high-risk group was significantly lower than that in the low-risk group (P<0.001). The statistical significance of this difference verified the effectiveness and reliability of MLP in distinguishing different risk groups. The MLP model can effectively process multidimensional, nonlinear biomarker data, especially cardiac function, renal function, and inflammation-related indicators. These indicators jointly affect the survival rate of patients under complex interactions, and the multi-layer structure of the MLP model can integrate the association characteristics of these variables into risk stratification.

[0068] Compared with traditional models, MLP can model nonlinear and interactive effects between features, which is particularly important in risk stratification. By more accurately predicting survival rates and performing risk stratification, the MLP model supports medical personnel in timely early intervention or intensive treatment in high-risk patients, helps optimize the allocation of medical resources, and improves the pertinence and effectiveness of treatment, making individualized treatment plans more accurate. S306, model interpretation and clinical application.

[0069] In the disclosed embodiment, in order to make the prediction results of the MLP model more interpretable and clinically practical, the SHapley Additive exPlanations (SHAP) value is used to interpret the model. The SHAP value can quantify the contribution of each feature to the patient's survival prediction, so that clinicians can more clearly understand the impact of each key physiological parameter on the survival rate, thereby improving the transparency and practicality of the model.

[0070] When the MLP model processes medical data, the SHAP value analysis can reveal the degree of impact of key features on survival. For example, features such as left ventricular ejection fraction LVEF, GFR, pro-BNP, etc. are all important physiological indicators that affect survival. The SHAP value assigns positive and negative values ​​to each feature: a positive value indicates that the feature has a positive effect on survival, while a negative value indicates the possibility of increased risk. For example, higher LVEF and GFR generally contribute to improved survival, while an increase in pro-BNP may increase the risk of survival. With this information, clinicians can more clearly identify the specific risk factors for each patient and provide data support for individualized treatment.

[0071] The interpretability of the MLP model is crucial for clinical decision-making. The SHAP value not only reveals the relationship between survival rate and key physiological parameters, but also provides a specific risk factor assessment for each patient. This feature enables the MLP model to transcend the limitations of traditional "black box" neural networks and provides an intuitive and actionable interpretation for clinicians to understand and apply. For example, based on the results of the SHAP analysis, doctors can pay more attention to high-risk features, intervene or adjust treatment plans in a timely manner, thereby improving the patient's prognosis.

[0072] By integrating the interpretability of SHAP values, the MLP model can not only provide survival predictions, but also help clinicians understand the individual risk characteristics of patients more comprehensively. This individualized risk assessment based on explanatory analysis not only improves the credibility of the model, but also provides strong data support for clinical decision-making, helping to achieve precision medicine.

[0073] like Figure 1As shown in the figure, the processing of research data and the development process of machine learning models are demonstrated. The data were collected from 3175 patients with ICD or CRT-D implantation in four large hospitals. The original data contained 86 features. After missing value screening and feature selection, 44 features were finally determined and 7 key predictive features were further selected for model development. These features include left ventricular ejection fraction (LVEF), glomerular filtration rate (GFR), N-terminal pro-brain natriuretic peptide (pro-BNP), uric acid (UA), left atrial diameter (LAD), serum creatinine (CR) and lymphocyte count (LYMPH), which have the most significant effect on the prediction of survival rate. During the model development process, multiple machine learning algorithms (such as RF, LR, MLP, etc.) were optimized through 5-fold cross validation to find the optimal model. In the end, the multilayer perceptron (MLP) model showed excellent performance in predicting three-year survival rate.

[0074] Figure 1 The English terms and their Chinese explanations are as follows.

[0075] 1. Zhongshan Hospital

[0076] 2. Fuwa Hospital

[0077] 3. Shanghai Chest Hospital

[0078] 4.Changhai Hospital

[0079] 5. Patients excluded for less than 3-year follow up after implantation

[0080] 6. Derivation cohort

[0081] 7. External test cohort

[0082] 8.Data preprocessing and random allocation

[0083] 9. Training cohort

[0084] 10. Internal test cohort

[0085] 11.Clinical features from database

[0086] 12.Features with missing values>40%——Features with more than 40% missing values

[0087] 13. Using custom bootstrap method to select 14. Machine-learning Algorithm

[0088] 15.5-fold cross validation

[0089] 16. Predictive models established

[0090] 17.Internal test

[0091] 18. External test

[0092] 19.removed——removed

[0093] Figure 1 The prediction model development and validation process based on machine learning algorithm in the present invention includes the following steps: S101, data collection. A certain amount of patient data was collected from Zhongshan Hospital, Fuhua Hospital, Shanghai Chest Hospital and Changhai Hospital.

[0094] S102, exclude some patient data. Patients with less than 3 years of follow-up after implantation were excluded to ensure the integrity and reliability of the data.

[0095] S103, the remaining patient data after data exclusion is set as a derivative cohort, which is used for model development and internal validation; at the same time, an external test cohort is designed for external validation of the model.

[0096] S104, preprocessed the data of the derivative cohort and randomly assigned the data to the training cohort and the internal test cohort; extracted 86 clinical features from the database, removed features with more than 40% missing values, identified 44 features, and further screened out 7 key clinical features through a custom bootstrap method.

[0097] S105, use a variety of machine learning algorithms (such as random forest RF, logistic regression LR, multi-layer perceptron MLP, extreme gradient boosting XGB, support vector machine SVM, K nearest neighbor KNN) to establish a prediction model.

[0098] S106, a five-fold cross-validation method was used to evaluate the performance of the model to ensure the generalization ability of the model.

[0099] S107, based on the cross-validation results, establish the final prediction model.

[0100] S108, the established model is tested on the internal test cohort and the external test cohort to verify the predictive performance of the model.

[0101] In the embodiment of the present disclosure, the best hyperparameters of each model are screened by the grid search method to ensure that the model achieves the best performance on different data sets (Table 2). For the MLP model, the grid search tested different numbers of hidden layer neurons (such as 50, 100), activation functions (such as tanh, relu) and initial learning rates (such as 0.001, 0.01), and finally selected a single layer of 50 neurons, tanh activation function and a learning rate combination of 0.001.

[0102] The LR model tested different regularization penalty coefficients and solvers, and the best settings were penalty coefficient 0.001, L2 regularization, and liblinear solver; the RF model selected a combination of 200 trees from different numbers of decision trees (such as 50, 100, and 200). For the KNN model, the grid search selected a configuration with 9 neighbors and a weight parameter of 'distance', and the SVM model selected a combination with the best value of 0.01 under different regularization parameters (such as 0.001, 0.01, and 0.1). These optimization settings ensure the stable performance of each model on internal and external test sets.

[0103] Table 2 Optimal hyperparameters for the model

[0104]

[0105] The English terms and their Chinese explanations in Table 2 are as follows:

[0106] 1.Classifier models

[0107] 2. Hyperparameters

[0108] 3. The best-tuned hyperparameters

[0109] Table 2 shows the hyperparameter selection process for different classifier models and the optimal hyperparameters finally selected. The following is a description of the models and hyperparameters.

[0110] (1) Logistic Regression:

[0111] C——Regularization parameter, controls the complexity of the model. The smaller the value, the stronger the regularization.

[0112] Penalty——Regularization type, 'l2' means L2 regularization.

[0113] Solver - Optimization algorithm used to solve the parameters of the logistic regression model.

[0114] (2) Random Forests:

[0115] n_estimators – The number of decision trees in the forest. A larger number usually results in a more stable and accurate model.

[0116] (3) Multilayer Perceptron and eXtreme Gradient Boosting:

[0117] Hidden layer sizes: The number of neurons in the hidden layer, which affects the complexity and learning ability of the model.

[0118] Activation——Activation function, 'tanh' is the hyperbolic tangent function, which is used to introduce nonlinearity.

[0119] learning rate init——initial learning rate, which affects the convergence speed and stability of model training.

[0120] learning rate——Learning rate controls the step size of model updates during training.

[0121] (4) Support Vector Machine:

[0122] C——Regularization parameter, controls the complexity of the model. The smaller the value, the stronger the regularization.

[0123] (5) K-Nearest Neighbors:

[0124] n_neighbors – The number of neighbors, i.e. the number of nearest neighbors to consider while doing classification.

[0125] Weights——weight function, 'distance' means weighting according to distance, the closer the neighbor, the greater the weight.

[0126] The "Best Tuned Hyperparameters" column in Table 2 shows the optimal parameter values ​​determined during the hyperparameter tuning process, such as through cross-validation. These parameter values ​​enable the model to achieve the best performance on a given dataset. For example, the logistic regression model performs best when C=0.001, using 'l2' regularization and 'liblinear' solver; while the random forest model performs best when there are 200 decision trees.

[0127] In the disclosed embodiments, the model performance was evaluated. Figure 2 Table 1 shows the performance evaluation of the multi-layer perceptron (MLP) model on the internal and external test sets, further illustrating the predictive stability and effectiveness of the model. Figure 2 The detailed performance indicators in the MLP model are AUC (area under the curve) of 0.70 (95% CI: 0.64-0.76) on the internal test set and AUC of 0.72 (95% CI: 0.62-0.82) on the external test set, indicating that the model has good discrimination ability on different data sets. In addition, the AUPRC (area under the precision-recall curve) of MLP on the external test set is 0.41, which is better than other models, especially showing strong generalization ability on complex clinical data.

[0128] Figure 2 The English terms and their Chinese interpretations are as follows:

[0129] 1. True Positive Rate (TPR)

[0130] 2.False Positive Rate (FPR)——False Positive Rate

[0131] 3. ROC Curve——Receiver Operating Characteristic Curve

[0132] 4.AUC——Area Under the Curve

[0133] Figure 2 The classification performance of different machine learning models on two different datasets is shown, evaluated by the receiver operating characteristic curve (ROC curve) and the area under the curve (AUC). Figure 2 A and Figure 2 B, respectively represent two different data sets or test environments. The curves in each figure represent the performance of different machine learning models.

[0134] The ROC curve is a graphical tool used to evaluate the performance of a binary classification model. It shows the classification ability of the model by plotting the true positive rate (TPR) and false positive rate (FPR) at different thresholds. The closer the ROC curve is to the upper left corner, the better the performance of the model. The AUC is the area under the ROC curve and is used to quantify the overall performance of the model. The AUC value ranges from 0 to 1, and the closer the value is to 1, the better the classification performance of the model. An AUC of 0.5 means that the performance of the model is comparable to random guessing.

[0135] Figure 2 A variety of machine learning models are listed in the figure, including K nearest neighbors (knn), logistic regression (lr), multi-layer perceptron (mlp), random forest (rf), support vector machine (svm), and extreme gradient boosting (xgb). The AUC value of each model is shown in the legend to compare their performance.

[0136] exist Figure 2 In A, the logistic regression (lr) model has the highest AUC value (0.71), indicating that it performs best on this dataset. Figure 2 In B, the random forest (rf) model has the highest AUC value (0.67), indicating that it performs best on this dataset.

[0137] In addition to AUC and AUPRC, Table 1 also shows the performance of the MLP model on multiple performance indicators: Internal test set: MLP has an accuracy of 0.70, a sensitivity of 0.67, a specificity of 0.70, and an F1-Score of 0.44, indicating that the model has a good performance in balancing sensitivity and specificity. At the same time, the Brier score is 0.13, reflecting the high calibration of the prediction.

[0138] External test set: In the external test set, the MLP achieved an accuracy of 0.76, a specificity of 0.80, and an F1-Score of 0.45. These data show the robustness of the model on external validation data, especially in the test of real clinical data, which maintained stable predictive performance.

[0139] Compared with other models (such as KNN, SVM and XGB), the MLP model performs well in multiple key indicators, especially in accuracy, AUC and F1-Score. The performance evaluation results of the model verify its effectiveness in survival prediction, and further prove the advantages of the MLP model over traditional statistical models and other machine learning models in nonlinear and high-dimensional data analysis. This superior performance makes the MLP model an ideal choice for three-year survival prediction of ICD and CRT-D patients, providing clinicians with a more reliable decision support tool.

[0140] Table 1 Model performance evaluation;

[0141]

[0142] The English terms and their Chinese explanations in Table 1 are as follows:

[0143] 1.AUROC (Area Under the Receiver Operating Characteristic Curve)

[0144] 2.AUPRC (Area Under the Precision-Recall Curve)

[0145] 3.Sensitivity——Sensitivity (Recall Rate)

[0146] 4. Specificity

[0147] 5.Accuracy

[0148] 6.Precision

[0149] 7. F1-Score

[0150] 8. Brier score

[0151] Table 1 shows the performance indicators of different machine learning models on internal test sets and external test sets.

[0152] AUROC is a measure of the model's ability to distinguish between positive and negative samples. The closer the value is to 1, the better the model performance.

[0153] AUPRC - Particularly suitable for imbalanced datasets, it measures the balance between the precision and recall of the model at different thresholds.

[0154] Sensitivity (Recall Rate) – The ability of the model to correctly identify positive samples, that is, the true rate.

[0155] Specificity — The ability of the model to correctly identify negative samples, i.e., the true negative rate.

[0156] Accuracy — The proportion of model predictions that are correct.

[0157] Precision — The proportion of samples that are actually positive among all samples predicted to be positive by the model.

[0158] F1-Score - the harmonic average of precision and recall, used to measure the balance between precision and recall of the model.

[0159] Brier score – a measure of the accuracy of a probabilistic forecast, with values ​​closer to 0 indicating more accurate forecasts.

[0160] From Table 1, we can see that

[0161] In the internal test set, the logistic regression (LR) model has the highest AUROC and AUPRC values, indicating that it performs best in distinguishing positive and negative samples.

[0162] In the external test set, the Multilayer Perceptron (MLP) model has the highest AUROC value, but on AUPRC, the KNN model performs best.

[0163] The performance of the support vector machine (SVM) on both test sets is relatively poor, especially in terms of recall and precision.

[0164] The performance of Extreme Gradient Boosting (XGB) is relatively stable in both test sets, but not optimal.

[0165] The Brill scores are relatively low in all models, indicating that there is some discrepancy between the predicted probability distribution and the actual outcomes.

[0166] The disclosed embodiment performed risk stratification and Kaplan-Meier survival analysis based on the Youden index. In the disclosed embodiment, the Youden index was used to determine the optimal risk stratification threshold based on the prediction results of the multilayer perceptron (MLP) model to divide ICD or CRT-D patients into high-risk and low-risk groups. The Youden index is used to determine the optimal threshold, which can achieve the best balance between sensitivity and specificity, thereby improving the classification performance of the model and making the prediction results more accurate in distinguishing patients with different risk levels. After determining the risk stratification, the stratification effect was further verified by Kaplan-Meier survival analysis. The Kaplan-Meier curve showed that the survival rate of patients in the high-risk group during the three-year follow-up was significantly lower than that of the low-risk group (P<0.001), indicating that the MLP model can effectively identify patient groups with higher survival risks. This result not only illustrates the effectiveness of the model in distinguishing patients with different survival risks, but also reflects its value in clinical applications.

[0167] Figure 3 The Chinese and English terms and their Chinese explanations are as follows:

[0168] 1. Kaplan-Meier Survival Curves

[0169] 2. Survival Probability

[0170] 3. Follow-up (Days)

[0171] 4. Number at risk

[0172] 5.predict label——predict label

[0173] 6. High

[0174] 7. Low

[0175] Figure 3 The Kaplan-Meier survival curve based on the predicted signature is used to compare the changes in the survival probability of two groups with different risk levels (high risk and low risk) over time.

[0176] Kaplan-Meier survival curve is a nonparametric statistical method for estimating the survival function, commonly used in survival analysis. It shows the distribution of survival data by plotting the change of survival probability over time.

[0177] Survival probability - represents the probability that the subject is still alive at a specific point in time.

[0178] Follow-up time – refers to the length of time from the start of the study to the occurrence of an event (such as death) or the end of the study.

[0179] Number of people at risk – the number of individuals who remain at risk at each point in time. Over time, as events occur, the number of people at risk decreases.

[0180] Predictive labeling - based on a certain predictive model, the research subjects are divided into two groups: high risk and low risk.

[0181] Among them, the blue curve represents the high-risk group and the orange curve represents the low-risk group. It can be seen that the survival probability of the high-risk group decreases faster than that of the low-risk group over time, indicating that the high-risk group has a shorter survival time. The shaded part of the curve represents the confidence interval of the survival probability, reflecting the uncertainty of the estimate. Figure 3 Through the method of survival analysis, the changing trend of survival probability of different risk groups over time is demonstrated, which is helpful to evaluate the effectiveness of the prediction model and the impact of risk stratification on survival prognosis.

[0182] The disclosed embodiments provide for model explanatory analysis, such as Figure 4The model interpretability analysis based on SHapley Additive ex Planations (SHAP) value (Shapley additive interpretation) was demonstrated to quantify the impact of each feature on survival prediction. The results showed that eGFR and pro-BNP ranked at the top of the feature importance, with significant positive and negative effects on patient survival, respectively. This interpretable analysis not only improves the transparency of the model, but also provides important reference indicators for clinicians, allowing them to better understand the model's prediction results and assist in individualized decision-making.

[0183] Figure 4 The English terms and their Chinese interpretations are as follows:

[0184] 1. eGFR——estimated Glomerular Filtration Rate

[0185] 2. pro-BNP——pro-B-type Natriuretic Peptide

[0186] 3.UA——Uric Acid

[0187] 4.LAD——Left Anterior Descending artery

[0188] 5.CR——Creatinine

[0189] 6.LYMPH - Lymphocytes

[0190] 7. LVEF——Left Ventricular Ejection Fraction

[0191] Figure 4 Demonstrates the use of SHAP (SHapley Additive exPlanations) values ​​to explain machine learning model predictions. Figure 4 A shows the average impact of different features on the model output (average of SHAP values). The larger the SHAP value of each feature, the greater the impact of the feature on the model prediction results. Figure 4 As can be seen in A, eGFR and pro-BNP are the two features that have the greatest impact on the model output. Figure 4 B shows the SHAP value distribution of each feature. The positive or negative SHAP value indicates the direction of the feature's impact on the model output: a positive value indicates that the feature increases the predicted value, and a negative value indicates that the feature decreases the predicted value. Figure 4Each point in B represents the SHAP value of a sample, and the color indicates the high or low feature value. Figure 4 C shows the prediction results and feature contributions for a specific sample. Figure 4 In C, the SHAP value of each feature shows the specific contribution of the feature to the prediction result. For example, for a specific sample, a high value of LAD (1.823) has a positive contribution to the prediction result, while a low value of LVEF (-1.079) has a negative contribution to the prediction result. Figure 4 D is similar to Figure 4 C, but shows the feature contribution of another sample. In this sample, the high value of UA (0.881) contributes positively to the prediction result, while the low value of LYMPH (-0.9728) contributes negatively to the prediction result.

[0192] Figure 4 SHAP values ​​are used to explain the prediction results of machine learning models, helping us understand how each feature affects the output of the model and how the high and low feature values ​​affect the prediction results.

[0193] In summary, the beneficial effect of the present disclosure is that, based on the multi-layer perceptron (MLP) model, seven key features such as glomerular filtration rate (GFR), N-terminal pro-brain natriuretic peptide (pro-BNP), left ventricular ejection fraction (LVEF), and uric acid (UA) are selected, and through data integration and the capture of nonlinear relationships, efficient and personalized three-year survival prediction is provided. By integrating machine learning models, especially multi-layer perceptron (MLP) models, complex patient data can be deeply mined to provide highly targeted survival predictions, thereby achieving more scientific risk stratification and decision support, and improving patients' long-term health management.

[0194] It should be understood that in the embodiments of the present invention, the term "and / or" is only a description of the association relationship of the associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0195] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0196] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0197] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for predicting the survival rate of patients with cardiac implant devices, characterized in that: Predicting survival of patients with cardiac implantable devices using a trained cardiac implant prognostic model.

2. The method according to claim 1, characterized in that The method for constructing a cardiac implantation prognosis model comprises the following steps: A101, obtaining clinical data of the patient with the cardiac implant device and preprocessing the clinical device; A102, Selecting predictive features from existing clinical features for predicting survival in patients with cardiac implantable devices; A103, using the patient data having the predicted features as a training sample set to train a prognosis model based on a multilayer perceptron model, thereby obtaining a prognosis model.

3. The method according to claim 2, characterized in that The method for constructing a cardiac implantation prognosis model further comprises the steps of: A104, the prognostic model was optimized using 5-fold cross validation.

4. The method according to claim 2, characterized in that: The trained prognostic model was used to predict the survival rate of patients, and the patients were divided into high-risk and low-risk groups based on the threshold determined by the Youden index, thereby obtaining a prediction of the patient's survival rate within three years.

5. The method according to claim 4, characterized in that Kaplan-Meier survival curve analysis was used to stratify the patients' risk, and high-risk and low-risk patients were grouped according to the optimal threshold determined by the Youden index.

6. The method according to claim 2, characterized in that The SHAP value was used to analyze the influence of each prediction feature on the prediction results of the prognostic model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The processor runs the computer program to implement the method according to any one of claims 1 to 6.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer program product, comprising a computer program, characterized in that The computer program is executed by a processor to implement the method according to any one of claims 1 to 6.

Images