Implantable cardiac defibrillator patient dynamic death risk prediction method and system

Abstract

The invention discloses a dynamic death risk prediction method and system for an implantable cardiac defibrillator patient. The death risk prediction method comprises the following steps: constructing an implantable cardiac defibrillator patient survival prediction model based on a random survival forest model; for an implantable cardiac defibrillator patient, baseline data and a survival state of the patient are collected; carrying out data preprocessing on the collected baseline data and survival state of the patient to obtain preprocessed baseline data of the patient; feature selection is carried out on the preprocessed patient baseline data, and survival prediction high-correlation features are screened out; the patient survival prediction model of the implantable cardiac defibrillator is trained and optimized; and carrying out survival prediction on the implantable cardiac defibrillator patient by adopting the trained and optimized survival prediction model for the implantable cardiac defibrillator patient. According to the death risk prediction method and system, the survival probability of the implantable cardiac defibrillator patient in any day within ten years can be predicted.

Description

Technical Field

[0001] The present invention relates to a method for predicting the death risk of a patient, and in particular to a method and system for predicting the dynamic death risk of an implantable cardioverter defibrillator patient. Background Art

[0002] For an implantable cardioverter defibrillator patient, the patient's long-term survival probability is affected by multiple factors, including comorbidities, individual physiological characteristics, and heterogeneity of disease progression. Therefore, accurate risk stratification is crucial for optimizing personalized treatment strategies.

[0003] Traditional survival analysis methods, such as the Cox proportional hazards model, often use linear assumptions for risk assessment, which makes it difficult to effectively capture complex nonlinear relationships and interactions in high-dimensional data.

[0004] Some existing prediction models are mainly used to assess the risk of ventricular arrhythmias in patients with ICD / CRT-D implants, or to predict mortality based on competing risk models, but existing models only provide static risk estimates.

[0005] To sum up, the current problems are:

[0006] When it comes to predicting the survival probability of patients with implantable cardioverter defibrillators, existing prediction models are mostly based on estimates at static time points, which makes it difficult to reflect the dynamic changes in patients' survival risks over time, especially the inability to "accurately predict the survival probability of patients with implantable cardioverter defibrillators on any day." Summary of the Invention

[0007] The purpose of the present invention is to provide a dynamic death risk prediction method for implantable cardioverter defibrillator patients. The death risk prediction method and system can predict the survival probability of implantable cardioverter defibrillator patients on any day within ten years.

[0008] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0009] A method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators, comprising:

[0010] S1, constructing a survival prediction model for patients with implantable cardioverter defibrillators based on the random survival forest model;

[0011] S2, for patients with implantable cardioverter defibrillators, collect the patient's baseline data and survival status;

[0012] S3, performing data preprocessing on the collected patient baseline information and survival status to obtain preprocessed patient baseline data;

[0013] S4, performing feature selection on the pre-treated patient baseline data to screen out features with high correlation with survival prediction;

[0014] S5, training and optimizing the implantable cardioverter defibrillator patient survival prediction model;

[0015] S6, using the trained and optimized implantable cardioverter defibrillator patient survival prediction model to predict the survival of implantable cardioverter defibrillator patients.

[0016] Furthermore, the highly correlated features for survival prediction include 28 feature data, namely left ventricular end-diastolic diameter, New York Heart Association functional class, uric acid, blood urea nitrogen, N-terminal pro-brain natriuretic peptide, left ventricular ejection fraction, age, serum creatinine, albumin, left atrial diameter, new oral anticoagulants, amiodarone, intraventricular conduction abnormalities, total cholesterol, sodium, atrial fibrillation, diuretics, QRS wave duration, digoxin, hypertension, gender, atrioventricular block, primary or secondary prevention, history of coronary artery intervention, diabetes, history of stroke, use of aldosterone receptor antagonists, and use of warfarin.

[0017] Furthermore, step S4 also includes: using the SHapley Additive Explanations method to further screen out key variables from the highly correlated survival prediction features, and the screened out key variables are used to explain the contribution of the features in the implantable cardioverter defibrillator patient survival prediction model.

[0018] Furthermore, the number of key variables screened out was 20, namely left ventricular end-diastolic diameter, New York Heart Association functional class, uric acid, blood urea nitrogen, N-terminal pro-brain natriuretic peptide, left ventricular ejection fraction, age, serum creatinine, albumin, left atrial diameter, new oral anticoagulants, amiodarone, intraventricular conduction abnormalities, total cholesterol, sodium, atrial fibrillation, diuretics, QRS duration, digoxin and hypertension.

[0019] Furthermore, the data preprocessing includes missing data filling, data normalization and one-hot encoding.

[0020] Furthermore, step S5 further includes:

[0021] During model training, the RSF model uses Bootstrap resampling technology to construct multiple survival decision trees and utilizes the Log-rank splitting criterion for feature partitioning, thereby learning and modeling the patient's long-term survival. The model integrates the prediction results of multiple decision trees to generate an individualized survival function, enabling it to output the survival probability at any given time point and possessing continuous time dynamic prediction capabilities.

[0022] Furthermore, step S5 further includes:

[0023] The optimal hyperparameter combination was determined using a grid search method, including a maximum number of decision trees of 100, a maximum tree depth of 40, and a minimum number of sample splits of 5. The final model output a complete survival probability curve for each patient. The model's predictive performance was evaluated and calibrated using C-index, Brier score, and time-dependent AUC metrics, and its clinical application value was verified using decision curve analysis.

[0024] Furthermore, step S6 further includes:

[0025] The implantable cardioverter defibrillator (ICD) patient survival prediction model was used to predict the five-year survival probability of ICD patients. Based on the five-year survival probability, ICD patients were divided into high-risk group, medium-risk group and low-risk group.

[0026] A dynamic death risk prediction system for patients with implantable cardiac defibrillators is provided, wherein the death risk prediction system is provided with a computer program capable of implementing the above-mentioned death risk prediction method.

[0027] Compared with the prior art, the mortality risk prediction method of the present invention has the following advantages:

[0028] The mortality risk prediction method and system of the present invention constructs a survival prediction model for patients with implantable cardioverter defibrillators (ICDs) based on a random survival forest model (RSF model). This model is trained and optimized using the RSF model. Cox regression is used to identify highly relevant features for survival prediction, including core indicators such as LVEDD and NYHA. The model uses bootstrap resampling technology to construct multiple decision trees, employs the Log-rank splitting criterion for feature partitioning, and integrates the prediction results to generate a personalized survival function. This method can predict the survival probability of an ICD patient on any given day within ten years. This approach overcomes the existing art's limitation of reflecting the dynamic changes in a patient's survival risk over time, achieving excellent results. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of data preprocessing and machine learning model development in the method for dynamic mortality risk prediction of implantable cardioverter-defibrillator patients of the present invention;

[0030] Figures 2 to 4 is a schematic diagram of the model performance evaluation in the mortality risk prediction method of the present invention,

[0031] in,

[0032] Figure 2 This is a schematic diagram comparing the time-dependent AUC performance of different models.

[0033] Figure 3 This is a schematic diagram comparing the time-dependent Brier scores of different models.

[0034] Figure 4 Schematic diagram of decision curve analysis for different models;

[0035] Figure 5 Schematic diagram of risk stratification and Kaplan-Meier survival analysis based on the five-year survival probability prediction value of the implantable cardioverter defibrillator patient survival prediction model in the death risk prediction method of the present invention;

[0036] Figure 6 Schematic diagram of the model explanatory analysis in the method for predicting mortality risk of the present invention;

[0037] Figure 7 A table showing the performance comparison of different models in predicting all-cause mortality;

[0038] Figure 8 Schematic diagram of the hyperparameter settings and tuning results for each survival analysis model. DETAILED DESCRIPTION

[0039] The present invention will be further described below with specific embodiments:

[0040] This embodiment provides a method for predicting the dynamic mortality risk of patients with implantable cardioverter defibrillators.

[0041] The death risk prediction method is based on the random survival forest model (RSF).

[0042] It is specifically designed to predict the long-term survival probability of patients with implantable cardioverter-defibrillators, such as those with an ICD or a cardiac resynchronization therapy defibrillator (CRT-D). By leveraging the non-parametric and ensemble learning properties of the random survival forest (RSF), it can comprehensively consider the patient's multidimensional physiological indicators, cardiac function data, and other key clinical variables to predict long-term survival risk with greater accuracy. Compared with traditional statistical methods, the RSF model not only captures complex nonlinear relationships but also provides personalized and dynamically updated survival probability predictions, significantly improving the accuracy of risk stratification.

[0043] Specifically,

[0044] The death risk prediction method of this embodiment includes steps S1 to S6.

[0045] S1, building the model

[0046] A prediction model is constructed based on the existing random survival forest model (RSF).

[0047] In order to facilitate accurate description, the constructed prediction model is defined as "implantable cardiac defibrillator patient survival prediction model".

[0048] In the implantable cardioverter defibrillator patient survival prediction model, by integrating multiple survival decision trees and utilizing bootstrap sampling and random feature selection, the interactive relationships among complex variables such as cardiac function, renal function, and biomarkers are effectively captured, thereby enabling good personalized survival prediction.

[0049] S2, data collection:

[0050] Data collected from patients with implantable cardioverter-defibrillators (ICDs) includes multidimensional physiological information such as cardiac function, renal function, inflammatory response, and medication use. This data comes from clinical data from patients with implantable cardiac devices (ICDs or CRT-Ds) at multiple medical centers. This clinical data, with continuous follow-up for up to ten years, enhances the model's long-term dynamic predictive capabilities and represents real-world clinical practice, making it applicable to a wider range of risk assessment cycles.

[0051] Data collected for patients with implantable cardioverter-defibrillators also include vital status data from the National Death Registry.

[0052] In this embodiment, the national death registration system is used to confirm the patient's survival status to ensure the accuracy of long-term survival prediction, and the contribution of each variable to survival prediction is analyzed in combination with SHapley Additive Explanations (SHAP) value to optimize the risk stratification strategy.

[0053] For the convenience of accurate expression, all data collected in step S2 are defined as "patient's baseline information and survival status".

[0054] The baseline information and survival status of the patients collected above provide data support for the subsequent construction of a comprehensive survival prediction model for implantable cardioverter defibrillator patients.

[0055] S3, data preprocessing

[0056] The data preprocessing is performed on the baseline information and survival status of the patients collected in step S2 in order to ensure the integrity and high quality of the medical data, thereby improving the prediction accuracy of the implantable cardioverter defibrillator patient survival prediction model.

[0057] The data preprocessing includes:

[0058] S31, filling in missing data

[0059] Specifically, in this embodiment, an iterative filling method is used to handle missing data. Since missing values in medical data may lead to information loss, thereby affecting the learning and prediction performance of the model, a more accurate filling method than simple mean filling or median filling is selected - the iterative filling method (IterativeImputer). This method fills missing values by iterative prediction, which can effectively reduce the deviation of missing data on the analysis results. For medical data, especially key physiological indicators, reasonable filling of missing values can maintain the integrity of the patient's physiological information, enabling the model to more comprehensively understand the patient's condition, thereby avoiding the deviation caused by missing data.

[0060] S32, data normalization

[0061] Specifically, because different medical indicators have widely varying numerical ranges, such as serum urea nitrogen (BUN) and left ventricular ejection fraction (LVEF), normalization is necessary. Normalization adjusts different indicators to similar ranges, preventing the model from assigning higher weights to certain features due to their excessively large values. Through normalization, the model treats each feature more fairly during training, ensuring that relative relationships between data are preserved, thereby helping to identify important features related to survival probability.

[0062] S33, one-hot encoding

[0063] Specifically, medical data often contains categorical variables, such as gender and underlying disease type. These variables need to be converted into numerical format through one-hot encoding (unique hot encoding) so that the model can process and analyze them. One-hot encoding converts categorical variables into binary features, avoiding bias caused by numerical differences, allowing the model to more accurately capture the potential impact of different categorical variables on the patient's survival probability. For medical data, this step ensures that the model can effectively utilize various types of patient information to achieve more accurate risk assessment.

[0064] For the convenience of accurate expression, the data after pre-processing in step S3 is defined as "pre-processed patient baseline data".

[0065] S4, feature selection

[0066] Feature selection is performed on the pre-processed patient baseline data obtained in step S3.

[0067] In this embodiment, a two-step feature screening strategy is adopted to ensure that the survival prediction model for implantable cardioverter defibrillator patients focuses on the variables with the most predictive value and optimizes computational efficiency.

[0068] Specifically,

[0069] First, Cox regression feature selection was used to preliminarily screen out 28 characteristic data variables most relevant to survival prediction (or survival outcomes) from the baseline data of pre-treatment patients, and irrelevant or redundant features were eliminated.

[0070] In order to facilitate accurate description, the features that are most relevant to survival prediction are defined as “highly relevant features for survival prediction”;

[0071] The 28 highly correlated features for survival prediction that were screened out are listed below:

[0072] 1. Left ventricular end-diastolic diameter

[0073] 2. New York Heart Association functional classification

[0074] 3. Uric acid

[0075] 4. Blood urea nitrogen

[0076] 5. N-terminal pro-brain natriuretic peptide

[0077] 6. Left ventricular ejection fraction

[0078] 7. Age

[0079] 8. Serum creatinine

[0080] 9. Albumin

[0081] 10.Left atrial diameter

[0082] 11. New oral anticoagulants

[0083] 12. Amiodarone

[0084] 13. Intraventricular conduction abnormalities

[0085] 14. Total cholesterol

[0086] 15. Sodium

[0087] 16. Atrial fibrillation

[0088] 17. Diuretics

[0089] 18.QRS duration

[0090] 19. Digoxin

[0091] 20. Hypertension

[0092] 21. Gender

[0093] 22. Atrioventricular block

[0094] 23. Primary or secondary prevention

[0095] 24. History of coronary artery intervention

[0096] 25. Diabetes

[0097] 26. History of stroke

[0098] 27. Use of aldosterone receptor antagonists

[0099] 28. Warfarin use

[0100] The above variables cover the patient's cardiac structure and function, biochemical indicators, electrophysiological status, medication use, and medical history, and are the core input variables for this embodiment to achieve dynamic survival probability prediction within ten years.

[0101] Next, we used the SHapley Additive Explanations (SHAP) method to evaluate the importance of the 28 highly relevant features identified previously. We then further screened out a final set of 20 key variables and ranked their importance. These 20 key variables significantly contributed to survival prediction and were used to explain the impact of each feature in the model.

[0102] See also Figure 6 The 20 key variables finally screened out are listed as follows:

[0103] 1. Left ventricular end-diastolic diameter

[0104] 2. New York Heart Association functional classification

[0105] 3. Uric acid

[0106] 4. Blood urea nitrogen

[0107] 5. N-terminal pro-brain natriuretic peptide

[0108] 6. Left ventricular ejection fraction

[0109] 7. Age

[0110] 8. Serum creatinine

[0111] 9. Albumin

[0112] 10.Left atrial diameter

[0113] 11. New oral anticoagulants

[0114] 12. Amiodarone

[0115] 13. Intraventricular conduction abnormalities

[0116] 14. Total cholesterol

[0117] 15. Sodium

[0118] 16. Atrial fibrillation

[0119] 17. Diuretics

[0120] 18.QRS duration

[0121] 19. Digoxin

[0122] 20. Hypertension.

[0123] The remaining 8 features of 28 highly correlated survival predictions after removing the 20 key variables are:

[0124] 21. Gender

[0125] 22. Atrioventricular block

[0126] 23. Primary or secondary prevention

[0127] 24. History of coronary artery intervention

[0128] 25. Diabetes

[0129] 26. History of stroke

[0130] 27. Use of aldosterone receptor antagonists

[0131] 28. Warfarin use.

[0132] It should be noted that

[0133] The 28 highly correlated features for survival prediction screened out will be used in the subsequent training and optimization of the survival prediction model for patients with implantable cardioverter defibrillators;

[0134] The 20 key variables selected will be used to explain the contribution of features to the survival prediction model of implantable cardioverter defibrillator patients and expressed in qualitative and quantitative ways, that is, Figure 6 The content presented.

[0135] S5, training and optimizing models

[0136] The constructed implantable cardioverter-defibrillator patient survival prediction model is trained and optimized using the 28 survival prediction high-correlation features obtained in step S4, so that the implantable cardioverter-defibrillator patient survival prediction model can achieve "predicting the survival probability of an implantable cardioverter-defibrillator patient on any day within ten years."

[0137] Specifically,

[0138] To predict the probability of survival for patients with implantable cardioverter-defibrillators on any given day within a decade, the mortality risk prediction method of this embodiment is based on training and optimization of a random survival forest (RSF) model. Prior to model construction, a variable screening step was performed using the Cox regression method to initially select 28 candidate variables (highly correlated with survival prediction), including core biochemical indicators such as LVEDD, NYHA, UA, BUN, and pro-BNP.

[0139] During model training, the RSF model uses Bootstrap resampling technology to construct multiple survival decision trees and employs the Log-rank splitting criterion for feature partitioning, thereby learning and modeling patients' long-term survival. By integrating the predictions from multiple decision trees, the model generates a personalized survival function, capable of outputting survival probabilities at any given time point (e.g., any day within a decade), demonstrating continuous and dynamic prediction capabilities.

[0140] To optimize model performance, the mortality risk prediction method of this embodiment uses a grid search method to determine the optimal hyperparameter combination, including a number of decision trees (n_estimators) of 100, a maximum tree depth (max_depth) of 40, and a minimum number of sample splits (min_samples_split) of 5. The final model can output a complete survival probability curve for each patient. The model's predictive performance is evaluated and calibrated using metrics such as C-index, Brier score, and time-dependent AUC, and its clinical application value is verified in conjunction with decision curve analysis (DCA), ensuring that the model can serve as an important reference tool for physicians to develop personalized follow-up and intervention strategies.

[0141] Through the above approach, the implantable cardioverter-defibrillator patient survival prediction model can effectively identify the complex relationship between physiological and clinical variables, dynamically update the survival probability of implantable cardioverter-defibrillator patients, and accurately predict the patient's long-term survival probability.

[0142] S6. Survival Prediction and Risk Stratification

[0143] S61, using the implantable cardioverter-defibrillator patient survival prediction model trained and optimized in step S5, to perform survival prediction on the implantable cardioverter-defibrillator patient, and to predict the survival probability of the implantable cardioverter-defibrillator patient on any day within ten years.

[0144] S62, using an implantable cardioverter defibrillator patient survival prediction model to predict the five-year survival probability of the implantable cardioverter defibrillator patient, and then dividing the implantable cardioverter defibrillator patient into a high-risk group, a medium-risk group, and a low-risk group based on the five-year survival probability.

[0145] Regarding model interpretation and clinical application, the SHapley Additive Explanations (SHAP) method was used to improve the interpretability and clinical practicality of the results of the implantable cardioverter-defibrillator patient survival prediction model. SHAP values quantify the contribution of each feature to patient survival prediction, allowing clinicians to more clearly understand the impact of key physiological parameters on survival probability, thereby enhancing the model's transparency and practicality.

[0146] It should be noted that in this embodiment, the Kaplan-Meier survival curve was used to further validate the risk stratification effectiveness of the survival prediction model for patients with implantable cardioverter defibrillators. The results showed that the survival probability of patients in the high-risk group was significantly lower than that of the low-risk group (hazard ratio (HR) 4.50, 95% confidence interval (CI) 3.10–6.55, P value < 0.0001), while the mortality risk of patients in the intermediate-risk group was also increased (HR 2.49, 95% CI 1.69–3.68, P value < 0.0001). The statistical significance of this difference validates the effectiveness and reliability of the RSF in distinguishing different risk groups. The RSF has powerful nonlinear modeling capabilities and can effectively handle multidimensional data such as cardiac function, renal function, and biomarkers, and capture the interactions between these indicators in survival probability prediction, thereby optimizing personalized risk assessment. Compared with traditional statistical models, the RSF has significant advantages in handling high-dimensional, nonlinear, and missing data, and can provide more accurate survival probability prediction and risk stratification. This helps medical staff identify high-risk patients and take timely intervention measures, thereby optimizing the allocation of medical resources, improving the accuracy and effectiveness of treatment, and making individualized treatment plans more precise.

[0147] In this embodiment, a SHAP value analysis of a survival prediction model for patients with implantable cardioverter defibrillators (ICDs) revealed the 20 most important features influencing patient survival prediction. Among them, left ventricular end-diastolic diameter (LVEDD) had the highest SHAP value, indicating its most significant impact on risk assessment. In addition, other features such as New York Heart Association class (NYHA), uric acid (UA), blood urea nitrogen (BUN), NT-proBNP, and left ventricular ejection fraction (LVEF) also play an important role in survival prediction. SHAP values assign positive and negative values to each feature: positive values indicate an increased risk of death, while negative values indicate an increased probability of survival. For example, a higher LVEF may be associated with a better prognosis, while elevated NT-proBNP levels may indicate an increased risk of heart failure. The interpretability of the RSF model is crucial for clinical decision-making. SHAP analysis not only reveals the relationship between survival probability and key physiological parameters but also provides personalized risk assessments. This capability enables the RSF model to transcend traditional black-box machine learning, providing intuitive and actionable insights. For example, based on the SHAP results, doctors can focus more specifically on patients with high-risk features, such as enlarged LVEDD or elevated NT-proBNP levels, and adjust treatment plans in a timely manner to improve patient prognosis.

[0148] This embodiment also provides a dynamic death risk prediction system for patients with implantable cardioverter defibrillators in response to the above-mentioned death risk prediction method. The death risk prediction system is essentially a computer system, which is provided with a computer program that can implement the above-mentioned death risk prediction method. That is, the death risk prediction system running program can execute the above-mentioned death risk prediction method.

[0149] The following summarizes how the mortality risk prediction method and system provided by this embodiment solve technical problems and achieve technical effects:

[0150] 1) The existing technology cannot accurately predict the patient's survival probability on a certain day;

[0151] In the death risk prediction method and system of this embodiment, the survival prediction model (RSF model) for patients with implantable cardioverter defibrillators is trained and optimized based on random survival forests. 28 candidate variables (highly correlated features for survival prediction) are screened through Cox regression, including core indicators such as LVEDD and NYHA. The model uses Bootstrap resampling technology to construct multiple decision trees, uses the Log-rank splitting criterion for feature partitioning, and integrates the prediction results to generate an individualized survival function, thereby achieving "predicting the survival probability of patients with implantable cardioverter defibrillators on any day within ten years."

[0152] 2) Existing survival analysis models (such as the Cox proportional hazards model) mostly use linear assumptions for risk assessment, which makes it difficult to effectively capture complex nonlinear relationships and interactions in high-dimensional data.

[0153] In the death risk prediction method and system of this embodiment, a random survival forest (RSF) model is adopted to construct multiple decision trees through an ensemble learning method, which can adaptively handle high-dimensional, nonlinear and missing data problems.

[0154] The RSF model integrates multiple decision trees and uses bootstrap sampling and random feature selection to effectively capture the interactive relationships between complex variables such as cardiac function, kidney function, and biomarkers, thereby improving the model's prediction accuracy and stability.

[0155] 3) Traditional survival analysis models only provide static risk estimates and are unable to reflect the dynamic changes in patients' survival risks over time.

[0156] In the death risk prediction method and system of this embodiment, the RSF model can dynamically update the patient's survival probability and provide an individualized survival probability prediction.

[0157] The RSF model dynamically updates the patient's survival probability, enhancing the timeliness and clinical practicality of the prediction. It can more accurately reflect the changes in the patient's survival risk over time and provide stronger support for clinical decision-making.

[0158] 4) When processing high-dimensional data, existing models are prone to increase model complexity due to too many features, which affects the model's interpretability and predictive performance.

[0159] In the death risk prediction method and system of this embodiment, the SHapley Additive Explanations (SHAP) method is used to interpret the RSF model, screen out key variables and rank their importance.

[0160] SHAP value analysis quantifies the contribution of each feature to survival prediction, making the model's decision-making process more transparent and explainable. This not only improves the interpretability of the model but also helps clinicians better understand which physiological indicators have the greatest impact on a patient's survival risk.

[0161] 5) When dealing with missing data, existing technologies often use simple mean filling or median filling, which may lead to information loss and bias.

[0162] In the death risk prediction method and system of this embodiment, in data preprocessing, an iterative imputation method is used to process missing data.

[0163] The iterative filling method is more accurate than simple mean filling or median filling by iteratively predicting missing values. It can reduce the deviation of missing data on analysis results and maintain the integrity of patients' physiological information.

[0164] 6) When existing technologies process different medical indicators, due to the large differences in numerical ranges, the model may assign too high weights to certain features.

[0165] In the death risk prediction method and system of this embodiment, the data is normalized to adjust different indicators to similar ranges.

[0166] Normalization ensures that the model treats each feature more fairly during training, avoids bias caused by differences in numerical ranges, and helps identify important features related to survival probability.

[0167] Here are some examples for reference:

[0168] Case 1: Data preprocessing and machine learning model development

[0169] Figure 1 The research data processing and machine learning model development process are demonstrated. The data were collected from 3,175 patients with ICD or CRT-D implants in four large hospitals in China. 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 predictive effect on survival probability. During the model development process, a variety of machine learning algorithms (such as RF, LR, MLP, etc.) were optimized to find the optimal model. In the end, the multi-layer perceptron (MLP) model showed excellent performance in predicting the three-year survival probability.

[0170] The original data contained 86 clinical features. During data preprocessing, 45 features with more than 60% missing values were removed, leaving 41 features for subsequent analysis. Subsequently, Cox regression feature selection was used to identify 28 key features, of which left ventricular end-diastolic diameter (LVEDD) showed the most significant predictive effect on survival probability. During model development, multiple machine learning algorithms were optimized to identify the optimal model. Ultimately, the RSF model demonstrated excellent performance in predicting dynamic survival probability.

[0171] Case 2: Optimal Hyperparameters for the Model

[0172] In this case, the optimal hyperparameters for each model are screened to ensure that the model achieves the best performance on different datasets (such as Figure 8 During the hyperparameter optimization process, grid search and tuning strategies were used to select the optimal parameter configuration for each model.

[0173] For the Cox proportional hazards model (Cox), different alpha values (0, 0.1, 0.01), maximum number of iterations (200, 1000, 2000), and convergence tolerances (1e-5, 1e-6, 1e-7) were tested, and finally alpha = 0.1, n_iter = 200, tol = 1e-6 were selected.

[0174] Coxnet survival analysis (CSA) was tuned using the L1 ratio parameter (0.5, 1, 0.1) and the maximum number of iterations (1000, 2000), and l1_ratio = 0.5, max_iter = 1000, tol = 1e-10 was finally selected.

[0175] The extreme survival tree (EST) was tuned between different maximum tree depths (5, 10, 20, 50), minimum number of leaf nodes (1, 2, 5, 10), and minimum number of sample splits (10, 20, 50). Finally, max_depth = 50, min_samples_leaf = 1, min_samples_split = 10 were selected.

[0176] Gradient boosting survival analysis (GBA) was performed using a grid search with different numbers of decision trees (50, 100, 200), maximum depths (2, 3, 5), minimum number of sample splits (1, 2, 5), and learning rates (0.1, 0.01, 0.001). The final result was n_estimators = 200, max_depth = 3, min_samples_split = 5, and learning_rate = 0.1.

[0177] Random survival forest (RSF) was optimized by varying the number of decision trees (100, 200), maximum tree depth (5, 10, 20, 40), and minimum number of sample splits (1, 2, 5, 10), and finally n_estimators = 100, max_depth = 40, and min_samples_split = 5 were selected.

[0178] These optimized settings ensured the stable performance of each model on both internal and external test sets, and demonstrated the superiority of RSF in predicting survival of ICD / CRT-D patients.

[0179] Case 3: Model Performance Evaluation

[0180] Figure 2 (A,B,C) and Figure 7 The performance evaluation of the RSF model on internal and external test sets is presented, further illustrating the predictive stability and effectiveness of the model. Figure 8 The detailed performance indicators of RSF are as follows: the C-index of RSF in the internal test set is 0.71 (95% CI: 0.67–0.75) and the C-index of RSF in the external test set is 0.70 (95% CI: 0.58–0.82), indicating that the model has good discrimination ability on different datasets. Figure 7 The performance of the RSF model in terms of time-dependent AUC is also shown: these results demonstrate the robustness of RSF in long-term survival prediction, especially showing strong generalization ability on complex clinical data.

[0181] Furthermore, compared to other models (such as Cox, CSA, EST, and GBA), RSF outperformed other models on multiple key metrics, particularly the C-index and time-dependent AUC. Cox and CSA performed poorly on all metrics, with CSA's C-index being only 0.52 (95% CI: 0.48–0.57) on the internal test set and even lower at 0.51 (95% CI: 0.39–0.64) on the external test set, indicating its poor adaptability to complex nonlinear data.

[0182] Case 4: Risk stratification and Kaplan-Meier survival analysis based on the five-year survival probability prediction value of the RSF model

[0183] In this study, patients with ICD or CRT-D were risk-stratified using the RSF model's predicted five-year survival probability, dividing them into high-risk, intermediate-risk, and low-risk groups. Risk stratification based on RSF-predicted survival probabilities ensures the model's high accuracy in distinguishing patients at different risk levels.

[0184] After determining the risk stratification, Kaplan-Meier survival analysis was performed to further validate the stratification results. The Kaplan-Meier curves showed that patients in the high-risk group had a significantly lower probability of survival during the five-year follow-up period than those in the low-risk group (HR: 4.50, 95% CI: 3.10–6.55, P < 0.0001), while patients in the intermediate-risk group also had an increased risk of death (HR: 2.49, 95% CI: 1.69–3.68, P < 0.0001). This result demonstrates that the RSF model can effectively identify patients at higher survival risk and further demonstrates its value in personalized risk assessment and clinical decision support.

[0185] Case 5: Model Interpretation Analysis

[0186] Figure 4 A model interpretability analysis based on SHapley Additive exPlanations (SHAP) values was presented to quantify the impact of each feature on survival prediction. The results showed that LVEDD ranked first in feature importance, with significant positive and negative impacts on patient survival probability. This interpretability analysis not only enhances model transparency but also provides important reference indicators for clinicians, helping them better understand the model's predictions and assist in personalized decision-making.

[0187] The above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the dynamic mortality risk of patients with implantable cardioverter defibrillators, characterized by: The death risk prediction method comprises: S1, constructing a survival prediction model for patients with implantable cardioverter defibrillators based on the random survival forest model; S2, for patients with implantable cardioverter defibrillators, collect the patient's baseline data and survival status; S3, performing data preprocessing on the collected patient baseline information and survival status to obtain preprocessed patient baseline data; S4, performing feature selection on the patient baseline data after pretreatment, and screening out features with high correlation with survival prediction using the COX method; S5, training and optimizing the implantable cardioverter defibrillator patient survival prediction model; S6, using the trained and optimized implantable cardioverter defibrillator patient survival prediction model to predict the survival of implantable cardioverter defibrillator patients.

2. The method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators according to claim 1, wherein: The highly correlated features for survival prediction include 28 feature data, namely left ventricular end-diastolic diameter, New York Heart Association functional class, uric acid, blood urea nitrogen, N-terminal pro-brain natriuretic peptide, left ventricular ejection fraction, age, serum creatinine, albumin, left atrial diameter, new oral anticoagulants, amiodarone, intraventricular conduction abnormalities, total cholesterol, sodium, atrial fibrillation, diuretics, QRS wave duration, digoxin, hypertension, gender, atrioventricular block, primary or secondary prevention, history of coronary artery intervention, diabetes, history of stroke, use of aldosterone receptor antagonists, and use of warfarin.

3. The method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators according to claim 1, wherein: Step S4 also includes: using the SHapley Additive Explanations method to further screen key variables from the highly correlated survival prediction features, and the screened key variables are used to explain the contribution of the features to the implantable cardioverter defibrillator patient survival prediction model.

4. The method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators according to claim 3, wherein: The number of key variables screened out was 20, including left ventricular end-diastolic diameter, New York Heart Association functional class, uric acid, blood urea nitrogen, N-terminal pro-brain natriuretic peptide, left ventricular ejection fraction, age, serum creatinine, albumin, left atrial diameter, new oral anticoagulants, amiodarone, intraventricular conduction abnormalities, total cholesterol, sodium, atrial fibrillation, diuretics, QRS duration, digoxin, and hypertension.

5. The method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators according to claim 1, wherein: The data preprocessing includes missing data filling, data normalization and one-hot encoding.

6. The method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators according to claim 1, wherein: Step S5 further includes: During model training, the RSF model uses Bootstrap resampling technology to construct multiple survival decision trees and utilizes the Log-rank splitting criterion for feature partitioning, thereby learning and modeling the patient's long-term survival. The model integrates the prediction results of multiple decision trees to generate an individualized survival function, enabling it to output the survival probability at any given time point and possessing continuous time dynamic prediction capabilities.

7. The method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators according to claim 6, wherein: Step S5 further includes: The optimal hyperparameter combination was determined using a grid search method, including a maximum number of decision trees of 100, a maximum tree depth of 40, and a minimum number of sample splits of 5. The final model output a complete survival probability curve for each patient. The model's predictive performance was evaluated and calibrated using C-index, Brier score, and time-dependent AUC metrics, and its clinical application value was verified using decision curve analysis.

8. The method for predicting dynamic mortality risk in patients with implantable cardioverter defibrillators according to claim 1, wherein: Step S6 further includes: The implantable cardioverter defibrillator (ICD) patient survival prediction model was used to predict the five-year survival probability of ICD patients. Based on the five-year survival probability, ICD patients were divided into high-risk group, medium-risk group and low-risk group.

9. A dynamic mortality risk prediction system for patients with implantable cardioverter defibrillators, characterized by: The death risk prediction system is provided with a computer program capable of implementing the death risk prediction method according to claim 1.

Images