Method and system for predicting in-hospital mortality risk of surgical patients with acute aortic dissection type a based on interpretable machine learning

Abstract

This invention discloses a method and system for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning. The method includes: collecting clinical data from patients undergoing acute type A aortic dissection surgery and constructing an original feature dataset; evaluating the feature importance of each feature in the original feature dataset to obtain a score for each feature, and filtering based on the scores to obtain a core feature dataset; inputting the core feature dataset corresponding to the target patient into a pre-trained in-hospital mortality risk prediction model, and outputting the predicted in-hospital mortality probability of the target patient; calculating the feature contribution corresponding to the predicted in-hospital mortality probability, and generating an interpretable result for the mortality risk prediction of the target patient. The technical solution of this invention can reduce model complexity, improve model generalization ability, and improve prediction accuracy and interpretability.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data analysis and clinical decision support technology, specifically to a method and system for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning. Background Technology

[0002] Acute Type A Aortic Dissection (ATAAD) is a rapidly progressing and extremely deadly cardiovascular emergency. Even with emergency surgical intervention, the in-hospital mortality rate can still reach approximately 20%–25%. Therefore, accurate assessment of patient mortality risk in the early perioperative period is a key issue for improving prognosis and optimizing resource allocation.

[0003] Existing technologies mainly include the following categories: ① Traditional statistical models: Risk scoring or nomogram models, represented by logistic regression, are usually built based on a small number of variables and assume a linear relationship between variables and outcomes. They struggle to characterize the nonlinear relationships and higher-order interactions prevalent in complex clinical data, resulting in limited model generalization ability. ② Conventional machine learning prediction methods: Some studies have attempted to improve prediction accuracy using machine learning algorithms such as random forests and gradient boosting machines. However, these models typically lack clear interpretable mechanisms and are considered "black box models." Clinicians find it difficult to understand the basis for the prediction results, limiting their application in actual clinical decision-making. They also have the following shortcomings: ① Insufficient variable selection methods: Existing prediction models often rely on empirically selected variables or univariate analysis results, potentially overlooking key factors with significant but nonlinear impacts on outcomes, or introducing redundant variables, leading to increased model complexity, decreased stability, and reduced generalization ability. ② Lack of external or time-based validation, resulting in insufficient generalization ability.

[0004] Therefore, there is an urgent need for a technical solution for predicting the risk of in-hospital mortality after ATAAD that combines high predictive performance with good interpretability and can be implemented based on routine clinical data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning, which reduces model complexity, improves model generalization ability, and enhances prediction accuracy and interpretability.

[0006] The present invention adopts the following technical solution.

[0007] The first aspect of the present invention provides a method for predicting the risk of in-hospital mortality in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning, comprising:

[0008] Clinical data of patients undergoing acute type A aortic dissection surgery were collected, and a raw feature dataset was constructed.

[0009] The features in the original feature dataset are evaluated for their importance to obtain a score for each feature. The core feature dataset is then selected based on the scores.

[0010] Input the core feature dataset corresponding to the target patient into the pre-trained in-hospital mortality risk prediction model, and output the in-hospital mortality prediction probability of the target patient; obtain the pre-trained in-hospital mortality risk prediction model by minimizing the loss function;

[0011] The feature contribution rate corresponding to the in-hospital mortality prediction probability is calculated to generate the interpretation results of the mortality risk prediction for the target patient.

[0012] Optionally, feature importance evaluation can be performed on the features in the original feature dataset to obtain a score for each feature, including:

[0013] A random forest model is used to evaluate the feature importance of features in the original feature dataset. The random forest model consists of multiple decision trees. Starting from the root node of each decision tree, training samples are extracted. At the non-leaf nodes of each decision tree, some features are selected from all features as candidate splitting features. The splitting threshold with the lowest Gini impurity after splitting is selected as the optimal splitting threshold. Node splitting is performed based on the optimal splitting threshold until there are no more splittable nodes in the decision tree. The decrease in Gini impurity caused by each feature in all decision trees is calculated. The feature importance is evaluated based on the decrease in Gini impurity of each feature to obtain the score of each feature.

[0014] Optionally, the core feature dataset obtained by filtering based on scores includes:

[0015] Using a random forest model as the base learner, in each round of feature elimination iteration, features are sorted from high to low scores, and a preset proportion of features are removed starting from the lowest score.

[0016] The predictive performance of the in-hospital mortality risk prediction model under the current feature subset is evaluated using multi-fold cross-validation.

[0017] The optimal feature combination is the subset of features that maximizes the average AUC value and whose AUC value, sensitivity, and specificity in each round of cross-validation meet the corresponding threshold requirements.

[0018] In each round of feature elimination iteration, the feature with a repetition rate exceeding a preset repetition rate threshold in the optimal feature combination is taken as the core prediction feature, and the data corresponding to the core prediction feature constitutes the core feature dataset.

[0019] Optionally, the core feature dataset obtained by filtering based on scores also includes:

[0020] The following comprehensive performance objective function is constructed to comprehensively evaluate the predictive performance and stability of the feature subset. The optimal feature combination is determined by maximizing the comprehensive performance objective function:

[0021] ,

[0022] in,

[0023] , , These are the mean values ​​of AUC, sensitivity, and specificity of the model on the validation set under K-fold cross-validation, respectively.

[0024] , , These are the standard deviations of AUC, sensitivity, and specificity, used to measure the stability of the model across different folds;

[0025] , , These are the weighting coefficients for the mean AUC, mean sensitivity, and mean specificity, respectively. For preset weighting coefficients, and .

[0026] Optionally, the loss function is calculated using the following formula:

[0027] ,

[0028] in, For the weight vector, For bias terms, Represents the slack variable vector. For the i-th sample Slack variables, For penalty parameters, Let i be the weight of the i-th sample. For kernel function, For sample labels.

[0029] Optionally, the calculation of the feature contribution corresponding to the in-hospital mortality prediction probability includes: weighting the marginal contribution of each feature in the prediction result of the in-hospital mortality risk prediction model under different feature combinations to obtain the corresponding feature contribution, wherein the marginal contribution of the feature is used to characterize the change in the prediction result of the in-hospital mortality risk prediction model after the feature is added or removed.

[0030] Optional clinical data may include at least: age, history of stroke, myocardial ischemia, hypotension, preoperative blood glucose, preoperative blood lactate level, cardiopulmonary bypass time, aortic cross-clamp time, circulatory arrest time, and total surgical time.

[0031] Optionally, the method further includes: constructing an in-hospital mortality risk prediction model based on the support vector machine algorithm, and using a network search method to jointly optimize the penalty parameter and kernel function parameter in the support vector machine algorithm, and selecting the parameter combination corresponding to the maximum cross-validation average AUC value as the optimal model parameter.

[0032] Optionally, the method further includes: jointly outputting the in-hospital mortality prediction probability and the feature contribution, so as to realize the simultaneous display of risk prediction and risk source.

[0033] A second aspect of the present invention provides a system for predicting the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery based on interpretable machine learning, implementing the aforementioned method for predicting the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery based on interpretable machine learning, the system comprising:

[0034] The data acquisition module is used to collect clinical data from patients undergoing acute type A aortic dissection surgery and to construct a raw feature dataset.

[0035] The feature filtering module is used to evaluate the feature importance of features in the original feature dataset to obtain a score for each feature, and then filter based on the score to obtain the core feature dataset.

[0036] The risk prediction module is used to input the core feature dataset corresponding to the target patient into a pre-trained in-hospital mortality risk prediction model and output the in-hospital mortality prediction probability of the target patient. The in-hospital mortality risk prediction model is constructed based on the support vector machine algorithm and obtained by minimizing the loss function.

[0037] The results interpretation module is used to calculate the feature contribution corresponding to the in-hospital mortality prediction probability and generate interpretation results for the mortality risk prediction of the target patient.

[0038] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded onto the processor, implements the above-described method for predicting the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery based on interpretable machine learning.

[0039] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning.

[0040] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0041] This invention integrates multi-dimensional perioperative clinical data and utilizes machine learning algorithms to construct a high-performance mortality risk prediction model. It employs a systematic feature selection method to screen a set of core, stable, and readily available predictive indicators from a large number of candidate variables, reducing redundant variables, lowering model complexity, and improving model simplicity and generalization ability. By introducing interpretable artificial intelligence technology, the prediction results are quantitatively interpreted, clarifying the direction and degree of contribution of each predictive variable to mortality risk. Finally, the model is validated using independent external population data to ensure its reliability and applicability in different clinical settings.

[0042] The prediction method proposed in this invention has high prediction accuracy and stability. The model maintains good performance in different regions and time periods. By introducing interpretability analysis, the clinical usability of the model is significantly improved. It can effectively assist in the perioperative risk management of patients with acute type A aortic dissection. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0044] Figure 1 This is a schematic diagram illustrating the construction and verification process of a postoperative in-hospital mortality risk prediction model provided in this embodiment of the disclosure;

[0045] Figure 2 This is a schematic diagram of a predictor variable selection process provided in an embodiment of this disclosure;

[0046] Figure 3 This is a schematic diagram of ROC curves for identifying in-hospital mortality after ATAAD surgery using different prediction models provided in this embodiment of the disclosure; Figure 3 In the diagram, A represents the ROC curve of the training set, B represents the ROC curve of the internal validation set, C represents the ROC curve of the external validation set, and D represents the feature descriptions of the six machine learning models.

[0047] Figure 4This is a schematic diagram illustrating the time-external validation results of a prediction model provided in an embodiment of this disclosure;

[0048] Figure 5 This is a schematic diagram of the internal verification results of a prediction model provided in an embodiment of this disclosure;

[0049] Figure 6 This is a SHAP interpretability analysis diagram of deceased and surviving patients provided in an embodiment of this disclosure. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0051] To accurately predict the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery and overcome the technical shortcomings of existing technologies, such as insufficient prediction accuracy, poor model generalization ability, and lack of clinical interpretability, this invention provides a method for predicting the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery based on interpretable machine learning, including the following:

[0052] Step 1: Collect clinical data from patients undergoing acute type A aortic dissection surgery and construct the original feature dataset.

[0053] Clinical data should include at least: age, history of stroke, myocardial ischemia, hypotension, preoperative blood glucose, preoperative blood lactate level, cardiopulmonary bypass time, aortic clamping time, circulatory arrest time, and total surgical time.

[0054] Hypotension is defined as at least one of the following: a sustained systolic blood pressure below 90 mmHg for more than 30 minutes, a mean arterial pressure below 60 mmHg, or an acute drop in systolic blood pressure of more than 40 mmHg from baseline.

[0055] It should be noted that the data collected are routine clinical information obtained from patients during the perioperative period, and do not rely on additional testing equipment or invasive procedures.

[0056] The clinical data shall include at least the following types:

[0057] 1) Demographic data: including basic information such as patient age and gender, with age entered into the system as a continuous numerical value.

[0058] 2) Past medical history data: including past stroke history, including but limited to whether there is a clear imaging or clinical diagnosis of stroke, past cardiovascular disease history, etc.

[0059] 3) Preoperative physiological and biochemical data: including preoperative blood glucose level, preoperative blood lactate level, systolic blood pressure, diastolic blood pressure and mean arterial pressure, etc.

[0060] 4) Perioperative status indicators: including whether there is myocardial ischemia and whether there is hypotension. Hypotension is defined as: systolic blood pressure below 90 mmHg for more than 30 minutes, mean arterial pressure below 60 mmHg, or systolic blood pressure dropping more than 40 mmHg from baseline.

[0061] 5) Surgical procedure parameters: including cardiopulmonary bypass time, aortic clamping time, circulatory arrest time, and total surgical time.

[0062] The aforementioned clinical data, after being manually entered, automatically acquired through electronic medical record systems or hospital information systems, is uniformly converted into a structured data format to form the original clinical feature dataset.

[0063] Step 2: Preprocess the data collected in Step 1. Preprocessing includes handling missing values, outliers, and standardization to obtain a standardized clinical feature dataset.

[0064] In step 2, the missing proportion of each clinical data is statistically analyzed. When the missing proportion of clinical data is less than the preset missing proportion threshold, the K-nearest neighbor imputation algorithm is used to imput the missing values. When the missing proportion of clinical data is greater than or equal to the preset missing proportion threshold, the clinical data is deleted.

[0065] In one embodiment of the present invention, data preprocessing includes the following steps:

[0066] Step 2.1: Missing value handling. The missing value proportion of each clinical variable is statistically analyzed. When the missing value proportion of a variable is less than a preset missing value proportion threshold, KNN (K-Nearest Neighbors) is used to impute the missing values. When the missing value proportion is greater than or equal to the preset missing value proportion threshold, the variable is removed to avoid introducing systematic bias. Preferably, the preset missing value proportion threshold is 10%.

[0067] Step 2.2: Outlier handling. For continuous clinical variables, the quantile truncation method is used to limit their values ​​to the range of the 1st percentile to the 99th percentile in order to reduce the adverse impact of extreme outliers on the model prediction results.

[0068] Step 2.3: Standardization processing. After handling missing and outlier values, the continuous variables are numerically standardized so that clinical indicators of different dimensions can participate in model training and prediction on a unified scale.

[0069] After the above processing, a standardized clinical feature dataset is obtained, which provides reliable input for subsequent feature selection and model construction, thereby improving model stability and prediction accuracy.

[0070] Step 3: Score the features in the original feature dataset according to their importance, and then filter them based on the scores to obtain the core feature dataset. Step 3 specifically includes:

[0071] Step 3.1: Assess the importance of features in the standardized clinical feature dataset to obtain a score for each feature, and then filter based on the scores to obtain the core feature dataset.

[0072] Specifically, a Random Forest (RF) model is used to preliminarily assess the feature importance of features in the standardized clinical feature dataset, obtaining feature scores. Training samples are selected from the standardized clinical feature dataset. The Random Forest model consists of multiple decision trees, each constructed by randomly sampling samples from the training data using a bootstrap sampling method. At each node split, a subset of features are randomly selected as candidate split features. Feature importance is assessed based on the Gini Impurity reduction method, which calculates the cumulative average Gini Impurity reduction value brought about by a feature when it is used as a split node across all decision trees. This value is used as the feature score. All clinical features are preliminarily ranked based on their scores, providing a basis for subsequent feature selection.

[0073] It should be noted that candidate splitting features are used to limit the search space and improve model diversity and feature scoring stability.

[0074] Specifically, the feature importance evaluation of the features in the original feature dataset yields a score for each feature, including:

[0075] Starting from the root node of the decision tree, multiple clinical features are randomly selected as candidate splitting features. For each candidate splitting feature, all possible splitting thresholds are traversed, and the Gini impurity of the left and right child nodes after splitting is calculated:

[0076] ;

[0077] in, This represents the Gini impurity of a node. This represents the proportion of the k-th class samples within a node to the total number of samples. Indicates the sample type, which includes live and dead samples;

[0078] The optimal splitting threshold is selected as the splitting threshold that minimizes the Gini impurity after splitting. Node splitting is then performed based on this optimal threshold, and the decrease in Gini impurity after each split is calculated.

[0079]

[0080] in, The Gini impurity decrease value represents the candidate splitting characteristic. This represents the number of samples in the left child node after the split. This represents the number of samples in the right child node after the split. This represents the number of samples in the parent node. This indicates the Gini impurity of the candidate splitting characteristic before splitting. This represents the Gini impurity of the left child node after splitting. This represents the Gini impurity of the right child node after splitting;

[0081] Record the Gini impurity decrease value of this splitting feature in this node split, and continue splitting child nodes until the decision tree has no more splittable nodes;

[0082] For each clinical feature, the cumulative Gini impurity decrease value across all N decision trees is summed to obtain the total cumulative value. The total cumulative value is then divided by the total number of decision trees N to obtain the score of the corresponding clinical feature. The higher the score, the greater the contribution of the clinical feature to distinguishing between survival and death, and the more core the feature.

[0083] More preferably, candidate splitting features are selected using a weighted approach based on feature importance. The importance scores of all features are normalized into weights, and weighted random sampling is used to extract features from all features according to these weights, forming a candidate splitting feature set for the node. If a feature is repeatedly selected as the best splitting feature, its weight is dynamically increased (e.g., multiplied by 1.1); conversely, its weight is decreased (e.g., multiplied by 0.9). Candidate splitting features represent the shortlist for node splitting, and the decrease in Gini impurity during node splitting serves as a validation metric for the candidate splitting features. Furthermore, the splitting results update the feature weights in reverse. This iterative optimization of feature weights improves the accuracy of candidate splitting feature selection and enhances the model's accuracy in identifying high-risk samples.

[0084] Step 3.2: Using the random forest model as the base learner, in each round of feature elimination iteration, sort the features from high to low scores and remove a preset proportion of features starting from the lowest score.

[0085] After obtaining the random forest importance scores for each feature, these scores are used as the ranking criteria for RFE (Recursive Feature Elimination). Specifically, in the RFE process, the random forest model is used as the base learner. In each iteration of feature elimination, the feature importance scores are recalculated based on the current model, and features with the lowest importance are gradually removed in ascending order of importance, according to a predetermined proportion. This approach combines the random forest model with the recursive feature elimination algorithm, enabling the feature selection process to simultaneously consider the ability to model nonlinear feature relationships and the ability to progressively optimize feature subsets.

[0086] Preferably, the preset ratio ranges from 5% to 10%.

[0087] Step 3.3: In each round of feature elimination iteration, K-fold cross-validation is used to evaluate the prediction performance of the prediction model under the current feature subset.

[0088] Specifically, in each round of recursive feature elimination, K-fold cross-validation is used to evaluate the predictive performance of the model under the current feature subset, where K is preferably 10. The specific implementation is as follows: the training data is randomly divided into 10 subsets of similar size, with 9 subsets used for model training and the remaining subset used for validation. This process is repeated 10 times, ensuring each subset serves as a validation set, and the average performance index of each round of validation is calculated. The predictive performance is quantified using the Area Under the Curve (AUC), supplemented by sensitivity, specificity, calibration curve, and clinical decision curve as reference indicators. During feature elimination, the feature subset that maximizes the average AUC and maintains stable performance is selected as the optimal feature combination. Stable performance means that the model's AUC, sensitivity, and specificity all meet preset threshold requirements in multi-fold cross-validation; preferably, a comprehensive performance objective function is constructed to constrain the predictive performance.

[0089] In the recursive feature elimination process, to comprehensively evaluate the predictive performance and stability of the feature subset, the following comprehensive performance objective function J can be constructed:

[0090]

[0091] in:

[0092] , , These are the mean values ​​of AUC, sensitivity, and specificity of the model on the validation set under K-fold cross-validation, respectively.

[0093] , , These are the standard deviations of AUC, sensitivity, and specificity, used to measure the stability of the model across different folds;

[0094] , , These are the weighting coefficients for the mean AUC, mean sensitivity, and mean specificity, respectively. For the preset weighting coefficients, satisfy and The importance of each indicator and the severity of the penalty for stability can be adjusted according to actual clinical needs.

[0095] The optimal feature subset is selected by maximizing the objective function J. This subset not only has high average prediction performance, but also shows stability in cross-validation, i.e., a small standard deviation, thereby ensuring the model's generalization ability and clinical reliability.

[0096] Step 3.4: Repeat steps 3.1-3.3, selecting features in the optimal feature combination from each iteration whose feature repetition rate exceeds a preset repetition rate threshold as core prediction features. Construct the core feature dataset from the data corresponding to these core prediction features. By repeating the above steps, the core feature set for in-hospital mortality risk prediction is finally determined.

[0097] In this embodiment, based on a random forest model and employing a recursive feature elimination algorithm, the standardized clinical feature dataset is ranked by feature importance, and a core feature set for in-hospital mortality risk prediction is selected. This allows the feature selection process to simultaneously consider the ability to model nonlinear feature relationships and the ability to progressively optimize feature subsets. This feature selection technique can reduce redundant variables and improve model generalization ability while retaining key clinical information.

[0098] Step 4: Input the core feature dataset corresponding to the target patient into the pre-trained in-hospital mortality risk prediction model, and output the in-hospital mortality prediction probability of the target patient.

[0099] A model for predicting in-hospital mortality risk is constructed based on the support vector machine algorithm and a radial basis function kernel. The optimal penalty parameters and kernel function parameters of the model are determined through parameter optimization.

[0100] The pre-training of the in-hospital mortality risk prediction model includes: using the core feature data of patients who have undergone historical acute type A aortic dissection surgery as the input vector, and using whether the patient died in hospital as the supervision label to construct a binary classification training sample set; based on the support vector machine algorithm, the input vector is mapped to a high-dimensional feature space through the radial basis function kernel, and the optimal in-hospital mortality risk prediction model for distinguishing between in-hospital and non-in-hospital mortality is obtained by minimizing the objective function.

[0101] In one embodiment of the present invention, training an in-hospital mortality risk prediction model includes:

[0102] 1) The Support Vector Machine (SVM) algorithm is used as the basic prediction model.

[0103] 2) Radial basis functions were selected as kernel functions to capture nonlinear relationships between clinical variables.

[0104] 3) After feature selection, a grid search approach is used to jointly optimize the penalty parameter C and kernel parameter γ in the support vector machine model. Specifically, a parameter search space is predefined, where the penalty parameter C ranges from 0.01 to 100, and the kernel parameter γ ranges from 0.0001 to 1, with values ​​discretized using a logarithmic scale. For each parameter combination, the model's predictive performance is evaluated using 10-fold cross-validation on the training data, with AUC as the primary optimization objective function. Finally, the parameter combination that maximizes the average AUC value of the cross-validation is selected as the optimal model parameters.

[0105] 4) Using the feature-selected and preprocessed training data, the Support Vector Machine (SVM) model is trained under supervision. The supervision label is a binary classification label indicating whether a patient has died in hospital. During training, an optimization objective function based on the Hinge Loss function is used, and a penalty parameter C is used to balance maximizing the classification margin with minimizing the classification error. The SVM training objective is to minimize the loss function including the improved loss term, thereby maximizing the classification margin while reducing the probability of missing high-risk patients. In this embodiment, the SVM model is trained using a mature quadratic programming algorithm without introducing an additional learning rate parameter, thus ensuring the stability and convergence of the model training process. The final in-hospital mortality risk prediction model is obtained through the above method.

[0106] The loss function is expressed as follows:

[0107] ,

[0108] in, For the weight vector, For bias terms, Represents the slack variable vector. For the i-th sample The slack variable, namely the classification bias of the i-th sample. A value of 0 indicates that the sample is unbiased. A value greater than 0 indicates that there is a deviation. , This is the penalty parameter, used to control the trade-off between maximizing the margin and penalizing misclassification; The weight of the i-th sample is set according to the patient's in-hospital mortality risk. High-risk samples are given a larger weight so that they are more severely penalized when misclassified. For the kernel function, this disclosure embodiment uses radial basis functions. The sample labels are used, with death being the positive class.

[0109] By introducing sample weights related to in-hospital mortality risk The model focuses more on high-risk samples during training, giving them a higher penalty for misclassification, effectively reducing the false negative rate and improving clinical applicability.

[0110] This model can stably learn complex clinical risk patterns under limited sample conditions and is suitable for risk prediction scenarios for critically ill patients.

[0111] After the model training is completed, the present invention uses the in-hospital mortality risk prediction model for individual patient risk assessment.

[0112] In one embodiment of the present invention, the core feature set corresponding to the target patient is input into a pre-trained in-hospital mortality risk prediction model; the in-hospital mortality risk prediction model outputs the predicted probability of the target patient experiencing an in-hospital mortality event; the predicted probability is compared with a preset risk threshold, and if the predicted probability is greater than or equal to the preset risk threshold, the patient is classified into a high-risk group; otherwise, the patient is classified into a low-risk group. Preferably, the preset risk threshold is 10%.

[0113] The risk stratification technology described above can provide clinicians with clear and intuitive risk level information to assist in perioperative decision-making.

[0114] It should be noted that the predictive model can be deployed on a server-side platform, building a browser-based human-computer interaction interface. Doctors can input 10 clinical indicators corresponding to the patient, and the system will output in real time: the predicted probability of in-hospital mortality; risk stratification results including but not limited to high risk / low risk; and visualization results of feature contributions generated based on the SHAP algorithm.

[0115] Step 5: Calculate the feature contribution corresponding to the in-hospital mortality prediction probability to generate the interpretation results of the mortality risk prediction for the target patient.

[0116] Step 5, calculating the feature contribution corresponding to the in-hospital mortality prediction probability, includes:

[0117] By statistically analyzing and weighting the predictive contributions of each feature in different feature combinations, the feature contribution at the individual patient level is obtained.

[0118] In the model prediction stage, the SHAP (Shapley Additive Explanations) algorithm is introduced to calculate the feature contribution of each prediction result, i.e., the probability of in-hospital mortality prediction. Specifically, the SHAP algorithm is based on the Shapley value theory in game theory, treating the model prediction result as the "payoff" of multiple features working together. By calculating the marginal contribution of a feature among all possible feature combinations, the contribution of that feature to the prediction result is quantified. In this embodiment, the prediction output of the support vector machine model is used as the target value. The SHAP value corresponding to each clinical feature is calculated by weighted averaging the changes in the prediction result before and after the addition or removal of features. The weighting method is based on the Shapley value theory, and the marginal contribution is weighted and averaged according to the combination probability of different feature subsets.

[0119] This study determines the positive or negative contribution of each clinical feature to the predicted risk of in-hospital mortality in individual patients. The feature contribution is termed the SHAP value, whose sign and absolute value reflect the direction and degree of influence of the feature on the predicted outcome for an individual patient. A positive SHAP value indicates that the feature has a positive effect on the prediction of in-hospital mortality risk; a negative SHAP value indicates that the feature has a protective or negative effect on the prediction of in-hospital mortality risk. The absolute value of the SHAP value reflects the strength of the feature's contribution to the prediction of an individual patient, thus enabling quantitative and interpretable analysis of the prediction results.

[0120] The feature contribution results are displayed in a visual format, enabling clinicians to intuitively understand the sources of risk.

[0121] This technical solution effectively improves model transparency and clinical acceptability, enabling prediction results to be translated into actionable clinical interventions and solving the problem of difficult-to-interpret prediction results.

[0122] This invention combines multi-source clinical data collection, systematic feature screening, nonlinear prediction model construction, and interpretability analysis to achieve quantitative assessment of individualized mortality risk and analysis of risk sources for patients.

[0123] Compared with the prior art, the present invention has the following beneficial effects:

[0124] High predictive performance: The constructed SVM model demonstrated high discriminative ability in internal validation, outperforming various control models.

[0125] High interpretability: The SHAP method makes the model decision-making process transparent and clarifies the contribution of each risk factor to the prediction results.

[0126] The number of variables is concise and easy to obtain: only 10 routine clinical indicators are required, making it easy to promote and apply in different medical institutions.

[0127] External and time validations show that the model maintains stable predictive performance in both independent external cohorts and time cohorts, demonstrating a certain degree of generalization ability.

[0128] High clinical application value: It can provide data support for perioperative risk stratification, resource allocation and precision intervention.

[0129] like Figure 1 As shown, this disclosure provides a specific application of a method for predicting the risk of in-hospital mortality in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning. In this embodiment, patients who underwent acute type A aortic dissection surgery at a tertiary hospital in Jiangsu Province between January 1, 2014 and December 31, 2021 were selected as the research subjects. Figure 1 In this context, n represents the number of patients, and specific methods include:

[0130] Step 1: Collect clinical data from patients undergoing acute type A aortic dissection surgery and construct the original feature dataset.

[0131] Inclusion criteria included: 1) age ≥ 18 years; 2) confirmed diagnosis of acute type A aortic dissection by aortic CT angiography or echocardiography; 3) surgical treatment.

[0132] Exclusion criteria included: 1) those who received conservative or non-surgical treatment; 2) those whose symptoms lasted for more than 14 days; and 3) those who lacked the primary outcome indicator.

[0133] After screening, a total of 452 patients were included, of whom 63 experienced in-hospital deaths. The following perioperative clinical variables were collected from each patient, all derived from the electronic medical record system: 1) Demographic variables: age (years); 2) Past medical history variables: history of stroke (yes / no); 3) Preoperative biochemical indicators: blood glucose (mmol / L); blood lactate (mmol / L); 4) Perioperative status indicators: myocardial ischemia (yes / no); hypotension (yes / no); 5) Surgical time parameters: cardiopulmonary bypass time (minutes); aortic clamping time (minutes); circulatory arrest time (minutes); total surgical duration (hours).

[0134] Step 2: Preprocess the data collected in Step 1. Preprocessing includes handling missing values, outliers, and standardization to obtain a standardized clinical feature dataset. Specifically, Step 2 includes:

[0135] Missing value handling: The missing value ratio of each variable is calculated. Variables with a missing value ratio ≤10% are imputed using the K-nearest neighbor imputation algorithm; variables with a missing value ratio >10% are removed. In this embodiment, K=5.

[0136] Outlier handling: Quantity truncation is performed on continuous variables to restrict the variables to the range of the 1st percentile to the 99th percentile.

[0137] Standardization: Perform Z-score standardization on continuous variables to make their mean 0 and standard deviation 1.

[0138] Step 3: Score the features in the original feature dataset according to their importance, and then filter them based on the scores to obtain the core feature dataset.

[0139] refer to Figure 2 As shown, Figure 2 In this diagram, A represents the feature selection path based on the recursive feature elimination algorithm, B represents the ranking result of the predictor variable importance calculated by the random forest model, and C represents the feature contribution summary diagram of the support vector machine model based on the SHAP algorithm. The predictor variable selection process in this embodiment includes the following steps:

[0140] First, based on the preprocessed clinical feature dataset, a random forest model is used to calculate the importance score of each predictor variable, such as... Figure 2 As shown in Figure B. The importance score is obtained by statistically analyzing the average magnitude of the decrease in Gini impurity caused by each variable in the random forest model, and the variables are then ranked from high to low accordingly.

[0141] Based on this, a recursive feature elimination algorithm is introduced, using a random forest model as the base learner. In each iteration, some predictive variables are gradually removed from low to high according to their importance. After each round of feature elimination, the model's predictive performance is evaluated using ten-fold cross-validation. The predictive performance is evaluated primarily by the area under the receiver operating characteristic curve (AUC).

[0142] like Figure 2 As shown in Figure A, as the number of predictor variables gradually increases, the model AUC value shows a trend of first rising and then stabilizing. When the AUC reaches the inflection point of the trend change, the corresponding subset of predictor variables is determined as the optimal feature combination.

[0143] Finally, the SHAP algorithm is used to perform a global interpretation and analysis of the contribution of the optimal feature subset to the support vector machine model, such as... Figure 2 As shown in Figure C, this is used to verify the actual role of the selected variables in the model prediction.

[0144] In the training set, for example, 80% of patients were randomly selected, and a random forest model combined with a recursive feature elimination algorithm was used for feature selection. The specific process is as follows: in-hospital mortality was used as the dependent variable; the importance score of each feature was calculated using the random forest model; 5% of low-importance features were recursively eliminated in each round; the model's AUC value was evaluated using 10-fold cross-validation; the above process was repeated 50 times, retaining the features selected in more than 70% of the repetitions. Finally, 10 core predictive features were selected, including: cardiopulmonary bypass time, age, blood glucose, history of stroke, aortic clamping time, circulatory arrest time, total operation time, blood lactate, myocardial ischemia status, and hypotension status. Meanwhile, such as Figure 3 As shown in Figures A, B, and C, referencing the AUC values ​​of the ROC curves of the six models on the training set, internal validation set, and external validation set, and... Figure 3 The best prediction model is selected from the model features (including sensitivity, specificity, etc.) in the internal validation set shown in D.

[0145] Step 4: Input the core feature dataset corresponding to the target patient into the pre-trained in-hospital mortality risk prediction model, and output the predicted in-hospital mortality probability of the target patient. The prediction model is constructed using a support vector machine algorithm, specifically including: selecting a radial basis function as the kernel function; the search range of the penalty parameter C is 0.01–100; the search range of the kernel function parameter γ is 0.0001–1; and using grid search combined with 10-fold cross-validation to determine the optimal parameter combination. After training, the in-hospital mortality risk prediction model is obtained.

[0146] like Figure 5 As shown, in this embodiment of the invention, the performance of the prediction model is further evaluated using an internal validation dataset. Wherein, Figure 5 The calibration curve shown in Figure A is used to compare the consistency between the model's predicted probability and the actual mortality rate; Figure 5 The clinical impact curve shown in Figure B is used to illustrate the relationship between the number of positive cases predicted by the model and the actual number of positive cases at different risk thresholds. Figure 5 The decision curve analysis shown in Figure C was used to evaluate the clinical net benefit of the model relative to the "all intervention" or "no intervention" strategies at different threshold probabilities. Model performance was evaluated on an internal validation set (20% of patients), with the following results: Area under the receiver operating characteristic (AUC) of 0.785; sensitivity of 0.769; specificity of 0.935; and Brier score of 0.104. These results indicate that the model has high discriminative power and good calibration performance.

[0147] Step 5: Calculate the feature contribution corresponding to the in-hospital mortality prediction probability to generate the interpretation results of the mortality risk prediction for the target patient.

[0148] like Figure 6 As shown, the SHAP algorithm is used to perform individual-level interpretability analysis on the prediction model; among which, Figure 6 In the diagram, A represents the characteristic contribution of deceased patients, and B represents the characteristic contribution of surviving patients; the positive or negative SHAP value indicates the promoting or protective effect of the corresponding characteristic on the risk of in-hospital mortality, respectively, and the absolute value of the SHAP value reflects the degree of characteristic contribution.

[0149] This disclosure provides an external validation example of a method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning. The external validation dataset consists of 381 patients who underwent acute type A aortic dissection surgery at another tertiary hospital in Shandong Province between January 1, 2015, and December 31, 2022. The inclusion and exclusion criteria are the same as in the previous embodiment. The validation method involves directly applying the trained support vector machine model to the external validation dataset without retraining or adjusting the model parameters. The validation results are as follows: the model's performance metrics on the external validation set are as follows: AUC: 0.682; the calibration curve shows good consistency between the predicted probability and the actual mortality rate; decision curve analysis shows a positive net benefit within the threshold range of 0.10–0.70. These results indicate that the model possesses a certain degree of cross-regional generalization ability.

[0150] This disclosure also provides an external validation example of a method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning, used to verify the model's stability and generalization ability in patients at different time stages. Figure 4 As shown, to verify the stability of the prediction model in patients at different time stages, this embodiment of the invention selects independent patient data from subsequent time periods as an external time validation dataset. Without retraining the model structure and parameters, the already trained support vector machine model is directly applied to the external time validation dataset, and ROC curves are plotted as follows: Figure 4 A, calibration curve as shown Figure 4 The results of the analysis of B and decision curves are as follows: Figure 4 The calibration curve (C) is an important graphical tool for evaluating the accuracy of predictive models. It reflects the degree of consistency between the model's predicted risk probabilities and the actual situation.

[0151] Decision curve analysis is used to evaluate the clinical utility of predictive models and to determine whether the benefits of using the model to guide clinical decisions outweigh the risks. Decision curve analysis quantifies the clinical value of the model by calculating its net benefit at different threshold probabilities. The curve illustrates the model's performance at different risk thresholds and compares it with two extreme strategies: a "full intervention" strategy (assuming all patients receive treatment) and a "no intervention" strategy (assuming all patients receive no treatment). In this embodiment, the decision curve analysis results show that the model has a stable net clinical benefit, indicating that using the predictive model to guide clinical decisions (such as whether to take more aggressive intervention measures) can achieve a good balance between avoiding unnecessary interventions and timely treatment of high-risk patients.

[0152] The threshold probability on the horizontal axis refers to the critical risk value for taking intervention measures in clinical decision-making. It reflects the trade-off between the relative weight of "false positives" (unnecessary intervention for non-high-risk patients) and "false negatives" (missed diagnosis and no intervention for truly high-risk patients). Specifically, when the model predicts a patient's in-hospital mortality risk ≥ the threshold probability, clinicians will take intervention measures, such as transferring to the ICU or adjusting the treatment plan; when the predicted risk < the threshold probability, no special intervention will be taken. This threshold reflects the level of caution in clinical decision-making—the lower the threshold, the greater the concern about missed diagnosis and the more inclined to intervene; the higher the threshold, the greater the concern about unnecessary intervention and the more inclined to be conservative.

[0153] The vertical axis represents net profit, calculated using the following formula:

[0154]

[0155] Where N is the total number of patients, The threshold probability is defined as follows: true positives are the number of patients predicted as high-risk by the model who actually died, false positives are the number of patients predicted as high-risk by the model who actually survived, and weighting factors are also defined. This reflects the degree of penalty for false positive results. A higher net benefit value indicates higher clinical utility of the model at that threshold. Decision curve analysis in this embodiment shows that the model has a stable clinical net benefit, indicating that the predictive model can achieve a good balance between avoiding unnecessary interventions and timely treatment of high-risk patients.

[0156] In addition, such as Figure 4As shown in Figure D, an online risk prediction system is constructed based on the aforementioned prediction model. By inputting patient clinical indicators through a human-computer interaction interface, the system can output in-hospital mortality risk prediction results in real time. This embodiment selects 202 consecutive patients who underwent surgery at the same medical institution in Application Example 1 between January 1, 2022, and December 31, 2024, as the time-out validation dataset. Keeping the model structure, parameters, and preprocessing procedures completely unchanged, the model is directly applied to time-independent patient data. The model's performance on the time-out validation set is as follows: AUC: 0.719; calibration curve close to the ideal 45° line; decision curve analysis shows stable clinical net benefit. The results indicate that the model can adapt to changes in patient characteristics at different time stages and has good temporal stability.

[0157] In practical use, high-risk patients showed a significantly higher incidence of in-hospital mortality. SHAP interpretation results clearly revealed the major contributions of factors such as cardiopulmonary bypass time, blood glucose, and age to individual risk.

[0158] This system significantly improves the efficiency of clinicians in identifying high-risk patients and provides technical support for enhanced perioperative monitoring and intervention.

[0159] This invention provides a system for predicting the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery based on interpretable machine learning. The system operates as described above using the same method for predicting the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery based on interpretable machine learning. The system includes:

[0160] The data acquisition module is used to collect clinical data from patients undergoing acute type A aortic dissection surgery and to construct a raw feature dataset.

[0161] The feature filtering module is used to evaluate the feature importance of features in the original feature dataset to obtain a score for each feature, and then filter based on the score to obtain the core feature dataset.

[0162] The risk prediction module is used to input the core feature dataset corresponding to the target patient into a pre-trained in-hospital mortality risk prediction model and output the in-hospital mortality prediction probability of the target patient. The in-hospital mortality risk prediction model is constructed based on the support vector machine algorithm and obtained by minimizing the loss function.

[0163] The results interpretation module is used to calculate the feature contribution corresponding to the in-hospital mortality prediction probability and generate interpretation results for the mortality risk prediction of the target patient.

[0164] In one embodiment of the present invention, the above method steps are implemented by a computer system, which includes at least a data acquisition module, a data preprocessing module, a feature screening module, a risk prediction module, a result interpretation module, and a result output module.

[0165] It is understood that the system can be deployed on hospital information systems, clinical decision support systems or cloud computing platforms, and can realize the real-time assessment and display of patients' in-hospital mortality risk through human-computer interaction interfaces.

[0166] Regarding the system in the above embodiments, the specific manner in which each unit performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0167] The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the above-mentioned method for predicting the in-hospital mortality risk of patients undergoing acute type A aortic dissection surgery based on interpretable machine learning.

[0168] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning.

[0169] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0170] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. An acute type A aortic dissection surgical patient in-hospital mortality risk prediction method based on explainable machine learning, characterized in that, include: Clinical data of patients undergoing acute type A aortic dissection surgery were collected, and a raw feature dataset was constructed. The features in the original feature dataset are evaluated for their importance to obtain a score for each feature. The core feature dataset is then selected based on the scores. Input the core feature dataset corresponding to the target patient into the pre-trained in-hospital mortality risk prediction model, and output the in-hospital mortality prediction probability of the target patient; obtain the pre-trained in-hospital mortality risk prediction model by minimizing the loss function; The feature contribution rate corresponding to the in-hospital mortality prediction probability is calculated to generate the interpretation results of the mortality risk prediction for the target patient.

2. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 1, characterized in that: The feature importance evaluation of the features in the original feature dataset yields a score for each feature, including: A random forest model is used to evaluate the feature importance of features in the original feature dataset. The random forest model consists of multiple decision trees. Starting from the root node of each decision tree, training samples are extracted. At the non-leaf nodes of each decision tree, some features are selected from all features as candidate splitting features. The splitting threshold with the lowest Gini impurity after splitting is selected as the optimal splitting threshold. Node splitting is performed based on the optimal splitting threshold until there are no more splittable nodes in the decision tree. The decrease in Gini impurity caused by each feature in all decision trees is calculated. The feature importance is evaluated based on the decrease in Gini impurity of each feature to obtain the score of each feature.

3. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 2, characterized in that: The core feature dataset obtained by filtering based on scores includes: Using a random forest model as the base learner, in each round of feature elimination iteration, features are sorted from high to low scores, and a preset proportion of features are removed starting from the lowest score. The predictive performance of the in-hospital mortality risk prediction model under the current feature subset is evaluated using multi-fold cross-validation. The optimal feature combination is the subset of features that maximizes the average AUC value and whose AUC value, sensitivity, and specificity in each round of cross-validation meet the corresponding threshold requirements. In each round of feature elimination iteration, the feature with a repetition rate exceeding a preset repetition rate threshold in the optimal feature combination is taken as the core prediction feature, and the data corresponding to the core prediction feature constitutes the core feature dataset.

4. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 3, characterized in that: The core feature dataset obtained by filtering based on scores also includes: The following comprehensive performance objective function is constructed to comprehensively evaluate the predictive performance and stability of the feature subset. The optimal feature combination is determined by maximizing the comprehensive performance objective function: ， in, , , AUC, sensitivity, specificity of the model on the validation set under K-fold cross-validation, respectively; , , AUC, sensitivity, specificity, respectively, standard deviation, to measure the stability of the model between different folds; , , are weight coefficients of AUC mean, sensitivity mean, and specificity mean, respectively, is a preset weight coefficient, and .

5. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 1, characterized in that: The loss function is calculated using the following formula: ， in, For the weight vector, For bias terms, Represents the slack variable vector. For the i-th sample Slack variables, For penalty parameters, Let i be the weight of the i-th sample. For kernel function, For sample labels.

6. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 1, characterized in that: The calculation of the feature contribution degree corresponding to the in-hospital mortality prediction probability includes: weighting the marginal contribution of each feature in the prediction result of the in-hospital mortality risk prediction model under different feature combinations to obtain the corresponding feature contribution degree. The marginal contribution of a feature is used to characterize the change in the prediction result of the in-hospital mortality risk prediction model after the feature is added or removed.

7. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 1, characterized in that: Clinical data should include at least: age, history of stroke, myocardial ischemia, hypotension, preoperative blood glucose, preoperative blood lactate level, cardiopulmonary bypass time, aortic clamping time, circulatory arrest time, and total surgical time.

8. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 1, characterized in that: The method further includes: constructing an in-hospital mortality risk prediction model based on the support vector machine algorithm, and using a network search method to jointly optimize the penalty parameter and kernel function parameter in the support vector machine algorithm, and selecting the parameter combination corresponding to the maximum cross-validation average AUC value as the optimal model parameter.

9. The method for predicting in-hospital mortality risk in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning according to claim 1, characterized in that: The method also includes: jointly outputting the in-hospital mortality prediction probability and the feature contribution, so as to realize the simultaneous display of risk prediction and risk source.

10. A system for predicting the risk of in-hospital mortality in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning, implementing the method for predicting the risk of in-hospital mortality in patients undergoing acute type A aortic dissection surgery based on interpretable machine learning as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition module is used to collect clinical data from patients undergoing acute type A aortic dissection surgery and to construct a raw feature dataset. The feature filtering module is used to evaluate the feature importance of features in the original feature dataset to obtain a score for each feature, and then filter based on the score to obtain the core feature dataset. The risk prediction module is used to input the core feature dataset corresponding to the target patient into a pre-trained in-hospital mortality risk prediction model and output the in-hospital mortality prediction probability of the target patient. The in-hospital mortality risk prediction model is constructed based on the support vector machine algorithm and obtained by minimizing the loss function. The results interpretation module is used to calculate the feature contribution corresponding to the in-hospital mortality prediction probability and generate interpretation results for the mortality risk prediction of the target patient.

Images