Prone position ARDS patient death risk prediction system based on interpretable machine learning

Abstract

The invention discloses a prone position ARDS patient death risk prediction system based on interpretable machine learning, which constructs an interpretable machine learning prediction model based on clinical data of an ARDS patient before the ARDS patient receives prone position ventilation treatment, and is used for identifying high-death-risk crowds in an ICU (Intensive Care Unit). According to the method, an optimal prediction model is determined by performing preprocessing and correlation and collinearity screening on original variables, selecting key features by adopting Lasso regression and combining multi-model training and cross validation. On the basis, an SHAP method is introduced to explain a model decision basis, a column graph is constructed, visual expression of a risk assessment result is realized, and clinical understandability and practicability of the model are improved, so that auxiliary support is provided for precise treatment and resource allocation.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical monitoring and management assistance, specifically to a mortality risk prediction system for prone ARDS patients based on interpretable machine learning. Background Technology

[0002] Acute respiratory distress syndrome (ARDS) is a common and severe pulmonary dysfunction characterized by rapid disease progression and high mortality. For patients with moderate to severe ARDS, prone positioning ventilation (PPV) is commonly used clinically as a supportive measure to improve the ventilation / perfusion ratio and oxygenation. Although relevant studies support its clinical value, significant individual differences exist in the response to prone positioning ventilation during actual use. Therefore, risk assessment and dynamic stratification management for patients receiving prone positioning ventilation are key tasks in clinical management.

[0003] Currently, traditional methods for risk identification in ARDS patients largely rely on single-dimensional or rule-based indicators such as oxygenation index, lung compliance, and acute physiological scores (e.g., APACHE II). However, these methods still have shortcomings in variable interaction modeling, utilization of complex data, and prediction accuracy, making it difficult to effectively address challenges such as large individual differences and complex variables in critically ill patients. Therefore, there is an urgent need to introduce more efficient technologies to achieve multi-dimensional modeling of ARDS patient data and intelligent assessment of risk levels, in order to assist medical staff in making non-invasive decisions such as resource allocation and adjustment of monitoring intensity.

[0004] With the development of artificial intelligence, interpretable machine learning models, due to their ability to model nonlinear data and visualize the importance of variables, have been widely used in the construction of auxiliary decision-making systems in medical scenarios. By analyzing structured clinical data, relevant models can provide medical staff with data-driven reference information to support risk identification and individual management, ensuring the interpretability of decisions while enhancing the technical support capabilities for patient management.

[0005] In view of this, the present invention proposes a mortality risk prediction system for prone ARDS patients based on interpretable machine learning. Summary of the Invention

[0006] The purpose of this invention is to provide a mortality risk prediction system for prone ARDS patients based on interpretable machine learning, which solves the problems of inaccurate identification of high-risk groups, lack of interpretability and clinical decision support capabilities in existing prone ventilation therapy for ARDS patients.

[0007] In a first aspect, the present invention provides a mortality risk prediction system for prone ARDS patients based on interpretable machine learning, comprising a data acquisition module, a target variable acquisition module, a prediction model screening module, and a target prediction module, wherein the modules are connected via wired and / or wireless means.

[0008] The data acquisition module extracts standardized, multidimensional clinical data from the target patients.

[0009] The target variable acquisition module performs Lasso regular regression modeling on the screened multidimensional clinical data and uses cross-validation to select the optimal penalty parameter to obtain a set of target variables with non-zero coefficients.

[0010] The prediction model selection module uses the target variable set as input to construct logistic regression models, random forest models, support vector machine models, and extreme gradient boosting models; it completes training using a unified training set partitioning and cross-validation method, and calculates the AUC, accuracy, sensitivity, specificity, and F1-score of each model based on model performance metrics; and determines the best-performing model as the target prediction model.

[0011] The target prediction module, based on the target prediction model, uses the SHAP interpretation method to perform feature contribution analysis on the target variable set, obtains the ranked variable explanation weights, and constructs a nomogram tool for visual assessment of mortality risk.

[0012] As a preferred embodiment of the present invention, the step of obtaining the target variable set through Lasso regular regression modeling includes:

[0013] Construct the input feature matrix and label vector corresponding to the training dataset, and standardize the features;

[0014] Set the number of cross-validation folds and determine the optimal λ parameter using the minimum mean square error criterion;

[0015] The variables with non-zero coefficients corresponding to λ were extracted as the target variable set, denoted as age, C-reactive protein, international normalized ratio, immune diseases, viral infection, and whether or not the patient received CRRT treatment.

[0016] As a preferred embodiment of the present invention, the setting of the cross-validation fold number includes:

[0017] The training set is divided into five non-overlapping subsets, and each subset is used as the validation set in turn.

[0018] Calculate the verification error for each λ group, and select the λ value with the smallest verification error and the sparsest number of coefficients.

[0019] The final set of variables is obtained by sieving the coefficient structure corresponding to λ.

[0020] As a preferred embodiment of the present invention, the step of obtaining the target prediction model includes:

[0021] A unified strategy for dividing the training and test sets is established, with the training set accounting for 70% and the test set accounting for 30%.

[0022] Triple-fold cross-validation was performed on the logistic regression model, random forest model, support vector machine model, and extreme gradient boosting model, respectively.

[0023] Record the AUC, F1-score, accuracy, sensitivity, and specificity of each model on the test set;

[0024] The model is ranked according to AUC as the main indicator. If the AUCs are the same, the model with the higher F1-score is selected first. This process is repeated until the model with the best performance is determined as the target prediction model.

[0025] As a preferred technical solution of the present invention, the target prediction model is preferably a Logistic regression model.

[0026] As a preferred embodiment of the present invention, the step of performing feature contribution analysis using the SHAP interpretation method includes:

[0027] Construct a SHAP interpreter for interpreting the target model using the TreeExplainer method;

[0028] Calculate the feature SHAP value for each sample to obtain the single-sample explanatory distribution of all variables;

[0029] Calculate the average absolute value of the SHAP of each variable in the sample set as a measure of its global importance;

[0030] After sorting the variables by their contribution, output them and generate a bar chart for visualization.

[0031] As a preferred embodiment of the present invention, the calculation of the contribution of the global variable includes:

[0032] Iterate through each sample in the sample set and extract the absolute value of the corresponding variable in its SHAP value matrix;

[0033] The average of the absolute values ​​of the SHAP for each variable is taken as its overall contribution value;

[0034] The contribution ranking results are used to drive the construction order of the axis scales in the nomogram.

[0035] As a preferred embodiment of the present invention, the step of constructing a nomogram tool for mortality risk assessment includes:

[0036] Based on the regression coefficients of the optimal model, construct a scale axis for each target variable;

[0037] Arrange each scale axis in order of variable contribution, construct a total-to-score axis, and map it to the risk probability axis.

[0038] The total score and predicted mortality risk probability for any combination of patient variables are visualized.

[0039] Secondly, the present invention provides a method for predicting the mortality risk of prone ARDS patients based on interpretable machine learning, which, based on the implementation of the first aspect, includes the following steps:

[0040] Extract standardized, multidimensional clinical data from the target patients;

[0041] The selected multidimensional clinical data were modeled using Lasso regular regression, and cross-validation was used to select the optimal penalty parameter to obtain a set of target variables with non-zero coefficients.

[0042] The target variable set is used as input to construct logistic regression model, random forest model, support vector machine model and extreme gradient boosting model; training is completed using a unified training set partitioning and cross-validation method, and the AUC, accuracy, sensitivity, specificity and F1-score of each model are calculated based on model performance indicators; and the model with the best performance is determined as the target prediction model.

[0043] Based on the target prediction model, the SHAP interpretation method is used to perform feature contribution analysis on the target variable set, obtain the ranked variable explanation weights, and construct a nomogram tool for visual assessment of mortality risk.

[0044] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0045] This invention introduces interpretable machine learning methods to construct a model for predicting ICU mortality risk in ARDS patients based on pre-prone clinical variables, offering the following advantages: First, compared to traditional methods relying on a limited number of scoring indicators, this method automatically selects multidimensional highly correlated variables through Lasso regularization, significantly improving the specificity and stability of the prediction model. Second, the introduction of multi-model comparison and cross-validation strategies ensures the objectivity and reliability of model construction. Third, the use of the SHAP method to interpret the model results clarifies the relative importance of each key variable in risk prediction, enhancing the model's clinical interpretability. Finally, the visualization of complex modeling results through nomograms provides clinicians with an intuitive and personalized mortality risk assessment tool, facilitating early identification of high-risk patients and optimizing intervention pace and resource allocation. The overall solution combines accurate predictive capabilities with good clinical usability, demonstrating broad application prospects. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0047] Figure 1 This is a framework diagram of the prone ARDS patient mortality risk prediction system of the present invention;

[0048] Figure 2 This is a flowchart of the prone position ARDS patient mortality risk prediction method of the present invention;

[0049] Figure 3 This is the ROC curve of the logistic regression model of this invention;

[0050] Figure 4 This is the ROC curve of the XGBoost model of this invention;

[0051] Figure 5 This is the ROC curve of the support vector machine model of this invention;

[0052] Figure 6 This is a plot of the receiver operating characteristic (ROC) curve of the XGBoost model of this invention.

[0053] Figure 7 This is a comparison chart of the ROC curves of the four machine learning models of this invention;

[0054] Figure 8 This is a calibration curve of the logistic regression model of this invention;

[0055] Figure 9 This is a decision curve analysis diagram of the logistic regression model of this invention;

[0056] Figure 10 This is a nomogram for predicting the risk of death in ICU patients with ARDS according to the present invention; Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0058] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0059] Example 1

[0060] Please see Figure 1 As shown, this embodiment provides a mortality risk prediction system for prone ARDS patients based on interpretable machine learning, including a data acquisition module, a target variable acquisition module, a prediction model screening module, and a target prediction module. The modules are connected to each other via wired and / or wireless means.

[0061] The data acquisition module extracts standardized, multidimensional clinical data from the target patients.

[0062] Specifically, the logic for acquiring the multidimensional clinical data is as follows:

[0063] Collect multidimensional raw clinical data of patients with moderate to severe ARDS before prone ventilation treatment. The multidimensional raw clinical data includes basic patient information, physiological indicators, laboratory tests, underlying diseases and treatment records.

[0064] Preprocessing of the raw clinical data includes:

[0065] Variables with a missing proportion greater than a set threshold are deleted, and the remaining variables containing missing data are filled with missing values ​​using an imputation algorithm.

[0066] Standardize all variables to have a mean of 0 and a variance of 1; perform correlation and multicollinearity tests on all standardized variables, and remove variables with correlations higher than a preset threshold or variance inflation factors greater than a set threshold; for example:

[0067] Variables with a correlation greater than 0.7 were screened out based on the Pearson correlation coefficient;

[0068] Variables with a VIF greater than 10 were further eliminated using the variance inflation factor (VIF) to avoid collinearity interference.

[0069] It should be noted that: the independence between variables was ensured, providing stable input for subsequent feature selection and modeling. Based on interpretable machine learning technology, multidimensional clinical data of ARDS patients before and after receiving prone ventilation therapy were analyzed and mined.

[0070] The target variable acquisition module performs Lasso regular regression modeling on the screened multidimensional clinical data and uses cross-validation to select the optimal penalty parameter to obtain a set of target variables with non-zero coefficients.

[0071] It should be noted that, in order to identify features significantly associated with ICU mortality risk from numerous variables, this embodiment uses the Lasso (minimum absolute contraction and selection operator) regression method for variable screening. Integrating Lasso screening with a multi-model comparison mechanism addresses the issues of one-sided indicators and rigid models in traditional ICU scoring methods.

[0072] Specifically, the steps for obtaining the target variable set through Lasso regularized regression modeling include:

[0073] Construct the input feature matrix and label vector corresponding to the training dataset, and standardize the features;

[0074] Set the cross-validation fold number and determine the optimal λ parameter using the minimum mean square error criterion; select the simplest model corresponding to λ = 0.07079595 (lambda.1se);

[0075] The variables with non-zero coefficients corresponding to λ were extracted as the target variable set, denoted as age, C-reactive protein, international normalized ratio, immune diseases, viral infection, and whether or not the patient received CRRT treatment.

[0076] To further explain, the setting of the cross-validation fold number includes:

[0077] The training set is divided into five non-overlapping subsets, and each subset is used as the validation set in turn.

[0078] Calculate the verification error for each λ group, and select the λ value with the smallest verification error and the sparsest number of coefficients.

[0079] The final set of variables is obtained by sieving the coefficient structure corresponding to λ.

[0080] The prediction model selection module uses the target variable set as input to construct logistic regression models, random forest models, support vector machine models, and extreme gradient boosting models; it completes training using a unified training set partitioning and cross-validation method, and calculates the AUC, accuracy, sensitivity, specificity, and F1-score of each model based on model performance metrics; and determines the best-performing model as the target prediction model.

[0081] It should be noted that the target prediction model aims to accurately identify patients with high early mortality risk, thereby providing clinicians with more targeted treatment options and decision support, ultimately improving patient prognosis. Specifically, it uses data from before the prone position to develop a model that predicts the prognosis of ARDS patients using the prone position method, identifies patients with high mortality risk as early as possible, and screens patients who may benefit from prone ventilation, in order to guide treatment methods and improve patient outcomes.

[0082] Specifically, the steps for obtaining the target prediction model include:

[0083] A unified strategy for dividing the training and test sets is established, with the training set accounting for 70% and the test set accounting for 30%.

[0084] Triple-fold cross-validation was performed on the logistic regression model, random forest model, support vector machine model, and extreme gradient boosting model, respectively.

[0085] Record the AUC, F1-score, accuracy, sensitivity, and specificity of each model on the test set;

[0086] The model is ranked according to AUC as the main indicator. If the AUCs are the same, the model with the higher F1-score is selected first. This process is repeated until the model with the best performance is determined as the target prediction model.

[0087] like Figures 3 to 6 The figures shown are the ROC curves for each model; where:

[0088] Figure 3 This figure shows the receiver operating characteristic (ROC) curves for predicting the risk of ICU mortality in ARDS patients using a logistic regression model. The horizontal axis represents specificity, and the vertical axis represents sensitivity, which measures the balance between the model's ability to identify positive patients and its ability to control false positives. The shaded area in the figure represents the area under the curve (AUC), with a value of 0.821, indicating that the model has good discriminative performance. The values ​​corresponding to the marked points in the figure are the optimal cutoff points for the model, i.e., when the predicted probability is 0.205, the sensitivity is 0.667 and the specificity is 0.889, indicating that the model achieves a good balance of identification ability at this probability threshold.

[0089] Figure 4The receiver operating characteristic (ROC) curves of the mortality risk prediction model built based on the XGBoost algorithm on the validation set are shown, where the horizontal axis represents specificity and the vertical axis represents sensitivity. The shaded area in the figure is the area under the curve (AUC), with a value of 0.739, indicating that the model's overall predictive performance is good and it has a certain clinical auxiliary recognition capability. The marked points in the figure are the cutoff points where the model performs best at different thresholds. The corresponding probability threshold is 0.509, at which point the sensitivity is 0.792 and the specificity is 0.619, providing a reasonable balance of recognition performance parameters for the model in clinical applications.

[0090] Figure 5 The receiver operating characteristic (ROC) curves of the ARDS patient ICU mortality risk prediction model built using the Support Vector Machine (SVM) algorithm are shown on the test set. In the figure, the horizontal axis represents specificity, and the vertical axis represents sensitivity, used to evaluate the model's overall ability to identify high-risk patients. The area under the curve (AUC) of the model is 0.784, indicating that the model has good predictive ability. The point shown in the figure represents the optimal performance value of the model among various thresholds. At a threshold of 0.285, the sensitivity is 0.771 and the specificity is 0.762, which can be used as a recommended threshold for balancing sensitivity and specificity.

[0091] Figure 6 The receiver operating characteristic (ROC) curves of the ARDS patient ICU mortality risk prediction model built based on the XGBoost (Extreme Gradient Boosting) algorithm are shown on the test set. In the figure, the horizontal axis represents specificity, and the vertical axis represents sensitivity, used to evaluate the model's performance at different discrimination thresholds. The area under the curve (AUC) of this model is 0.739, indicating that it has moderately high predictive ability. The optimal cutoff point marked in the figure is 0.509, at which point the model's sensitivity is 0.792 and specificity is 0.619, which is one of the key parameters providing a reference for risk identification in real clinical scenarios.

[0092] like Figure 7 The curves show the overall performance comparison of the four models. Figure 7This figure compares the receiver operating characteristic (ROC) curves of Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost models in predicting the risk of ICU mortality in ARDS patients. In the figure, the horizontal axis represents specificity, the vertical axis represents sensitivity, and the area under each curve (AUC) represents the predictive accuracy of the corresponding model. The results show that the Logistic Regression model has the highest AUC (0.821), followed by the Support Vector Machine model (0.784), the XGBoost model (0.739), while the Random Forest model has relatively low predictive ability (0.678). This figure is used to comprehensively evaluate the discriminative performance of different models on real clinical data, assisting in the selection of the optimal model. From the performance of each indicator and the image results, it is clear that the Logistic Regression model has the best overall performance; therefore, it was selected as the final prediction model.

[0093] To further clarify, the preferred target prediction model is a Logistic regression model.

[0094] The target prediction module, based on the target prediction model, uses the SHAP interpretation method to perform feature contribution analysis on the target variable set, obtains the ranked variable explanation weights, and constructs a nomogram tool for visual assessment of mortality risk.

[0095] It should be noted that the introduction of the SHAP interpretation mechanism and the construction of line charts form an interpretable and quantifiable risk display closed loop, which is different from the traditional "black box" AI prediction.

[0096] Specifically, the steps for performing feature contribution analysis using the SHAP interpretation method include:

[0097] Construct a SHAP interpreter for interpreting the target model using the TreeExplainer method;

[0098] Calculate the feature SHAP value for each sample to obtain the single-sample explanatory distribution of all variables;

[0099] Calculate the average absolute value of the SHAP of each variable in the sample set as a measure of its global importance;

[0100] After sorting the variables by their contribution, output them and generate a bar chart for visualization.

[0101] The calculation of the contribution of global variables includes:

[0102] Iterate through each sample in the sample set and extract the absolute value of the corresponding variable in its SHAP value matrix;

[0103] The average of the absolute values ​​of the SHAP for each variable is taken as its overall contribution value;

[0104] The contribution ranking results are used to drive the construction order of the axis scales in the nomogram.

[0105] The steps for constructing a nomogram tool for mortality risk assessment include:

[0106] Based on the regression coefficients of the optimal model, construct a scale axis for each target variable;

[0107] Arrange each scale axis in order of variable contribution, construct a total-to-score axis, and map it to the risk probability axis.

[0108] The total score and predicted mortality risk probability for any combination of patient variables are visualized.

[0109] This embodiment is based on interpretable machine learning technology. By analyzing and mining clinical data of ARDS patients before and after receiving prone ventilation, a classification model for predicting mortality risk is constructed to identify high-risk patients as early as possible, thereby assisting clinicians in optimizing intervention decisions and resource allocation.

[0110] To improve the clinical interpretability and application visibility of the model, the Logistic model was further processed as follows:

[0111] Figure 8 This figure shows the calibration curve of a logistic regression model for predicting the risk of death in ICU patients with ARDS, used to assess the consistency between the model's predicted probabilities and the actual occurrence probabilities. The horizontal axis represents the predicted probability output by the model, and the vertical axis represents the actual observed probability. The dashed line in the figure is the apparent curve, reflecting the model's fit on the original training set; the solid line is the bias-corrected curve, obtained after 1000 repeated samplings using the bootstrap method, used to correct for overfitting bias. The two curves are close to a 45-degree diagonal, indicating that the model has good calibration ability, that is, the predicted risk value is relatively consistent with the actual incidence rate, supporting its use as a reliable risk assessment tool in clinical settings.

[0112] Figure 9The results of decision curve analysis of the constructed logistic regression model in the task of predicting the risk of death in the ICU of ARDS patients are presented, which are used to evaluate the net clinical benefit (Net Benefit) of the model at different risk thresholds. The horizontal axis represents the set high risk threshold, and the vertical axis represents the corresponding net benefit (Net Benefit). In the figure: the red curve (Model) represents the net benefit when intervention is carried out based on the predictive model; the gray curve (All) represents the extreme case assuming that all patients are intervened; the black curve (None) represents the other extreme case assuming that no patients receive intervention.

[0113] The figure shows that within the high-risk threshold range of approximately 0.01 to 0.30, the net benefit curve of the prediction model is consistently higher than that of the "All" and "None" strategies, indicating that the model has practical clinical decision-making value within this threshold range, which helps to improve patient screening efficiency and reduce unnecessary interventions.

[0114] Figure 10 A nomogram based on a logistic regression model is presented to comprehensively score the characteristics of ARDS patients before they receive prone ventilation and to estimate their probability of death in the ICU. The figure includes six clinical variables: C-reactive protein, viral infection status, age, whether the patient received CRRT treatment, international normalized ratio (INR), and whether the patient has an autoimmune disease.

[0115] The usage is as follows: Based on the value of each variable, find the corresponding position on its axis and read the corresponding score (Points); add up the scores of all variables to get the "Total Score"; then find the point corresponding to the total score below the "Totalpoints" axis, and map it down to the predicted mortality probability (Pr) axis at the bottom to obtain the individualized mortality risk prediction value for the patient. Figure 10 As part of the interpretable modeling results, it facilitates clinicians to quickly obtain risk estimates based on specific patient characteristics, which helps to support precise intervention and resource allocation.

[0116] In summary, this embodiment combines retrospective multivariate clinical data, uses Lasso regular regression to screen key variables, and constructs various machine learning models, ultimately selecting a mortality risk prediction model based on logistic regression. This model possesses good predictive performance, interpretability, and clinical applicability, providing strong support for early risk screening and precise intervention for ICU patients, and has clear potential for technology promotion.

[0117] Example 2

[0118] Please see Figure 2As shown, this embodiment provides a method for predicting the risk of ICU mortality in ARDS patients based on an interpretable machine learning model. It employs a combination of multi-model modeling and interpretability analysis to achieve high-accuracy prediction while improving the model's transparency and usability. The method includes the following steps:

[0119] Extract standardized, multidimensional clinical data from the target patients;

[0120] The selected multidimensional clinical data were modeled using Lasso regular regression, and cross-validation was used to select the optimal penalty parameter to obtain a set of target variables with non-zero coefficients.

[0121] The target variable set is used as input to construct logistic regression model, random forest model, support vector machine model and extreme gradient boosting model; training is completed using a unified training set partitioning and cross-validation method, and the AUC, accuracy, sensitivity, specificity and F1-score of each model are calculated based on model performance indicators; and the model with the best performance is determined as the target prediction model.

[0122] Based on the target prediction model, the SHAP interpretation method is used to perform feature contribution analysis on the target variable set, obtain the ranked variable explanation weights, and construct a nomogram tool for visual assessment of mortality risk.

[0123] In practice, a retrospective study was conducted on 207 ARDS patients admitted to a tertiary-level hospital between 2019 and 2024. Patients with moderate to severe ARDS who received prone ventilation were selected based on the inclusion criteria. Key variables were collected before prone ventilation, covering over 70 items including laboratory indicators, underlying diseases, and treatment methods.

[0124] For the raw data, a systematic preprocessing procedure is performed, including removing variables with a missing value rate exceeding 20% ​​and imputing missing values ​​for variables with a missing value rate less than 20% using multiple imputation methods to preserve data integrity as much as possible and reduce bias. Subsequently, based on correlation analysis and multicollinearity detection, highly correlated or collinear variables are removed to improve the stability of subsequent modeling.

[0125] To construct a high-quality predictive model, correlation and multicollinearity among variables were tested. Variables with a VIF > 10 were excluded, and variables with a correlation > 0.7 were further screened to ensure that there was no high degree of collinearity among variables in subsequent modeling, thereby improving the stability and interpretability of the model. In the variable screening stage, LASSO regression was used to screen clinical features to identify variables significantly associated with the risk of ICU mortality in ARDS patients. These variables were used as input variables for subsequent model training and optimization. Specifically, five-fold cross-validation (5-fold CV) was set in the LASSO regression to select the optimal penalty parameter λ (0.07079595; lambda.1se). In the final model, six variables had non-zero coefficients: age, C-reactive protein, international normalized ratio, immune diseases, viral infection, and whether CRRT was used. These variables maintained high feature importance in different models, demonstrating their crucial role in predicting the risk of ICU mortality in ARDS patients.

[0126] The screening results were used as input features for modeling, and four models—Logistic Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost Extreme Gradient Model—were constructed for training and prediction of ICU mortality outcomes in ARDS patients after prone ventilation. To select the optimal prediction model for each algorithm under different parameter combinations, this embodiment used three-fold cross-validation to tune and evaluate the performance of each model. Model performance evaluation metrics included AUC, accuracy, recall, sensitivity, specificity, and F1 score. Calibration curves were used to assess the model's calibration capability, and decision curve analysis was used to verify the model's application value in clinical practice. The model with the best overall performance was selected as the prediction model for this embodiment. Furthermore, the SHAP method was used to rank the importance of its features, revealing the relative contributions of each key variable in predicting the mortality risk of ARDS patients from an interpretability perspective, thus contributing to clinical understanding and the exploration of intervention mechanisms.

[0127] The predictive weights of each key variable on the risk of death in the ICU were visualized, and the importance of the features was further analyzed. The results showed that the main variables affecting the model's predictions were, in order: C-reactive protein, viral infection, age, whether CRRT was used, international normalized ratio, and immune diseases.

[0128] Based on six key variables selected using Lasso regression, different machine learning prediction models were successfully constructed, and their performance was compared. The results showed that the Logistic regression model outperformed other models in prediction, and C-reactive protein, viral infection, and age were important predictors of patient outcomes. By using Lasso regression to screen variables, six variables were ultimately selected for model construction: age, C-reactive protein, international normalized ratio, immune disease, viral infection, and whether CRRT was used. This provides a reference for the early identification and risk stratification of high-risk patients.

[0129] Based on the above variables, four prediction models were established: Logistic Regression, Random Forest, Support Vector Machine and XGBoost. The AUC of each model was calculated to evaluate its prediction performance, as shown in Table 1.

[0130]

[0131] like Figures 3 to 6 The figures show the receiver operating characteristic (ROC) curves for four models: Logistic Regression, XGBoost, SVM, and Random Forest. The comparison results show that the Logistic Regression model has the best predictive performance, with an AUC of 0.78. A nomogram was constructed based on this optimal model. The calibration curve obtained after 1000 repeated bootstrap sampling shows that the model's predicted curve is consistent with the ideal curve and is close to the diagonal, suggesting that the model's prediction of the mortality risk of ARDS patients receiving PPV is consistent with the actual mortality risk, indicating high predictive accuracy. In the decision curve graph, the "None" curve represents the net clinical benefit when no intervention is given to all ARDS patients; the "All" curve represents the net benefit when PPV is given to all ARDS patients. The red curve represents the net benefit of the logistic regression prediction model constructed based on the nomogram. The decision curve analysis results show that, within the high-risk threshold probability range of 0.01 to 0.30, the net clinical benefit of this model is consistently higher than that of the two extreme strategies, "All" and "None," suggesting that the model can provide stable and superior clinical decision-making value.

[0132] like Figure 7 As shown, this is a comparison of the ROC curve fitting of the four models mentioned above. The purple curve represents logistic regression, which is generally above the other models, further validating its advantages. Regarding model interpretability, such as... Figure 8 The figure shows the calibration curve of the logistic regression model. The curve's trend is close to the diagonal, indicating a good match between the model's predicted probabilities and the actual observed probabilities. Figure 9The diagram shows the decision curve analysis (DCA). Within the high-risk threshold range of 0.01 to 0.3, the net return of this model is higher than that of the two extreme strategies, "All" and "None," demonstrating good clinical decision support value. To improve the convenience and transparency of the model in practical applications, a further model based on logistic regression was constructed, as shown below. Figure 10 The nomogram shown is a graph where doctors can locate corresponding points based on the patient's specific values. By accumulating scores and mapping them to the risk axis below, individualized ICU mortality risk can be derived, providing quantitative evidence for the early identification of high-risk patients.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A mortality risk prediction system for prone ARDS patients based on interpretable machine learning, characterized in that, It includes a data acquisition module, a target variable acquisition module, a prediction model selection module, and a target prediction module, with each module connected via wired and / or wireless means. The data acquisition module extracts standardized, multidimensional clinical data from the target patients. The target variable acquisition module performs Lasso regular regression modeling on the screened multidimensional clinical data and uses cross-validation to select the optimal penalty parameter to obtain a set of target variables with non-zero coefficients. The prediction model selection module uses the target variable set as input to construct logistic regression models, random forest models, support vector machine models, and extreme gradient boosting models; it completes training using a unified training set partitioning and cross-validation method, and calculates the AUC, accuracy, sensitivity, specificity, and F1-score results for each model based on model performance metrics; The model with the best performance was then selected as the target prediction model. The target prediction module, based on the target prediction model, uses the SHAP interpretation method to perform feature contribution analysis on the target variable set, obtains the ranked variable explanation weights, and constructs a nomogram tool for visual assessment of mortality risk.

2. The prone ARDS patient mortality risk prediction system based on interpretable machine learning according to claim 1, characterized in that, The steps for obtaining the target variable set through Lasso regular regression modeling include: Construct the input feature matrix and label vector corresponding to the training dataset, and standardize the features; Set the number of cross-validation folds and determine the optimal λ parameter using the minimum mean square error criterion; The variables with non-zero coefficients corresponding to λ were extracted as the target variable set, denoted as age, C-reactive protein, international normalized ratio, immune diseases, viral infection, and whether or not the patient received CRRT treatment.

3. The prone ARDS patient mortality risk prediction system based on interpretable machine learning according to claim 2, characterized in that, The number of cross-validation folds is defined as follows: The training set is divided into five non-overlapping subsets, and each subset is used as the validation set in turn. Calculate the verification error for each λ group, and select the λ value with the smallest verification error and the sparsest number of coefficients. The final set of variables is obtained by sieving the coefficient structure corresponding to λ.

4. The prone ARDS patient mortality risk prediction system based on interpretable machine learning according to claim 1, characterized in that, The steps for obtaining the target prediction model include: A unified strategy for dividing the training and test sets is established, with the training set accounting for 70% and the test set accounting for 30%. Tri-fold cross-validation was performed on the logistic regression model, random forest model, support vector machine model, and extreme gradient boosting model, respectively. Record the AUC, F1-score, accuracy, sensitivity, and specificity of each model on the test set; The model is ranked according to AUC as the main indicator. If the AUCs are the same, the model with the higher F1-score is selected first. This process is repeated until the model with the best performance is determined as the target prediction model.

5. The prone ARDS patient mortality risk prediction system based on interpretable machine learning according to claim 4, characterized in that, The preferred prediction model for the target is a Logistic regression model.

6. The prone ARDS patient mortality risk prediction system based on interpretable machine learning according to claim 1, characterized in that, The steps for feature contribution analysis using the SHAP interpretation method include: Construct a SHAP interpreter for interpreting the target model using the TreeExplainer method; Calculate the feature SHAP value for each sample to obtain the single-sample explanatory distribution of all variables; Calculate the average absolute value of the SHAP of each variable in the sample set as a measure of its global importance; After sorting the variables by their contribution, output them and generate a bar chart for visualization.

7. The prone ARDS patient mortality risk prediction system based on interpretable machine learning according to claim 6, characterized in that, The calculation of the contribution of global variables includes: Iterate through each sample in the sample set and extract the absolute value of the corresponding variable in its SHAP value matrix; The average of the absolute values ​​of the SHAP for each variable is taken as its overall contribution value; The contribution ranking results are used to drive the construction order of the axis scales in the nomogram.

8. The prone ARDS patient mortality risk prediction system based on interpretable machine learning according to claim 1, characterized in that, The steps for constructing a nomogram tool for mortality risk assessment include: Based on the regression coefficients of the optimal model, construct a scale axis for each target variable; Arrange each scale axis in order of variable contribution, construct a total-to-score axis, and map it to the risk probability axis. The total score and predicted mortality risk probability for any combination of patient variables are visualized.

9. A method for predicting mortality risk in prone ARDS patients based on interpretable machine learning, comprising the implementation of the prone ARDS patient mortality risk prediction system based on any one of claims 1-8, characterized in that, Includes the following steps: Extract standardized, multidimensional clinical data from the target patients; The selected multidimensional clinical data were modeled using Lasso regular regression, and cross-validation was used to select the optimal penalty parameter to obtain a set of target variables with non-zero coefficients. The target variable set is used as input to construct logistic regression model, random forest model, support vector machine model and extreme gradient boosting model; training is completed using a unified training set partitioning and cross-validation method, and the AUC, accuracy, sensitivity, specificity and F1-score of each model are calculated based on the model performance indicators; The model with the best performance was then selected as the target prediction model. Based on the target prediction model, the SHAP interpretation method is used to perform feature contribution analysis on the target variable set, obtain the ranked variable explanation weights, and construct a nomogram tool for visual assessment of mortality risk.

Images