Inplanatable ML method and system for predicting death risk of ICU stroke patient in hospital

Abstract

The invention provides an interpretable ML method and system for predicting the death risk of an ICU stroke patient in a hospital, and belongs to the technical field of medical care informatics and electric digital data processing. The ICU cerebral arterial thrombosis patient in-hospital death risk prediction method comprises the steps of S1, inputting variables, related to the in-hospital death risk of the ICU cerebral arterial thrombosis patient, of a current patient; s2, performing in-hospital death risk prediction on the current patient by adopting a pre-constructed optimal prediction model for the in-hospital death risk of the ICU cerebral arterial thrombosis patient to obtain an in-hospital death prediction probability; and S3, performing explanatory analysis on the optimal prediction model by using an SHAP method, and analyzing the importance of variables to a hospital death risk prediction result. According to the method, the hospital death risk of the ICU cerebral arterial thrombosis patient can be quickly and accurately predicted, and an explanatory analysis can be provided for the prediction result, so that a clinical decision is guided.

Description

Technical Field

[0001] The present invention relates to the technical fields of medical care informatics and electrical digital data processing, and in particular to a method and system for constructing an explainable machine learning model for predicting the risk of in-hospital death of ICU ischemic stroke patients. Background Art

[0002] Stroke has become the leading cause of death and disability among adults in my country, with five major characteristics: high morbidity, high disability rate, high recurrence rate, high mortality rate, and high economic burden. Among them, ischemic stroke is the most common type of stroke, accounting for about 70% of all strokes. Although the disability-adjusted life years (DALYs) of patients have been declining with the continuous updating of drugs and methods for the treatment of ischemic stroke, patients who need to enter the intensive care unit (ICU) for treatment still face a high risk of death. However, there is currently a lack of early specific detection indicators and means for the prognosis identification of severe ischemic stroke. Constructing a prediction model to accurately identify ischemic stroke patients with a high risk of in-hospital death, timely warning, and taking targeted prevention and treatment measures are of great significance to reducing the risk of death and improving the prognosis of patients.

[0003] With the popularization of modern hospital information systems and the further improvement of computing power, the integration of machine learning (ML) and the medical field has become increasingly close. Compared with traditional mortality risk scoring models, complex ML models usually have higher prediction accuracy because they can capture nonlinear relationships and interactions between variables. Although ML models have good predictive accuracy, most of them are regarded as "black boxes", that is, the models lack interpretable analysis results, so their performance and application are questioned. Simply using machine learning algorithms to build clinical prediction models can no longer meet the needs of disease prevention and medical decision-making in the current context of refined medical care. Summary of the invention

[0004] In order to overcome the above technical defects, the first aspect of the present invention provides an interpretable ML method for predicting the risk of in-hospital mortality of ICU stroke patients, which comprises:

[0005] Step S1: Enter the variables related to the risk of in-hospital mortality in patients with ischemic stroke in the ICU for the current patient;

[0006] Step S2: using the pre-constructed optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients to predict the in-hospital mortality risk of the current patient, and obtaining the predicted probability of in-hospital mortality;

[0007] Step S3: Use the SHAP method to perform an explanatory analysis on the optimal prediction model and analyze the importance of variables to the prediction results of in-hospital mortality risk; analyze the importance of variables to the prediction results from three aspects: weight analysis of each variable, dependency relationship between each variable and outcome, and individualized analysis of patients.

[0008] The variables associated with the risk of in-hospital death of ICU ischemic stroke patients are updated based on the training set data. For example, the variables associated with the risk of in-hospital death of ICU ischemic stroke patients include mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate and systolic blood pressure.

[0009] Furthermore, in step S2, the method for constructing the optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients includes:

[0010] Step S2.1: According to the inclusion and exclusion criteria, the clinical data and prognostic outcomes of ischemic stroke patients on the first day of ICU admission were retrospectively collected to construct an analysis data set.

[0011] Step S2.2: Data preprocessing of the analysis data set: ① Divide the variables into categorical variables and continuous variables; ② Use the multiple imputation method to fill in missing data; ③ Use the minimum and maximum value standardization method to normalize the continuous variables: and randomly divide the data set into a training set and a test set in a ratio of 7:3;

[0012] Step S2.3: In the training set, three algorithms, namely LASSO regression, Boruta algorithm and recursive feature elimination, are used to screen the variables for constructing the in-hospital death prediction model respectively, and the intersection of the variables screened by the three algorithms is taken as the predictive variable for constructing the in-hospital death prediction model for ICU ischemic stroke patients;

[0013] Step S2.4: constructing an ICU ischemic stroke in-hospital mortality risk prediction model using different machine learning methods on the training set after data preprocessing, the machine learning methods including logistic regression (LR), k-nearest neighbor (KNN), Naive Bayesian (NB), decision tree (DT), support vector machine (SVM), random forest (RF), XGBoost (extreme Gradient Boosting), LightGBM (Light Gradient Boosting Machine) and artificial neural network (ANN);

[0014] Step S2.5: Perform hyperparameter search on the model constructed in step S2.4 using a random grid search method with 5-fold cross validation repeated 50 times to obtain the optimal model hyperparameters;

[0015] Step S2.6: Using the area under the receiver operating characteristic curve as an indicator, the receiver operating characteristic (ROC) curve was drawn, and the accuracy, sensitivity, specificity and F1 value were selected to evaluate the discrimination performance of the model. The calibration curve was drawn and the Brill score of the model was calculated to indicate the calibration degree of the model. The decision curve analysis was used to reflect the clinical net benefit of the model.

[0016] Step S2.7: Taking all the above indicators into consideration, the model with the best comprehensive performance is finally selected as the in-hospital mortality risk prediction model for ICU ischemic stroke patients.

[0017] In step S2.1, the clinical data include general information, vital signs, laboratory test results, complications, drug treatment, clinical information on life support, and clinical scores.

[0018] Furthermore, in step S2.7, the random forest model was finally selected as the in-hospital mortality risk prediction model for ICU ischemic stroke patients.

[0019] Furthermore, the variables associated with the risk of in-hospital death in ICU ischemic stroke patients include mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate and systolic blood pressure; the model hyperparameters of the random forest model are n_estimators (maximum number of iterations of weak learners) of 500, max_depth (maximum depth of decision tree) of 9, min_samples_split (minimum number of samples required for internal node re-division) of 7, and min_samples_leaf (minimum number of samples for leaf nodes) of 4.

[0020] The second aspect of the present application provides an explainable ML system for predicting the risk of in-hospital mortality in ICU stroke patients, comprising:

[0021] A user interaction module, the user interaction module is used to input variables related to the in-hospital mortality risk of ICU ischemic stroke patients of the current patient; and is also used to output the predicted probability of in-hospital mortality and explanatory analysis results;

[0022] A risk prediction module, wherein the risk prediction module is used to predict the in-hospital death risk of the current patient using a pre-constructed optimal prediction model for the in-hospital death risk of ICU ischemic stroke patients, and obtain a predicted probability of in-hospital death;

[0023] An explanatory analysis module is used to perform an explanatory analysis on the optimal prediction model using the SHAP method to analyze the importance of variables to the prediction results of in-hospital mortality risk; the explanatory analysis module is used to analyze the importance of variables to the prediction results from three aspects: weight analysis of each variable, dependency relationship between each variable and outcome, and individualized analysis of patients.

[0024] The variables associated with the risk of in-hospital death of ICU ischemic stroke patients are updated based on the training set data. For example, the variables associated with the risk of in-hospital death of ICU ischemic stroke patients include mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate and systolic blood pressure.

[0025] Furthermore, in the risk prediction module, the method for constructing the optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients includes:

[0026] Step S2.1: According to the inclusion and exclusion criteria, the clinical data and prognostic outcomes of ischemic stroke patients on the first day of ICU admission were retrospectively collected to build an analysis data set;

[0027] Step S2.2: Data preprocessing of the analysis data set: ① Divide the variables into categorical variables and continuous variables; ② Use the multiple imputation method to fill in missing data; ③ Use the minimum and maximum value standardization method to normalize the continuous variables: and randomly divide the data set into a training set and a test set in a ratio of 7:3;

[0028] Step S2.3: In the training set, three algorithms, namely LASSO regression, Boruta algorithm and recursive feature elimination, are used to screen the variables for constructing the in-hospital death prediction model respectively, and the intersection of the variables screened by the three algorithms is taken as the predictive variable for constructing the in-hospital death prediction model for ICU ischemic stroke patients;

[0029] Step S2.4: constructing an ICU ischemic stroke in-hospital mortality risk prediction model using different machine learning methods on the training set after data preprocessing, wherein the machine learning methods include logistic regression, K nearest neighbor, naive Bayes, decision tree, support vector machine, random forest, XGBoost, LightGBM, and artificial neural network;

[0030] Step S2.5: Perform hyperparameter search on the model constructed in step S2.4 using a random grid search with 5-fold cross validation repeated 50 times to obtain the optimal parameter model;

[0031] Step S2.6: Using the area under the receiver operating characteristic curve as an indicator, the receiver operating characteristic curve was drawn, and the accuracy, sensitivity, specificity and F1 value were selected to evaluate the discrimination performance of the model. The calibration curve was drawn and the Brill score of the model was calculated to indicate the calibration degree of the model. The decision curve analysis was used to reflect the clinical net benefit of the model.

[0032] Step S2.7: Taking all the above indicators into consideration, the model with the best comprehensive performance is finally selected as the in-hospital mortality risk prediction model for ICU ischemic stroke patients.

[0033] In step S2.1, the clinical data include general information, vital signs, laboratory test results, complications, drug treatment, clinical information on life support, and clinical scores.

[0034] Furthermore, in step S2.7, the random forest model was finally selected as the in-hospital mortality risk prediction model for ICU ischemic stroke patients.

[0035] Furthermore, the variables associated with the risk of in-hospital death in ICU ischemic stroke patients include mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate and systolic blood pressure; the model hyperparameters of the random forest model are n_estimators (maximum number of iterations of weak learners) of 500, max_depth (maximum depth of decision tree) of 9, min_samples_split (minimum number of samples required for internal node re-division) of 7, and min_samples_leaf (minimum number of samples for leaf nodes) of 4.

[0036] Compared with the prior art, the above technical solution has the following beneficial effects:

[0037] The present application can quickly and accurately predict the in-hospital mortality risk of ICU ischemic stroke patients, and can also provide explanatory analysis of the prediction results. The present invention helps to improve the accurate and rapid assessment of the prognosis of ICU ischemic stroke patients, assist clinical decision-making, and improve patient prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flowchart of the method for constructing the optimal prediction model for the risk of in-hospital mortality in ICU ischemic stroke patients in this implementation;

[0039] Figure 2 A and B in the figure are LASSO algorithm variable screening diagrams;

[0040] Figure 3 Filter graph for Boruta algorithm variables;

[0041] Figure 4 Variable screening graph for recursive feature elimination method;

[0042] Figure 5 Receiver operating characteristic curves for each prediction model;

[0043] Figure 6 is the calibration curve of each prediction model;

[0044] Figure 7 is the decision curve of each prediction model;

[0045] Figure 8 It is the bee swarm diagram of SHAP value of each variable;

[0046] Fig. 9 It is the SHAP dependency graph of each variable;

[0047] Fig.10This is a local explanation diagram for a single sample. Because the SHAP value of the SBP variable here is too small, the impact on sample prediction is limited, and due to the size of the picture, it is impossible to display all variables. Only a few variables that are important for sample prediction are displayed, and the SBP variable is not shown in the figure.

[0048] Fig.11 Example screenshot of the user interface of the user interaction module of the explainable ML system for in-hospital mortality risk prediction for ICU stroke patients. DETAILED DESCRIPTION

[0049] The advantages of the present invention are further described below in conjunction with the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following specific description is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.

[0050] Example 1: Building an explainable ML system for predicting the risk of in-hospital mortality in ICU stroke patients

[0051] The method for constructing an explainable ML system for predicting the risk of in-hospital mortality of ICU stroke patients in this implementation includes the following steps:

[0052] Step 1: Construct an optimal prediction model for the risk of in-hospital mortality in patients with ischemic stroke in the ICU.

[0053] like Figure 1 As shown in Figure 2, the optimal prediction model for the risk of in-hospital death in ICU patients with ischemic stroke includes:

[0054] Step 1.1: According to the inclusion and exclusion criteria, the clinical data and prognosis outcomes (i.e., in-hospital mortality) of ischemic stroke patients on the first day of ICU admission were retrospectively collected to construct an analysis data set, and the data set was randomly divided into a training set and a test set in a ratio of 7:3.

[0055] When selecting patients with ischemic stroke in the ICU, patients must meet the following inclusion criteria and not include any of the exclusion criteria;

[0056] Inclusion criteria included: ① adult patients diagnosed with ischemic stroke; ② clear in-hospital death / survival outcomes.

[0057] Exclusion criteria included: ① ICU stay of less than 24 hours; ② only the data of the first ICU stay were used for multiple ICU stays; ③ incomplete clinical data.

[0058] The variables included included general information, vital signs, laboratory test results, complications, drug treatment, clinical information about life support, and clinical scores. General information included gender, age, weight, smoking history, and drinking history. The average values ​​of vital signs within 24 hours before admission to the ICU were used, including heart rate, diastolic blood pressure, systolic blood pressure, respiratory rate, body temperature, and blood oxygen saturation. For laboratory test results, the variables with the maximum values ​​within 24 hours before admission to the ICU were blood glucose, white blood cell count, blood urea nitrogen, serum creatinine, sodium ion, international normalized ratio, coagulation activation enzyme time, and partial thromboplastin time; the minimum values ​​within 24 hours before admission to the ICU were selected for hematocrit, platelets, anion gap, bicarbonate, calcium ion, chloride ion, and potassium ion. Related complications included hypertension, diabetes, hyperlipidemia, atrial fibrillation, cardiomyopathy, congestive heart failure, valvular heart disease, chronic obstructive pulmonary disease, and malignant tumors. Clinical information about life support, such as mechanical ventilation, renal replacement therapy, and mechanical thrombolysis. Drug treatment included aspirin, alteplase, warfarin, albumin, and statins. Clinical scores included Glasgow Coma Scale (GCS) and other data.

[0059] Then the dataset of ICU ischemic stroke patients collected above was randomly divided into training set and test set in a ratio of 7:3, and the missing data of the training set and test set were interpolated using the K-nearest neighbor algorithm, respectively.

[0060] Step 1.2: In the training set, three algorithms, namely LASSO regression, Boruta algorithm and recursive feature elimination, were used to screen the variables for constructing the in-hospital death prediction model. The intersection of the variables screened by the three algorithms was taken as the predictive variable for constructing the in-hospital death prediction model for ICU ischemic stroke patients.

[0061] In the training set, multiple methods were used to screen predictive variables using the included clinical data: LASSO regression ( Figure 2 )、Boruta algorithm( Figure 3 ) and recursive feature elimination ( Figure 4 ) three algorithms were used to screen the variables for constructing the in-hospital mortality prediction model respectively, and the intersection of the variables screened by the three algorithms was taken as the predictive variable for constructing the in-hospital mortality prediction model for ICU ischemic stroke patients.

[0062] The variables screened for the in-hospital mortality prediction model included mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate, and systolic blood pressure.

[0063] Step 1.3: In the training set, the final predictor variables screened out were preprocessed: ① The obtained variables were divided into categorical variables and continuous variables; ② The continuous variables were normalized using the minimum and maximum value standardization method; ③ In order to solve the problem that the imbalanced distribution of outcomes may cause the model to have bias in the classification prediction of the majority category, the training set was preprocessed using the Synthetic Minority Oversampling Technique (SMOTE).

[0064] Step 1.4: Different machine learning methods were used to construct the in-hospital mortality risk prediction model for ischemic stroke in ICU on the training set after data preprocessing. The machine learning methods included logistic regression (LR), K nearest neighbor (KNN), naive Bayes (NB), decision tree (DT), support vector machine (SVM), random forest (RF), XGBoost, LightGBM, and artificial neural network (ANN), respectively constructing the in-hospital mortality prediction model for ischemic stroke patients in ICU.

[0065] Step 1.5: Perform hyperparameter search on the model built in step 1.4 using a random grid search with 5-fold cross validation repeated 50 times to obtain the optimal model hyperparameters.

[0066] Step 1.6: In the test set, use the area under the receiver operating characteristic curve (AUC) as the main indicator to draw the receiver operating characteristic curve ( Figure 5 ), and selected accuracy, sensitivity, specificity and F1 value to evaluate the discrimination performance of the model. By drawing the calibration curve ( Figure 6 ), and the Brier Score of the model was calculated to indicate the degree of calibration of the model. The specific results are shown in Table 1. Decision Curve Analysis (DCA) was used to reflect the clinical net benefit of the model ( Figure 7 ).

[0067] Table 1 Prediction performance of each model on the test set

[0068]

[0069] Step 1.7: Taking all the above indicators into consideration, the model with the best overall performance (i.e., random forest algorithm (RF)) was finally selected to construct a risk prediction model for in-hospital mortality in ICU patients with ischemic stroke.

[0070] For example, the specific construction process of the optimal prediction model (i.e., random forest model) for the in-hospital mortality risk of ICU ischemic stroke patients in this embodiment includes:

[0071] (1) Data preprocessing

[0072] The data preprocessing process includes: 1.1) According to the inclusion and exclusion criteria, the clinical data and prognosis outcomes (i.e., in-hospital mortality rate) of ischemic stroke patients on the first day of admission to the ICU were retrospectively collected to build an analysis data set, and the data set was randomly divided into a training set and a test set in a ratio of 7:3. 1.2) In the training set, three algorithms, LASSO regression, Boruta algorithm, and recursive feature elimination, were used to screen the variables for constructing the in-hospital death prediction model, and the intersection of the variables screened by the three algorithms was taken as the predictive variable for constructing the in-hospital death prediction model for ICU ischemic stroke patients. The variables finally screened for the in-hospital death prediction model include: mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate, and systolic blood pressure. 1.3) In the training set, the final predicted variables screened out are processed: ① The obtained variables are divided into categorical variables and continuous variables; ② The continuous variables are normalized using the minimum and maximum value standardization method; ③ In order to solve the problem that the imbalanced distribution of outcomes may cause the model to have bias in the classification prediction of the majority category, the training set is preprocessed using the synthetic minority oversampling technique.

[0073] Next, the random forest algorithm was used to construct the optimal prediction model for the risk of in-hospital mortality of ischemic stroke in the ICU on the training set after data preprocessing.

[0074] (2) Bootstrap sampling: Multiple different data sets are formed by bootstrap sampling in the training set. Bootstrap sampling allows samples in the data set to be repeated, which means that the size of each data set is the same as the original training set, but contains some repeated samples.

[0075] (3) Feature selection: During the training process of each decision tree, a portion of the variables are randomly selected instead of considering all the variables. This step further increases the diversity of the model and helps reduce overfitting.

[0076] (4) Decision tree construction: Build a decision tree model based on the selected variables until each tree reaches its maximum depth or meets other stopping conditions.

[0077] (5) Prediction: For a new input sample, each tree will give a prediction result. A random grid search with 5-fold cross validation repeated 50 times was used to search for hyperparameters of the random forest model to obtain the optimal parameter model. The optimal model hyperparameters in this embodiment are: n_estimators (maximum number of iterations of weak learners) is 500, max_depth (maximum depth of decision tree) is 9, min_samples_split (minimum number of samples required for internal node re-division) is 7, and min_samples_leaf (minimum number of samples for leaf nodes) is 4.

[0078] (6) Result summary: For classification problems, the final prediction result is the majority vote of all decision tree prediction results.

[0079] Step 2: Use the SHAP method to perform interpretability analysis on the optimal prediction model.

[0080] The Shapley Additive Explanations (SHAP) method is used to evaluate the importance of the model's variables, ranking them according to their contribution to the prediction results, so as to gain a deeper understanding of which factors have a significant impact on the prediction. Specifically, from the analysis of the weight of each variable ( Figure 8 ), the dependence relationship between each variable and the outcome ( Fig. 9 )、Patient individualized analysis( Fig.10 ) Analyze the prediction results from three aspects and draw corresponding pictures to achieve the interpretability of the best machine learning model.

[0081] Step 3: Construct the optimal prediction model of step 1 and the interpretable method of step 2 into an in-hospital mortality risk prediction system for ICU ischemic stroke patients.

[0082] The in-hospital mortality risk prediction system for ICU ischemic stroke patients (for example, it can be in the form of an APP or a web page, etc.) includes: a user interaction module, a risk prediction module, and an explanatory analysis module. The risk prediction module is used to predict the in-hospital mortality risk of the current patient using a pre-built optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients, and obtain the predicted probability of in-hospital mortality. The explanatory analysis module is used to perform an explanatory analysis on the optimal prediction model using the SHAP method to analyze the importance of the variables to the in-hospital mortality risk prediction results. The user interaction module is used to input the variables of the current patient that are related to the in-hospital mortality risk of ICU ischemic stroke patients, and is also used to output the predicted probability of in-hospital mortality and the explanatory analysis results. The user interaction module provides user interaction functions in the form of an interactive user interface. The interactive user interface consists of three main parts, namely, a variable input area, a predicted probability output area, and a predicted result interpretation area ( Fig.11The variable input area in the sidebar allows users to enter information about the variables related to the patient's ICU admission. This interactive panel allows users to customize input based on the patient's variable values ​​and treatment plan. The predicted probability output area displays the predicted probability of in-hospital death after the user enters the specified variables. The prediction result interpretation area provides detailed information about the model itself and the prediction results, aiming to improve model transparency and allow users to gain in-depth understanding of how the patient's outcome prediction is generated.

[0083] Example 2 An interpretable ML method for predicting the risk of in-hospital mortality in ICU stroke patients

[0084] Based on the interpretable ML system for predicting the risk of in-hospital death of ICU stroke patients that has been constructed in Example 1, further, an interpretable ML method for predicting the risk of in-hospital death of ICU stroke patients using the system includes the following steps:

[0085] Step 1: Enter the current patient's variables associated with the risk of in-hospital mortality in patients with ischemic stroke in the ICU.

[0086] The user inputs variables related to the risk of in-hospital death of ICU ischemic stroke patients in the variable input area of ​​the user interaction interface of the user interaction module of the in-hospital death risk prediction system for ICU ischemic stroke patients based on interpretable machine learning, for example, including mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate and systolic blood pressure.

[0087] Step 2: Use the pre-constructed optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients to predict the in-hospital mortality risk of the current patient and obtain the predicted probability of in-hospital mortality.

[0088] The optimal prediction model is an in-hospital mortality risk prediction model for ischemic stroke patients in ICU constructed using the random forest algorithm (RF).

[0089] Step 3: Use the SHAP method to perform explanatory analysis on the optimal prediction model and analyze the importance of variables in the prediction results of in-hospital mortality risk.

[0090] The Shapley Additive Explanations (SHAP) method was used to evaluate the importance of the model's variables, and the variables were ranked according to their contribution to the prediction results, so as to gain a deeper understanding of which factors have a significant impact on the prediction. Specifically, the prediction results were analyzed and corresponding pictures were drawn from three aspects: the weight analysis of each variable, the dependency relationship between each variable and the outcome, and the individualized analysis of patients, to achieve the interpretability of the best machine learning model.

[0091] Step 4: Output the predicted probability of in-hospital death obtained in step 2 and the explanatory analysis results obtained in step 3.

[0092] The predicted probability output area of ​​the user interaction interface in the user interaction module of the in-hospital mortality risk prediction system for ICU ischemic stroke patients based on interpretable machine learning automatically outputs the predicted probability of in-hospital mortality after the user inputs the specified variables. The predicted result explanation area of ​​the user interaction interface in the user interaction module of the in-hospital mortality risk prediction system for ICU ischemic stroke patients based on interpretable machine learning automatically outputs detailed information about the model itself and the predicted results, aiming to improve the transparency of the model and allow users to have a deep understanding of how the patient's outcome prediction is generated. The Shapley Additive Explanations (SHAP) method is used to evaluate the importance of the model's variables, and the variables are ranked according to their contribution to the prediction results, so as to gain a deep understanding of which factors have an important impact on the prediction. Specifically, the prediction results are analyzed and the corresponding pictures are drawn from three aspects: the weight analysis of each variable, the dependency relationship between each variable and the outcome, and the individualized analysis of patients, to achieve the interpretability of the best machine learning model.

[0093] Step 5: Determine the corresponding risk level based on the predicted probability of in-hospital death and make corresponding clinical decisions.

[0094] In the present invention, the patient's risk of in-hospital death is divided into three categories: low risk, medium risk, and high risk according to the output probability of the model, and the corresponding predicted probabilities are less than 33.3%, between 33.3% and 66.6%, and greater than 66.6%, respectively. Low-risk patients should undergo routine monitoring, regularly check vital signs and changes in their condition, and use conventional treatment plans for treatment. Medium-risk patients should monitor vital signs and changes in their condition more frequently, and adjust the treatment plan in time according to the patient's specific situation. High-risk patients should closely and continuously monitor the patient's condition, and take more active medical interventions. For example, a patient is an 82.53-year-old female with chronic heart failure and coronary heart disease. She was admitted to the intensive care unit due to ischemic stroke ( Fig.10 ). The model predicts that the patient's probability of in-hospital death is 65%, which is a medium risk. The SHAP method further shows that factors that increase the patient's risk of death include not using statins, receiving mechanical ventilation, age of 82.53 years, bicarbonate of 19mEq / L, and not using warfarin; while protective factors are hematocrit 21.4%, white blood cell count 8.8 (109 / L), etc. Doctors are advised to treat modifiable factors, such as increasing the use of statins and warfarin.

[0095] It should be noted that the embodiments of the present invention have better practicability and do not impose any form of limitation on the present invention. Any technician familiar with the field may use the technical content disclosed above to change or modify it into an equivalent effective embodiment. However, any modification or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An interpretable ML method for predicting the risk of in-hospital mortality in ICU stroke patients, characterized by: include: Step S1: Enter the variables related to the risk of in-hospital mortality in patients with ischemic stroke in the ICU for the current patient; Step S2: using the pre-constructed optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients to predict the in-hospital mortality risk of the current patient, and obtaining the predicted probability of in-hospital mortality; Step S3: Use the SHAP method to perform an explanatory analysis on the optimal prediction model and analyze the importance of variables to the prediction results of in-hospital mortality risk; analyze the importance of variables to the prediction results from three aspects: weight analysis of each variable, dependency relationship between each variable and outcome, and individualized analysis of patients.

2. The interpretable ML method for predicting the risk of in-hospital mortality in ICU stroke patients as claimed in claim 1, characterized in that In step S2, the method for constructing the optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients includes: Step S2.1: According to the inclusion and exclusion criteria, the clinical data and prognostic outcomes of ischemic stroke patients on the first day of ICU admission were retrospectively collected to build an analysis data set; Step S2.2: Data preprocessing of the analysis data set: ① Divide the variables into categorical variables and continuous variables; ② Use the multiple imputation method to fill in missing data; ③ Use the minimum and maximum value standardization method to normalize the continuous variables: and randomly divide the data set into a training set and a test set in a ratio of 7:3; Step S2.3: In the training set, three algorithms, namely LASSO regression, Boruta algorithm and recursive feature elimination, are used to screen the variables for constructing the in-hospital death prediction model respectively, and the intersection of the variables screened by the three algorithms is taken as the predictive variable for constructing the in-hospital death prediction model for ICU ischemic stroke patients; Step S2.4: constructing an ICU ischemic stroke in-hospital mortality risk prediction model using different machine learning methods on the training set after data preprocessing, wherein the machine learning methods include logistic regression, K nearest neighbor, naive Bayes, decision tree, support vector machine, random forest, XGBoost, LightGBM, and artificial neural network; Step S2.5: Perform hyperparameter search on the model constructed in step S2.4 using a random grid search method with 5-fold cross validation repeated 50 times to obtain the optimal model hyperparameters; Step S2.6: Using the area under the receiver operating characteristic curve as an indicator, the receiver operating characteristic curve was drawn, and the accuracy, sensitivity, specificity and F1 value were selected to evaluate the discrimination performance of the model. The calibration curve was drawn and the Brill score of the model was calculated to indicate the calibration degree of the model. The decision curve analysis was used to reflect the clinical net benefit of the model. Step S2.7: Taking all the above indicators into consideration, the model with the best comprehensive performance is finally selected as the in-hospital mortality risk prediction model for ICU ischemic stroke patients.

3. The interpretable ML method for predicting the risk of in-hospital mortality in ICU stroke patients as claimed in claim 2, characterized in that: In step S2.7, the random forest model was finally selected as the risk prediction model for in-hospital mortality in ICU patients with ischemic stroke.

4. The interpretable ML method for predicting the risk of in-hospital mortality in ICU stroke patients as claimed in claim 3, characterized in that The variables associated with the risk of in-hospital death in ICU ischemic stroke patients include mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate and systolic blood pressure; the model hyperparameters of the random forest model are: the maximum number of iterations of the weak learner is 500, the maximum depth of the decision tree is 9, the minimum number of samples required for internal node re-division is 7, and the minimum number of samples of leaf nodes is 4.

5. An explainable ML system for predicting the risk of in-hospital mortality in ICU stroke patients, characterized by: include: A user interaction module, wherein the user interaction module is used to input variables of the current patient related to the risk of in-hospital death of ICU ischemic stroke patients; It is also used to output the predicted probability of in-hospital death and explanatory analysis results; A risk prediction module, wherein the risk prediction module is used to predict the in-hospital death risk of the current patient using a pre-constructed optimal prediction model for the in-hospital death risk of ICU ischemic stroke patients, and obtain a predicted probability of in-hospital death; An explanatory analysis module is used to perform an explanatory analysis on the optimal prediction model using the SHAP method to analyze the importance of variables to the prediction results of in-hospital mortality risk; the explanatory analysis module is used to analyze the importance of variables to the prediction results from three aspects: weight analysis of each variable, dependency relationship between each variable and outcome, and individualized analysis of patients.

6. The interpretable ML system for predicting the risk of in-hospital mortality in ICU stroke patients as claimed in claim 5, characterized in that In the risk prediction module, the method for constructing the optimal prediction model for the in-hospital mortality risk of ICU ischemic stroke patients includes: Step S2.1: According to the inclusion and exclusion criteria, the clinical data and prognostic outcomes of ischemic stroke patients on the first day of ICU admission were retrospectively collected to build an analysis data set; Step S2.2: Data preprocessing of the analysis data set: ① Divide the variables into categorical variables and continuous variables; ② Use the multiple imputation method to fill in missing data; ③ Use the minimum and maximum value standardization method to normalize the continuous variables: and randomly divide the data set into a training set and a test set in a ratio of 7:3; Step S2.3: In the training set, three algorithms, namely LASSO regression, Boruta algorithm and recursive feature elimination, are used to screen the variables for constructing the in-hospital death prediction model respectively, and the intersection of the variables screened by the three algorithms is taken as the predictive variable for constructing the in-hospital death prediction model for ICU ischemic stroke patients; Step S2.4: constructing an ICU ischemic stroke in-hospital mortality risk prediction model using different machine learning methods on the training set after data preprocessing, wherein the machine learning methods include logistic regression, K nearest neighbor, naive Bayes, decision tree, support vector machine, random forest, XGBoost, LightGBM, and artificial neural network; Step S2.5: Perform hyperparameter search on the model constructed in step S2.4 using a random grid search with 5-fold cross validation repeated 50 times to obtain the optimal model hyperparameters; Step S2.6: Using the area under the receiver operating characteristic curve as an indicator, the receiver operating characteristic curve was drawn, and the accuracy, sensitivity, specificity and F1 value were selected to evaluate the discrimination performance of the model. The calibration curve was drawn and the Brill score of the model was calculated to indicate the calibration degree of the model. The decision curve analysis was used to reflect the clinical net benefit of the model. Step S2.7: Taking all the above indicators into consideration, the model with the best comprehensive performance is finally selected as the in-hospital mortality risk prediction model for ICU ischemic stroke patients.

7. The explainable ML system for predicting the risk of in-hospital mortality in ICU stroke patients as claimed in claim 6, characterized in that In step S2.7, the random forest model was finally selected as the risk prediction model for in-hospital mortality in ICU patients with ischemic stroke.

8. The interpretable ML system for predicting the risk of in-hospital mortality in ICU stroke patients as claimed in claim 7, characterized in that The variables associated with the risk of in-hospital death in ICU ischemic stroke patients include mechanical ventilation, age, statins, white blood cell count, blood urea nitrogen, hematocrit, warfarin, bicarbonate and systolic blood pressure; the model hyperparameters of the random forest model are the maximum number of iterations of the weak learner is 500, the maximum depth of the decision tree is 9, the minimum number of samples required for internal node re-division is 7, and the minimum number of samples of leaf nodes is 4.

Images