Method for constructing a neural function prognosis prediction model after cardiopulmonary resuscitation and application thereof

Abstract

The application discloses a construction method and application of a post-cardiopulmonary resuscitation neural function prognosis prediction model, and a method, which comprises the following steps: acquiring clinical data of patients who have achieved spontaneous circulation recovery after cardiopulmonary resuscitation, and screening clinical patients meeting the inclusion criteria; taking brain function classification as a dependent variable, respectively adopting LASSO regression analysis and Boruta feature screening to screen characteristic variables, and selecting common characteristic variables; based on the common characteristic variables, constructing multiple machine learning models, and drawing a ROC curve and a decision curve corresponding to each machine learning model; adopting AUC to evaluate the discrimination of the machine learning models, adopting the decision curve to evaluate the clinical benefits of the machine learning models, and screening an optimal prediction model; and adopting SHAP tools to explain the optimal prediction model and the common characteristic variables. The prediction model constructed by the application can conveniently and quickly predict the neural function prognosis within 24 hours after cardiopulmonary resuscitation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of prediction model construction, in particular to a construction method of a neural function prognosis prediction model after cardiopulmonary resuscitation and application thereof. BACKGROUND

[0002] Cardiac arrest (CA) is a leading cause of death and severe neurological disability worldwide. After cardiopulmonary resuscitation (CPR) to achieve resumption of spontaneous circulation (ROSC), 45%~70% of surviving CA patients develop hypoxic ischemic encephalopathy, showing severe neurological impairment or even death. According to the results of a 2023 epidemiological survey on cardiac arrest in China, the survival to hospital discharge / survival rate at 30 days of out-of-hospital cardiac arrest (OHCA) is 1.2%, and the rate of good neurological prognosis is only 0.8%; the survival to hospital discharge / survival rate at 30 days of in-hospital cardiac arrest (IHCA) is 9.4%, and the rate of good neurological prognosis is only 6.7%.

[0003] Early and accurate prediction of the neurological prognosis of CA patients is crucial for guiding clinical decision-making and optimizing resource allocation. Currently available clinical assessment tools, such as the modified Rankin (mRS) scale, the good outcome for cardiac arrest (GO-FAR), the cardiac arrest survival prediction instrument (CASPRI), and the cardiac arrest outcome score (CAOS), have limitations in accuracy, specificity, and timeliness, with an evaluation time window generally within 72 hours or 1 week after ROSC or even at the time of hospital discharge.

[0004] Cardiac arrest survival prediction instrument (CASPRI), cardiac arrest outcome score (CAOS), etc. The evaluation time window is generally within 72 hours or 1 week after ROSC or even at the time of hospital discharge, with limitations in accuracy, specificity, and timeliness.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The technical task of the present application is to provide a method for constructing a neural function prognosis prediction model after cardiopulmonary resuscitation and an application thereof, the prediction model constructed by the present application can conveniently and quickly predict the neural function prognosis within 24 hours after cardiac arrest resuscitation, provide a new strategy for early identification of adverse neural function prognosis and precise intervention, thereby reducing the incidence of adverse neural function and improving the prognosis of CA patients.

[0007] To achieve the above object, the present application provides the following technical scheme:

[0008] According to one aspect of the present application, a method for constructing a neural function prognosis prediction model after cardiopulmonary resuscitation is provided, comprising: obtaining clinical data of patients who have achieved spontaneous circulation recovery after cardiopulmonary resuscitation, and screening clinical patients who meet the inclusion criteria; taking brain function classification as the dependent variable, using LASSO regression analysis and Boruta feature screening method to screen feature variables respectively, and selecting common feature variables; based on the common feature variables, constructing multiple machine learning models, and drawing the ROC curve and decision curve corresponding to each machine learning model; using the area under the receiver operating characteristic curve to evaluate the discrimination of the machine learning model, using the decision curve to evaluate the clinical benefit of the machine learning model, and screening the optimal prediction model; using SHAP tool to explain the optimal prediction model and the common feature variables.

[0009] In some embodiments, the SHAP tool is used to perform explanatory analysis on the optimal prediction model, comprising: calculating the SHAP value of each common feature variable, ranking the importance of the common feature variables according to the SHAP value, and drawing a dependency contribution graph.

[0010] In some embodiments, the inclusion criteria include: inclusion criteria of age ≥ 18 years old, hospitalization time > 24 hours after successful resuscitation; exclusion criteria of having a history of neurological function system related diseases, and more than 20% of missing clinical data.

[0011] In some embodiments, the common feature variables are age, non-perfusion time, ROSC time, initial defibrillation rhythm, neuron-specific enolase, S100 calcium binding protein β, lactic acid, abnormal electroencephalogram, activated partial thromboplastin time, and gray matter / white matter ratio of head CT.

[0012] In some embodiments, the machine learning models are extreme gradient boosting model, support vector machine model, random forest model, K nearest neighbor algorithm model, decision tree model and logistic regression model.

[0013] In some embodiments, the optimal prediction model is an extreme gradient boosting model.

[0014] According to another aspect of the present application, the application also provides an application of the extreme gradient boosting prediction model constructed by the method in the prediction of neurological function prognosis after cardiopulmonary resuscitation.

[0015] In some embodiments, the learning rate of the extreme gradient boosting prediction model is 0.03, the global bias is -1.782, the number of trees is 300, and the maximum depth of the tree is 4.

[0016] In some embodiments, the prediction method of the application comprises: defining a feature vector of common characteristic variables of the patient; inputting the common characteristic variables into the extreme gradient boosting prediction model; calculating the probability of neurological adverse prognosis; comparing the calculated probability with an adverse prognosis threshold, if the probability of neurological adverse prognosis is greater than or equal to the adverse prognosis threshold, outputting a high risk of neurological adverse prognosis; if the probability of neurological adverse prognosis is less than the adverse prognosis threshold, outputting a low risk of neurological adverse prognosis.

[0017] In some embodiments, the adverse prognosis threshold is 0.42.

[0018] Compared with the prior art, the application has the advantages and positive effects that the prediction model constructed by the application can conveniently and quickly predict the neurological function prognosis within 24 hours after cardiopulmonary resuscitation, provides a new strategy for early identification of adverse neurological function prognosis and precise intervention, thereby reducing the incidence of adverse neurological function and improving the prognosis of CA patients.

[0019] Further, by using LASSO regression analysis and Boruta feature screening method to screen the characteristic variables and selecting common characteristic variables, the influence of screening characteristic variables by a single method can be reduced, and the accuracy of the common characteristic variables can be improved; by constructing six machine learning models, the optimal prediction model can be selected, and by using SHAP tool to explain the optimal prediction model and the common characteristic variables, the relationship between the common characteristic variables and the neurological function prognosis after cardiopulmonary resuscitation can be visually explained. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 A flowchart of the method in the embodiments of the present application is shown.

[0022] Figure 2 The characteristic variables screened by LASSO regression analysis in the embodiments of the present application are shown.

[0023] Figure 3 Feature variables of Boruta feature selection in the embodiment of the application are shown.

[0024] Figure 4 The intersection graph of feature variables of LASSO and Boruta feature selection in the embodiment of the application is shown.

[0025] Figure 5 The receiver operating characteristic curves of the six machine learning models in the embodiment of the application are shown.

[0026] Figure 6 The performance index heat map of the six machine learning models in the embodiment of the application is shown.

[0027] Figure 7 The decision curve of the six machine learning models in the embodiment of the application is shown.

[0028] Figure 8 The ROC curve comparison of the XGBoost prediction model and the GCS score and the mRS score in the embodiment of the application is shown.

[0029] Figure 9 The importance matrix graph of the feature variables in the development of the XGBoost model in the embodiment of the application is shown.

[0030] Figure 10 The SHAP visualization example in the embodiment of the application is shown.

[0031] Figure 11 The ROC curve of the six machine learning models in the embodiment of the application is shown.

[0032] Figure 12 The performance index heat map of the six machine learning models in the embodiment of the application is shown.

[0033] Figure 13 The decision curve of the six machine learning models in the embodiment of the application is shown. DETAILED DESCRIPTION

[0034] In order to enable the above-mentioned purposes, features and advantages of the application to be more clearly understood, the application will be further described below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments of the application and the features in the embodiments can be combined with each other without conflict.

[0035] The application will be further described below with reference to the accompanying drawings and specific embodiments. Figure 1 The method flowchart in the embodiment of the application is shown.

[0036] S100. Obtain clinical data of patients who have achieved spontaneous circulation after cardiopulmonary resuscitation and screen clinical patients who meet the exclusion criteria.

[0037] Clinical data of patients who achieved ROSC after CPR and were admitted to a hospital in Tianjin from January 2014 to December 2023 were retrospectively collected. Inclusion criteria: (1) age ≥ 18 years; (2) hospital stay > 24 hours after successful resuscitation. Exclusion criteria: (1) history of neurological disorders; (2) missing clinical data > 20%. Patients were randomly assigned to the training group and the validation group at a ratio of 7:3 (randomization seed set to 123). According to the Glasgow-Pittsburgh Brain Function Performance Scale (CPC), patients were divided into a good neurological function group (CPC 1-2) and a poor neurological function group (CPC 3-5) at discharge. Variables with missing data > 20% were removed; variables with missing data ≤ 20% were supplemented, with normal distributions supplemented by the mean and skewed distributions supplemented by the median.

[0038] Ultimately, 44 variables were selected, including 42 independent variables and 2 rating variables, as follows:

[0039] (1) Demographic data: age, sex, body mass index (BMI);

[0040] (2) Basic diseases and family history: coronary heart disease, hypertension, diabetes, smoking history, drinking history, family history of CA;

[0041] (3) Cardiopulmonary resuscitation related information: no perfusion time (i.e., the time from CA to CPR), initial defibrillation rhythm, mechanical ventilation, adrenaline dose, ROSC time;

[0042] (4) Laboratory test data: white blood cell count, hemoglobin, platelet count, alanine aminotransferase (ALT), aspartate aminotransferase (AST), albumin, globulin, total cholesterol, triglycerides, serum creatinine, blood urea nitrogen, serum sodium, serum potassium, serum calcium, blood glucose, glycated hemoglobin, lactate, troponin, brain natriuretic peptide (BNP), prothrombin time (PT), activated partial thromboplastin time (APTT), D-dimer, C-reactive protein, Acute Physiological and Chronic Health Assessment (APACHE-II score), head CT gray matter / white matter ratio (GWR), S100 calcium-binding protein β (S100β), neuron-specific enolase (NSE), abnormal electroencephalogram (EEG).

[0043] (5) Scoring variables: Glasgow Coma Scale (GCS) was assessed 72 hours after ROSC, and the modified Rankin Scale (mRS) was assessed at discharge.

[0044] SPSS 26.0 and R (version 4.3.2) software were used for data statistics and analysis. Normally distributed quantitative data were analyzed using... Data are expressed in mean ± standard deviation (±s) form. Comparisons between two groups were performed using the independent samples t-test. Non-normally distributed continuous data are expressed as median (interquartiles) [M(Q1, Q3)], and comparisons between two groups were performed using the Mann-Whitney U test. Count data are expressed as number of cases and / or percentages, and comparisons between two groups were performed using the χ² test. P < 0.05 was considered statistically significant.

[0045] This study included 519 patients, of whom 102 had a good neurological outcome (19.6%). Patients were randomly assigned to a training group (n=363) and a validation group (n=156) at a ratio of 7:3. There were no significant differences in baseline characteristics between the training and validation groups (p>0.05), making them comparable (see Table 1).

[0046] Table 1. Comparison of baseline characteristics between the training and validation groups

[0047]

[0048] Note: BMI is Body Mass Index, CA is Cardiac Arrest, ROSC is Recovery of Spontaneous Circulation, ALT is Alanine Aminotransferase, AST is Aspartate Aminotransferase, BNP is B-type Brain Natriuretic Peptide, PT is Prothrombin Time, APTT is Activated Partial Thromboplastin Time, APACHE-II score is Acute Physiology and Chronic Health Assessment, GWR is Gray / White Matter Ratio on Head CT, S100β is S100 Calcium-Binding Protein β, NSE is Neuron-Specific Enolase, EEG is Electroencephalogram, GCS is Glasgow Coma Scale, and mRS is Modified Rankin Scale.

[0049] S200. Using brain function classification as the dependent variable, LASSO regression analysis and Boruta feature selection were used to screen feature variables, and common feature variables were selected.

[0050] Using brain function classification (CPC) grouping as the dependent variable, the least absolute contraction and selection operator (LASSO) regression screening method was first used to analyze 44 independent variables. Figure 2 The following are the characteristic variables selected by LASSO regression analysis in the embodiments of this application, wherein, Figure 2 In the figure, A is the estimated curve of the LASSO predictor variable (the left vertical line represents the position of the optimal lambda, and the right vertical line corresponds to a distance of one standard error from the optimal lambda). Figure 2 In this context, B represents the identification of the optimal Lambda penalty estimate. For example...Figure 2 As shown in the figure, 16 characteristic variables were screened out: age, ROSC time, non-perfusion time, initial defibrillation rhythm, mechanical ventilation, adrenaline dosage, NSE, S100β, lactic acid, abnormal EEG, PT, APTT, serum creatinine, hemoglobin, GWR, and troponin.

[0051] Then, Boruta feature screening method was used for analysis. Figure 3 The characteristic variables of Boruta feature screening in the embodiment of the application are shown (green boxes are selected variables, yellow boxes are suspicious variables, and red boxes are excluded variables). As shown in the figure, Figure 3 As shown in the figure, 14 characteristic variables were screened out: non-perfusion time, S100β, BNP, abnormal EEG, NSE, ROSC time, age, GWR, lactic acid, APTT, APACHE-II score, D-dimer, platelet count, and initial defibrillation rhythm (see Figure 3 ).

[0052] Finally, two groups of the same 10 variables were selected as characteristic variables. Figure 4 The intersection diagram of LASSO and Boruta feature screening characteristic variables in the embodiment of the application is shown, in which the 10 characteristic variables are age, non-perfusion time, ROSC time, initial defibrillation rhythm, NSE, S100β, lactic acid, abnormal EEG, APTT, and GWR.

[0053] S300, based on the 10 common characteristic variables, 6 kinds of machine learning models were constructed, and the ROC curve and decision curve corresponding to each machine learning model were drawn.

[0054] Based on the 10 characteristic variables screened in step S200, 6 kinds of machine learning (ML) models were constructed in the training group and the validation group to predict the neurological functional prognosis of cardiac arrest (CA) patients.

[0055] S400, the area under the receiver operating characteristic curve (AUC) was used to evaluate the discrimination of the machine learning model, the decision curve was used to evaluate the clinical benefit of the machine learning model, and the optimal prediction model was screened.

[0056] Figure 5 The receiver operating characteristic curve (ROC curve) of the 6 kinds of machine learning models in the embodiment of the application is shown, in which Figure 5 A in the figure is the training group, Figure 5 B in the figure is the validation group. As shown in the figure, Figure 5As shown in FIG. 6, in the training group, the area under the receiver operating characteristic curve (AUC) of the extreme gradient boosting (XGBoost) model was the highest (AUC = 0.983), followed by the support vector machine (SVM) model (AUC = 0.947), the random forest (RF) model (AUC = 0.943), the K-nearest neighbor algorithm (KNN) model (AUC = 0.912), the decision tree (DT) model (AUC = 0.908), and the logistic regression (LR) model (AUC = 0.906). In the validation group, the AUC of the extreme gradient boosting (XGBoost) model was the highest (AUC = 0.939), followed by the support vector machine (SVM) model (AUC = 0.934), the logistic regression (LR) model (AUC = 0.925), the random forest (RF) model (AUC = 0.862), the K-nearest neighbor algorithm (KNN) model (AUC = 0.858), and the decision tree (DT) model (AUC = 0.829).

[0057] Figure 6 FIG. 7 shows the performance index heat map of the six machine learning models in the embodiments of the present application, wherein Figure 6 A in FIG. 7 is the training group, Figure 6 B in FIG. 7 is the validation group. As shown in FIG. 7, Figure 6 the evaluation index heat map of the model shows that the specificity (0.865) and sensitivity (0.892) of the LR model in the training group are the highest (A in FIG. 7); Figure 6 the sensitivity (0.942) and F1 score (0.953) of the DT model in the validation group are the highest (B in FIG. 7). The XGBoost model has the highest precision in the training group, with a value of 0.957 (A in FIG. 7); and in the validation group, it has a value of 0.985 (B in FIG. 7). Figure 6 Figure 6 Figure 6

[0058] Figure 7 FIG. 8 shows the decision curve of the six machine learning models in the embodiments of the present application, wherein, Figure 7 A in FIG. 8 is the training group, Figure 7 B in FIG. 8 is the validation group, ALL is all intervention; None is no intervention at all. The decision curve (DCA) is used to evaluate the clinical benefit of the prediction model. As shown in FIG. 8, Figure 7 the threshold range of the XGBoost model in the training group is about 0-0.95, which is significantly wider than the applicable range of other models. In addition, the DCA partially overlaps with other models in most threshold ranges, and the difference in net benefit is not obvious. However, in the threshold range of 0.8-0.95, the net benefit of the XGBoost model is higher than that of other models. In the validation group, the XGBoost model also shows better clinical applicability. After analyzing the above indexes comprehensively, the overall prediction performance of the XGBoost model is the best.​​​

[0059] Figure 8 This application compares the ROC curves of the XGBoost prediction model with those of the GCS score and mRS score in the embodiments. For example... Figure 8 As shown, the AUC of the traditional neurological function assessment tools, GCS score, was 0.869 (95% CI: 0.823-0.915); the AUC of the mRS score was 0.912 (95% CI: 0.877-0.947). The AUC of the XGBoost prediction model was 0.983 (95% CI: 0.969-0.997). Compared with the GCS score, the Delong test value Z=4.85, P<0.05, indicating a statistically significant difference; compared with the mRS score, the Delong test value Z=4.31, P<0.05, also indicating a statistically significant difference. These results indicate that the XGBoost model has significantly better predictive power than traditional neurological function assessment tools.

[0060] S500. Use the SHAP tool to interpret the optimal prediction model and the common feature variables.

[0061] The Shapley Plus Interpretation (SHAP) tool was used to explain how the 10 selected feature variables predicted the neurological prognosis of CA patients in an XGBoost model. Figure 9 This illustrates a matrix diagram showing the importance of feature variables in XGBoost model development in embodiments of this application, wherein... Figure 9 In this context, A represents the SHAP interpretation of feature importance ranking. Figure 9 In the graph, B represents the SHAP dependency contribution (yellow dots represent higher feature values, and purple dots represent lower feature values), and the SHAP value on the X-axis indicates the contribution of the XGBoost model. For example... Figure 9 As shown, the characteristic variables are arranged from highest to lowest importance: ROSC time, NSE, no-perfusion time, S100β, initial defibrillable rhythm, abnormal EEG, lactate, age, GWR, and APTT. SHAP-dependent contribution plot ( Figure 9 (B) indicates the relationship between individual trait values ​​and SAP values ​​to predict different neurological prognoses. For each trait, we determined a threshold to distinguish between positive and negative SAP values. That is, the horizontal position indicates whether the value's effect is associated with a higher or lower prediction; a positive increase has a positive effect and will predict a poor neurological prognosis; a negative increase has a negative effect and will predict a good neurological prognosis. Within each trait importance line, all patients' attributions to the outcome are plotted with dots of different colors, where yellow dots represent high-risk values, i.e., features that increase the risk of a poor neurological prognosis; and purple dots represent low-risk values, i.e., features that decrease the risk of a poor neurological prognosis.

[0062] Two examples are used to illustrate the SHAP explainability of the prediction model, Figure 10 A SHAP visualization example in the embodiments of the application is shown, wherein Figure 10 A in the formula (1) is used to predict patients with good neurological function, Figure 10 B in the formula (1) is used to predict patients with poor neurological function.

[0063] Example 1 (predicting patients with good neurological function): patient NSE = 18.1, abnormal EEG = no, no perfusion time = 3.5, APTT = 33.2, S100β = 1.15, ROSC time = 31, lactic acid = 7.3, initial defibrillation heart rate = yes, GWR = 1.21, age = 62, output prediction value f(x) = -0.0356 (see Figure 10 A in the formula (1)).

[0064] Example 2 (predicting patients with poor neurological function): age = 65, initial defibrillation heart rate = no, no perfusion time = 6.5, APTT = 31.3, ROSC time = 27, lactic acid = 4.6, abnormal EEG = no, S100β = 1.62, NSE = 41.7, GWR = 1.12, output prediction value f(x) = 1.7 (see Figure 10 B in the formula (1)). The yellow direction increases the risk of poor neurological function, and the purple direction reduces the risk of poor neurological function. The length of the arrow represents the weight of the influence of each variable on the outcome.

[0065] The same exclusion criteria as in step S100 were used to collect clinical data of 133 patients who survived CPR in a hospital in Shenzhen from March 2020 to March 2024, to verify the universality and extrapolation of the model.

[0066] Based on the 10 characteristic variables screened in step S200, 6 kinds of ML models were constructed to predict the neurological function prognosis of CA patients. AUC was used to evaluate the discrimination of the prediction model. Figure 11 The ROC curve of the external validation of the 6 kinds of machine learning models in the embodiments of the application is shown, as shown in Figure 11 The AUC value of the XGBoost model is the highest, AUC = 0.831 (95% CI: 0.764-0.898). Figure 12 The performance index heat map of the 6 kinds of machine learning models in the embodiments of the application in the external validation group is shown. As shown in Figure 12 XGBoost model: specificity = 0.86, sensitivity = 0.91, F1 score = 0.86, precision = 0.97, and the overall prediction performance is better than other prediction models.

[0067] Figure 13 The decision curve of the 6 kinds of machine learning models in the embodiments of the application in the external validation group is shown. As shown in Figure 13As shown, the threshold range of the XGBoost model is about 0.2-0.8, which is obviously wider than the applicable range of other models. In addition, DCA partially overlaps with other models in most threshold ranges, and the net benefit difference is not obvious. However, in the threshold range of 0.6-0.8, the net benefit of the XGBoost model is higher than that of other models.

[0068] The method for predicting poor neurological prognosis is as follows:

[0069] (1) Define the patient feature vector:

[0070] Let x = [x1, x2, x3..., x 10 ]ᵀ; wherein x1= age (years); x2= perfusion time (min); x3= ROSC time (min); x4= initial defibrillation rhythm (yes / no); x5= NSE (ng / mL); x6= S100β (ng / mL); x7= lactic acid (mmol / L); x8= abnormal EEG (yes / no); x9= APTT (s); x 10 =GWR(HU).

[0071] (2) Input 10 common feature variable indicators into the XGBoost prediction model, wherein the model is composed of 300 regression trees with a maximum depth of 4; learning rate η=0.03; global bias γ = -1.782.

[0072] (3) Calculate the probability P of poor neurological prognosis: ;

[0073] wherein, , is the prediction output of the mth regression tree.

[0074] (4) When P≥0.42, output "high risk of poor neurological prognosis" prompt.

[0075] Through the above specific embodiments, the skilled in the art can easily implement the present application. However, it should be understood that the present application is not limited to the above specific embodiments. On the basis of the disclosed embodiments, the skilled in the art can arbitrarily combine different technical features to realize different technical solutions.

Claims

1. A method for constructing a predictive model for neurological functional prognosis after cardiopulmonary resuscitation, characterized in that, include: We obtained clinical data of patients who achieved spontaneous circulation after cardiopulmonary resuscitation and screened clinical patients who met the inclusion criteria. Using brain function performance grading as the dependent variable, LASSO regression analysis and Boruta feature selection were used to screen characteristic variables, and common characteristic variables were selected. Specifically, using brain function performance grading as the dependent variable, LASSO regression analysis was first used to screen 16 characteristic variables: age, ROSC time, time without perfusion, initial defibrillable rhythm, mechanical ventilation, adrenaline dosage, neuron-specific enolase, S100 calcium-binding protein β, lactate, abnormal EEG, prothrombin time, activated partial thromboplastin time, serum creatinine, hemoglobin, gray / white matter ratio on head CT, and troponin. Then, Boruta feature selection was used for analysis to screen... Fourteen characteristic variables were selected: time without perfusion, S100 calcium-binding protein β, brain natriuretic peptide, abnormal EEG, neuron-specific enolase, ROSC time, age, gray / white matter ratio on head CT, lactate, activated partial thromboplastin time, acute physiology and chronic health assessment, D-dimer, platelet count, and initially defibrillable rhythm. Finally, ten common characteristic variables were selected from two groups: age, time without perfusion, ROSC time, initially defibrillable rhythm, neuron-specific enolase, S100 calcium-binding protein β, lactate, abnormal EEG, activated partial thromboplastin time, and gray / white matter ratio on head CT. Based on common feature variables, multiple machine learning models were constructed to predict the neurological functional prognosis of cardiopulmonary resuscitation patients, and the ROC curve and decision curve corresponding to each machine learning model were plotted. The area under the ROC curve was used to evaluate the discriminative power of multiple machine learning models, and the decision curve was used to evaluate the clinical efficacy of multiple machine learning models. The optimal prediction model was selected as the extreme gradient boosting model. The SHAP tool is used to interpret the optimal prediction model and the common feature variables. The SHAP value of each common feature variable is calculated, and the common feature variables are ranked in importance according to the SHAP values. A dependency contribution plot is then drawn.

2. The method according to claim 1, characterized in that, The exclusion criteria include: Inclusion criteria were age ≥18 years and hospital stay >24 hours after successful resuscitation; exclusion criteria were history of neurological disorders and >20% of clinical data missing.

3. The method according to claim 1, characterized in that, The various machine learning models mentioned are the extreme gradient boosting model, the support vector machine model, the random forest model, the K-nearest neighbor algorithm model, the decision tree model, and the logistic regression model.

4. A method for predicting neurological functional prognosis after cardiopulmonary resuscitation, characterized in that, The prediction method uses an extreme gradient boosting model constructed by the construction method described in any one of claims 1-3 to make predictions; The prediction method includes: Define the feature vector of common patient characteristic variables; Input the feature vectors of the common feature variables into the extreme gradient boosting model; Calculate the probability of a poor prognosis in neurological dysfunction; The calculated probability is compared with the poor prognosis threshold. If the probability of poor neurological function prognosis is greater than or equal to the poor prognosis threshold, a high risk of poor neurological function prognosis is output; if the probability of poor neurological function prognosis is less than the poor prognosis threshold, a low risk of poor neurological function prognosis is output.

5. The prediction method according to claim 4, characterized in that, The learning rate of the extreme gradient boosting model is 0.03, global bias is -1.782, number of trees is 300, and maximum tree depth is 4.

6. The prediction method according to claim 5, characterized in that, The threshold for poor prognosis is 0.42.

Images