Construction method of hospitalization death risk prediction model for sepsis blood coagulation disorder patient
By constructing a predictive model of hospitalized death risk for patients with septic coagulation disorders, using the MIMIC-IV database and urinary system dataset to screen key characteristics and predictors, the high mortality problem caused by septic coagulation dysfunction was solved, and the high mortality rate caused by septic coagulation dysfunction was achieved. Accurate prediction and management of SIC patients reduces the risk of death.
Patent Information
- Application Number
- CN202411055774.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-08
AI Technical Summary
Coagulation dysfunction caused by sepsis leads to high hospitalization mortality rates for patients, and it is difficult for the existing technology to predict the condition in advance, making it difficult to adjust treatment in a timely manner.
Construct a predictive model of hospitalized death risk in patients with septic coagulation disorders. Using the MIMIC-IV database and urinary system dataset, use a variety of machine learning models to train and verify key features and predictors to determine the most suitable model to reduce death risk.
The prediction of hospitalization mortality in patients with SIC through models provides clinical insights to help accurately manage, reduce patient mortality and improve the prediction accuracy of treatment.
Smart Images

Figure CN118919084A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning, and in particular to a method for constructing a risk prediction model for hospitalized death in patients with sepsis and coagulation disorders. Background Art
[0002] Sepsis-induced coagulopathy (SIC) is a common and serious complication in critically ill patients with sepsis, often leading to adverse outcomes.
[0003] The coagulation dysfunction caused by sepsis is usually manifested clinically as disseminated intravascular coagulation, abnormal coagulation factors, platelet dysfunction, abnormal fibrin degradation, and imbalance of the coagulation and fibrinolytic systems. In the clinical treatment of such diseases, the above coagulation dysfunction will significantly increase the in-hospital mortality of patients with sepsis. Therefore, in clinical management, it is necessary to closely monitor the coagulation function of patients, adjust the treatment strategy in time, and take appropriate intervention measures to improve the prognosis. However, since it is too late after the above dysfunction occurs, if the patient's preclinical behavior cannot be predicted in advance, it will cause unavoidable situations. Therefore, a method for constructing a risk prediction model for in-hospital mortality in patients with sepsis coagulation disorders is proposed. Summary of the invention
[0004] The purpose of the present invention is to provide a method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulation disorders, so as to solve the problems raised in the above-mentioned background technology.
[0005] To achieve the above object, the present invention provides the following technical solution: a method for constructing a prediction model for hospitalized mortality risk in patients with sepsis and coagulopathy, the method comprising the following steps:
[0006] S1, establishing a training cohort, wherein the data content of the training cohort is composed of a data set of patients diagnosed with sepsis recorded in the MIMIC-IV database;
[0007] S2, establishing an external validation cohort, wherein the data content of the validation cohort is composed of a data set of patients with urinary sepsis in the hospital within 3-5 years before the prediction time;
[0008] S3, extracting data, using Navicat Premium software to extract data, and preliminarily obtaining raw data through structured query language, and then screening the raw data;
[0009] S4, candidate variables and predictors were selected. After the data were screened, the continuous variables of the clinical characteristics of sepsis patients were standardized, and the categorical variables were encoded using the one-hot encoding method;
[0010] S5, building a machine learning model, training and validating through multiple machine learning models, and each model is configured to utilize specific features;
[0011] S6, develop and compare models, divide the MIMIC-IV dataset into 7:3 ratio, 70% of the data is used for model training, and the remaining 30% is used to test model performance. The most suitable hyperparameters for each model are determined through internal 5-fold cross-validation, and the machine learning models are compared through external 5-fold cross-validation to determine the best performing model as the final prediction model.
[0012] Preferably, the data content of the MIMIC-IV database in S1 is clinical information of sepsis patients, and the clinical information includes demographic data, vital signs, imaging results, laboratory tests, data dictionaries, documents containing International Classification of Disease codes, and hourly physiological data verified by ICU nurses.
[0013] Preferably, the sepsis-induced coagulopathy scoring system is used to determine whether eligible sepsis patients have coagulopathy, and patients are considered to have coagulopathy if their total score is greater than or equal to four within the first day of ICU admission and the combined score of the prothrombin time international normalized ratio and platelet count parameters exceeds two.
[0014] Preferably, the structured query language in S3 is used to extract patient information, including sociodemographic characteristics, vital signs, laboratory results, complications and microbiological data.
[0015] Preferably, the sociodemographic characteristics include age at admission, gender, race, weight, height, ICU stay time, and in-hospital mortality markers of sepsis patients at the time of first ICU admission;
[0016] The vital signs are those of sepsis patients within 24 hours before admission to the ICU, including heart rate, body temperature, respiratory rate and blood oxygen saturation;
[0017] The laboratory results included routine blood tests, liver and kidney function tests, blood glucose, and arterial blood gas analysis parameters;
[0018] The complications were identified using the Charlson table;
[0019] The microbiological data included mechanical ventilation and renal replacement therapy.
[0020] Preferably, if there are missing data in the patient information features, the K nearest neighbor classification algorithm in the "DMwR2" package of R is used to process the missing data for each feature with less than 30% missing values.
[0021] Preferably, the criteria for screening and eliminating the S3 data include patients under 18 years old, patients not admitted for the first time with sepsis, pregnant women, patients with congenital coagulation disorders, patients who died or left the ICU within 24 hours after admission, and patients without coagulation dysfunction.
[0022] Preferably, after the prediction factor encoding in S4 is completed, the coefficient estimate is shrunk to zero using least absolute shrinkage and selection operator regression, and the degree of shrinkage is determined by an additional parameter penalty factor.
[0023] Preferably, the types of models in S5 include support vector machine model, logistic regression model, random forest model, lightweight gradient boosting machine model, recursive partitioning and regression tree model, generalized linear model through elastic net, K nearest neighbor model, neural network model and extreme gradient boosting model.
[0024] Preferably, the evaluation indicators of the final model in S6 include the area under the receiver operating characteristic curve, the area under the precision-recall curve, the F1 score, accuracy, positive predictive value, precision and recall;
[0025] And the receiver operating characteristic curve, precision-recall curve, and Kaplan-Meier calibration curve were used for statistical analysis. Descriptive statistics were used to summarize the baseline characteristics of the patients;
[0026] The chi-square test was used to compare categorical variables, and the t-test was used to compare continuous variables. The statistical significance level was set at p < 0.05. Data analysis was performed using R and Python programming languages.
[0027] Technical effects and advantages of the present invention:
[0028] The present invention analyzes the predictive performance of various machine learning models in predicting the in-hospital mortality of SIC patients, uses the MIMIC-IV dataset for model evaluation, and uses the urinary system dataset for external validation. For each algorithm in the training cohort, a five-fold cross-validation grid search is performed to determine the optimal hyperparameters, and the result is a model with the highest predictive accuracy. This optimal model is then evaluated in the test cohort. The in-hospital mortality of SIC patients is predicted by this model, which provides clinical insights, helps to accurately manage patients, and reduces patient mortality. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 An operational flow chart is designed for the model study of the present invention.
[0030] Figure 2 Figure 2 ROC curves of nine models in predicting the prognosis of in-hospital mortality in SIC patients.
[0031] Figure 3PR curves of nine models in predicting the prognosis of in-hospital mortality in SIC patients.
[0032] Figure 4 The area under the PR curve predicts the performance probability graph.
[0033] Figure 5 Net benefit curves of prognostic models were analyzed for decision curves.
[0034] Figure 6 SHAP summary plot of feature importance analysis for a random forest model.
[0035] Figure 7 Partial dependence plot of the SHAP-based random forest model.
[0036] Figure 8 Waterfall plots for local model interpretation via SHAP and LIME methods.
[0037] Fig. 9 LIME plots for local model interpretation by the LIME method for six patients. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] Embodiment 1, the present invention provides as follows Figure 1 The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulation disorders shown in FIG. 1 includes the following steps:
[0040] S1, establishing a training cohort, wherein the data content of the training cohort is composed of a data set of patients diagnosed with sepsis recorded in the MIMIC-IV database;
[0041] It should be noted that the MIMIC-IV database contains extensive clinical data of patients admitted to Beth Israel Deaconess Medical Center from 2008 to 2019. It includes more than 200,000 emergency room visit records and more than 70,000 ICU admission records.
[0042] Specifically, the data content of the MIMIC-IV database in S1 is the clinical information of sepsis patients, and the clinical information includes demographic data, vital signs, imaging results, laboratory tests, data dictionaries, documents containing International Classification of Disease codes, and hourly physiological data verified by ICU nurses.
[0043] S2, establishing an external validation cohort, wherein the data content of the validation cohort is composed of a data set of patients with urinary sepsis in the hospital within 3-5 years before the prediction time;
[0044] Specifically, according to the sepsis-induced coagulopathy scoring system, eligible sepsis patients were determined to have coagulopathy. If the total score of the patient was greater than or equal to four within the first day of ICU admission, and the combined score of the prothrombin time international normalized ratio and platelet count parameters exceeded two, the patient was considered to have coagulopathy.
[0045] S3, extracting data, using Navicat Premium software to extract data, and preliminarily obtaining raw data through structured query language, and then screening the raw data;
[0046] Specifically, the structured query language in S3 was used to extract patient information, including sociodemographic characteristics, vital signs, laboratory results, comorbidities, and microbiological data.
[0047] Furthermore, the sociodemographic characteristics included age at admission, sex, race, weight, height, ICU length of stay, and in-hospital mortality markers of sepsis patients at first ICU admission;
[0048] The vital signs are those of sepsis patients within 24 hours before admission to the ICU, including heart rate, body temperature, respiratory rate and blood oxygen saturation;
[0049] The laboratory results included routine blood tests, liver and kidney function tests, blood glucose, and arterial blood gas analysis parameters;
[0050] The complications were identified using the Charlson table;
[0051] The microbiological data included mechanical ventilation and renal replacement therapy.
[0052] Furthermore, if there were missing data in the patient information features, the K nearest neighbor (KNN) classification algorithm in the “DMwR2” package of R was used to process the missing data for each feature with less than 30% missing values.
[0053] It should be noted that KNN is a non-parametric machine learning algorithm that estimates missing values based on the values of neighboring data points. The interpolated values are calculated by averaging the measured values of neighboring data points. This method estimates missing values by averaging the non-missing values of neighboring data points, and also handles the case where all neighboring points are missing by averaging the overall column of the feature. This technique is useful for datasets with missing values that contain many variables.
[0054] Specifically, the criteria for S3 data screening and exclusion include patients under 18 years old, patients not admitted for the first time with sepsis, pregnant women, patients with congenital coagulation disorders, patients who died or left the ICU within 24 hours after admission, and patients without coagulation dysfunction.
[0055] S4, candidate variables and predictors were selected. After the data were screened, the continuous variables of the clinical characteristics of sepsis patients were standardized, and the categorical variables were encoded using the one-hot encoding method;
[0056] Specifically, after the predictor encoding in S4 is completed, the coefficient estimates are shrunk to zero using the least absolute shrinkage and selection operator regression, and the degree of shrinkage is determined by the additional parameter penalty factor (lambda).
[0057] It should be noted that in order to determine the optimal penalty factor (lambda), a ten-fold cross-validation error plot was constructed for the LASSO model. The Boruta algorithm, a widely used feature selection method, played a key role in this process. The algorithm creates a set of "shadow features" from the original dataset and compares their Z scores with the original features. If the Z score of a feature exceeds the maximum possible Z score of the shadow feature, the feature is considered significant and retained; otherwise, it is excluded. By combining LASSO regression and the Boruta method, this study determined the most favorable feature combination.
[0058] S5, constructing a machine learning model, training and validating multiple machine learning models separately, and each model is configured to utilize specific features to distinguish the good and bad short-term prognosis of SIC patients;
[0059] Specifically, the types of models in S5 include support vector machine models, logistic regression models, random forest models, lightweight gradient boosting machine models, recursive partitioning and regression tree models, generalized linear models through elastic net, K nearest neighbor models, neural network models, and extreme gradient boosting models.
[0060] S6, develop and compare models, divide the MIMIC-IV dataset into 7:3 ratio, 70% of the data is used for model training, and the remaining 30% is used to test model performance. The most suitable hyperparameters for each model are determined through internal 5-fold cross-validation, and the machine learning models are compared through external 5-fold cross-validation to determine the best performing model as the final prediction model.
[0061] Specifically, the evaluation metrics of the final model in S6 include the area under the receiver operating characteristic curve, the area under the precision-recall curve, the F1 score, accuracy, positive predictive value, precision, and recall, which are used to evaluate the performance of each model;
[0062] And the receiver operating characteristic curve, precision-recall curve, and Kaplan-Meier calibration curve were used for statistical analysis. Descriptive statistics were used to summarize the baseline characteristics of the patients;
[0063] The chi-square test was used to compare categorical variables, and the t-test was used to compare continuous variables. The statistical significance level was set at p < 0.05. Data analysis was performed using R and Python programming languages, using packages such as scikit-learn, TensorFlow, and LightGBM. R's "tidymodels" package is used for machine learning. It is a meta-package for modeling and statistical analysis that shares the underlying design concepts, syntax, and data structures.
[0064] It should be noted that the recursive feature elimination (RFE) method was used to determine the predictors of in-hospital mortality. The model performance was evaluated by receiver operating characteristic (ROC) curve and decision curve analysis (DCA). The SHapley Additive exPlanations (SHAP) method was used to provide model interpretability.
[0065] Example 2, reference Figure 1 As shown, 2244 elderly SIC patients from the MIMIC-IV database and 759 patients from the WMU database were recruited as training and external validation datasets, respectively. After excluding cases with missing follow-up data in the MIMIC-IV training cohort, 35.0% (785 / 2244) of patients died in the hospital. In addition, the external validation cohort from the WMU showed an in-hospital mortality rate of 24.8% (188 / 759).
[0066] In all cohorts, the mean age of patients who survived was significantly lower than that of patients who did not survive. In the training cohort, the mean age of survivors was 60.61 years, while that of nonsurvivors was 62.99 years, a significant difference (P = 0.004). This trend was consistent in the validation cohort (60.77 years vs. 65.55 years, P = 0.003) and the external cohort of the WMU (62.48 years vs. 66.63 years, P = 0.003), indicating that old age is a potential risk factor for mortality. Body mass index (BMI) did not show significant differences in most cohorts, indicating that BMI alone may not be a strong predictor of outcome in these patient populations. However, in the WMU cohort, the mean BMI of survivors (23.05) was significantly higher than that of nonsurvivors (22.03), a significant difference (P = 0.026).
[0067] Regarding comorbidities, the prevalence of chronic obstructive pulmonary disease (COPD) was significantly higher in nonsurvivors than in survivors in the training cohort (18% vs. 12%, P = 0.001). Although similar trends were observed in the validation and WMU cohorts, the differences were not statistically significant, indicating the variability of the effect of COPD on mortality in different populations. History of stroke was another important factor, showing significant differences in both the training cohort (9% in nonsurvivors vs. 5.3% in survivors, P = 0.004) and the WMU cohort (16% vs. 8.2%, P = 0.005). This consistent pattern across cohorts highlights history of stroke as an important factor in mortality risk. The prevalence of malignancies was significantly higher in nonsurvivors in both the training cohort (25% in nonsurvivors vs. 20% in survivors, P = 0.028) and validation cohorts (28% vs. 19%, P = 0.017). Although the WMU cohort showed a higher prevalence of malignancies among nonsurvivors, the difference was not statistically significant (19% vs. 13%, P = 0.092). These findings suggest that malignancies may lead to worse outcomes, but with some cohort variability.
[0068] Acute kidney injury (AKI) within 48 hours of admission was strongly associated with mortality. In the training cohort, 92% of nonsurvivors had AKI compared with 82% of survivors (P < 0.001). This association was also significant in the validation cohort (90% vs. 81%, P = 0.007) and the WMU cohort (55% vs. 33%, P < 0.001), highlighting AKI as an important predictor of mortality.
[0069] In conclusion, baseline demographic and clinical characteristics showed significant differences between survivors and nonsurvivors in terms of age, stroke history, presence of malignancy, AKI within 48 hours of admission, and SIC score. These factors are critical for understanding patient outcomes and guiding clinical interventions.
[0070] Example 3, Results The training set includes 1682 SIC patients, of which 1094 survived and 588 did not survive. The validation set includes 562 patients, of which 365 survived and 197 did not survive. In addition, the external validation cohort includes 759 patients, of which 571 survived and 188 did not survive. 40 variables were selected for model development. The ROC AUC values of each model were: logistic regression 0.820, rpart 0.737, glmnet 0.803, KNN 0.851, nnet 0.816, LightGBM 0.817, random forest 0.920, XGBoost 0.863, SVM 0.817. The random forest model performed best in the training set with an AUC of 0.920 and maintained good performance in the external validation cohort. The summary graph and force graph based on SHAP provide global and local interpretability of the random forest model. It can be seen that machine learning is a reliable tool for predicting in-hospital mortality in SIC patients. The SHAP method clarified the internal mechanism of the RF model and provided valuable clinical insights to help with precise patient management.
[0071] Example 4, reference Figure 2 As shown in Figure 2, the ROC curves of the prognosis of in-hospital death in SIC patients were predicted using 9 machine learning algorithms for the training set, test set, and WMU set. LASSO regression and Boruta algorithm were used to select features from the data of the training cohort. Specifically, we identified 26 feature variables from the 46 variable pools.
[0072] For each algorithm in the training cohort, a five-fold cross-validation grid search was performed to determine the best hyperparameters, resulting in the model with the highest predictive accuracy. This best model was then evaluated in the test cohort. The best predictive performance in the training cohort was achieved by the random forest (RF) model, with an AUROC of 0.920, AUPRC of 0.87, F1 score of 0.783, accuracy (ACC) of 0.831, positive predictive value (PPV) of 0.708, precision of 0.708, and recall of 0.876. In addition, in the external validation cohort, the simplified RF model showed good SIC discrimination ability in elderly surviving patients using the first 26 features, with an AUROC of 0.779, AUPRC of 0.714, F1 score of 0.520, ACC of 0.543, PPV of 0.641, precision of 0.732, and recall of 0.878.
[0073] Combination Figure 2 and Figure 3 The ROC and PR curves of nine machine learning algorithms on different data sets are shown, indicating that the random forest model consistently outperforms other models in predictive performance, with higher AUROC and AUPRC values. Figure 4The effectiveness of the model in predicting in-hospital mortality was further emphasized by the area under the precision-recall curve. Figure 5 A decision curve analysis is presented showing the net benefit of the models in a clinical decision scenario. The random forest model provides the highest net benefit at various threshold probabilities.
[0074] Example 5. Clinicians often have difficulty trusting predictive models that lack transparency and explainability. To improve the interpretability of the final model, we used the SHapley Additive exPlanations (SHAP) method to explain the contribution of each variable to the model prediction. This method provides two types of explanations: global explanations at the feature level and local explanations at the individual level, providing comprehensive insights into the overall model behavior.
[0075] Figure 6 and Figure 7 The SHAP method is used to perform feature importance analysis. Figure 6 ) identifies the top 20 features that influence the model’s predictions and visually displays the impact of each feature. Figure 7 Partial dependence plots are shown, indicating the relationship between specific features and the predicted outcomes.
[0076] exist Figure 7 In the study, the actual values and SHAP values of 24 features were compared. A SHAP value above zero indicates a positive class prediction in the model, indicating a higher risk of in-hospital death. For example, SIC patients aged 90 years or older or patients with an anion gap (AG) of 30 or more had a SHAP value above zero, indicating a higher likelihood of being classified into the “death” category. Similarly, low albumin values (≤3) or high SOFA scores (≥10) also favored a decision toward the “death” category.
[0077] Figure 8 Local explanations of individual predictions are provided by SHAP and Local Explanations Model-Independent Explanations (LIME) methods. Waterfall plots ( Figure 8 ) and LIME graphs ( Fig. 9 ) provides model insights into the decision-making process of individual patients, distinguishing between low-risk and high-risk cases.
[0078] Figure 8 A case in which a patient did not die during hospitalization is presented. In this case, alkaline phosphatase (ALP), albumin, bicarbonate, SOFA, red cell distribution width (RDW), anion gap, prothrombin time (PT), and 17 other features favored the decision for the “death” category, while lactate dehydrogenase (LDH) and mean corpuscular volume (MCV) did not. Conversely, lower-than-normal LDH and MCV may reduce the patient’s risk of death, even if other factors indicate otherwise.
[0079] Fig. 9 Examples of local explanations for six patients are shown, depicting false positive, true positive, false negative, and true negative cases via LIME. In these illustrations, blue bars represent features that are positively correlated with the output, and red bars represent features that are negatively correlated with the output.
[0080] For example, in Fig. 9 In case 2, low bicarbonate values (less than 20), low body mass index (BMI) (less than 25.2), and infection site value of 0 were positively associated with death, according to the LIME description.
[0081] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulation disorders, characterized in that: The construction method comprises the following steps: S1, establishing a training cohort, wherein the data content of the training cohort is composed of a data set of patients diagnosed with sepsis recorded in the MIMIC-IV database; S2, establishing an external validation cohort, wherein the data content of the validation cohort is composed of a data set of patients with urinary sepsis in the hospital within 3-5 years before the prediction time; S3, extracting data using Navicat Premium software, and preliminarily obtaining raw data through structured query language, and then screening the raw data; S4, candidate variables and predictors were selected. After the data were screened, the continuous variables of the clinical characteristics of sepsis patients were standardized, and the categorical variables were encoded using the one-hot encoding method; S5, building a machine learning model, training and validating through multiple machine learning models, and each model is configured to utilize specific features; S6, develop and compare models, divide the MIMIC-IV dataset into 7:3 ratio, 70% of the data is used for model training, and the remaining 30% is used to test model performance. The most suitable hyperparameters for each model are determined through internal 5-fold cross-validation, and the machine learning models are compared through external 5-fold cross-validation to determine the best performing model as the final prediction model.
2. The method for constructing a risk prediction model for hospitalized death in patients with sepsis and coagulopathy according to claim 1, characterized in that: The data content of the MIMIC-IV database in S1 is the clinical information of sepsis patients, and the clinical information includes demographic data, vital signs, imaging results, laboratory tests, data dictionaries, documents containing International Classification of Disease codes, and hourly physiological data verified by ICU nurses.
3. The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulopathy according to claim 2, characterized in that: Eligible sepsis patients were determined to have coagulopathy according to the Sepsis-Induced Coagulopathy Scoring System, and were considered to have coagulopathy if their total score was ≥ four and their combined score for the international normalized ratio of the prothrombin time and the platelet count parameters was greater than two within the first day of ICU admission.
4. The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulopathy according to claim 1, characterized in that: The structured query language in S3 was used to extract patient information, including sociodemographic characteristics, vital signs, laboratory results, comorbidities, and microbiological data.
5. The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulopathy according to claim 4, characterized in that: The sociodemographic characteristics included age, sex, race, weight, height, ICU length of stay, and in-hospital mortality markers of sepsis patients at admission; The vital signs are those of sepsis patients within 24 hours before admission to the ICU, including heart rate, body temperature, respiratory rate and blood oxygen saturation; The laboratory results included routine blood tests, liver and kidney function tests, blood glucose, and arterial blood gas analysis parameters; The complications were identified using the Charlson table; The microbiological data included mechanical ventilation and renal replacement therapy.
6. The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulopathy according to claim 5, characterized in that: If there were missing data in the patient information features, the K nearest neighbor classification algorithm in the "DMwR2" package of R was used to process the missing data for each feature with less than 30% missing values.
7. The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulopathy according to claim 1, characterized in that: The criteria for screening and eliminating the S3 data include patients under 18 years old, patients not admitted for the first time with sepsis, pregnant women, patients with congenital coagulation disorders, patients who died or left the ICU within 24 hours after admission, and patients without coagulation dysfunction.
8. The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulopathy according to claim 1, characterized in that: After the prediction factor encoding in S4 is completed, the coefficient estimates are shrunk to zero using the least absolute shrinkage and selection operator regression, and the degree of shrinkage is determined by the additional parameter penalty factor.
9. The method for constructing a risk prediction model for hospitalized mortality in patients with sepsis and coagulopathy according to claim 1, characterized in that: The types of models in S5 include support vector machine model, logistic regression model, random forest model, lightweight gradient boosting machine model, recursive partitioning and regression tree model, generalized linear model through elastic net, K nearest neighbor model, neural network model and extreme gradient boosting model.
10. The method for constructing a risk prediction model for hospitalized death in patients with sepsis and coagulopathy according to claim 1, characterized in that: The evaluation indicators of the final model in S6 include the area under the receiver operating characteristic curve, the area under the precision-recall curve, the F1 score, accuracy, positive predictive value, precision and recall; And the receiver operating characteristic curve, precision-recall curve, and Kaplan-Meier calibration curve were used for statistical analysis. Descriptive statistics were used to summarize the baseline characteristics of the patients; The chi-square test was used to compare categorical variables, and the t-test was used to compare continuous variables. The statistical significance level was set at p < 0.
05. Data analysis was performed using R and Python programming languages.
Citation Information
Cited By
Inplanatable ML method and system for predicting death risk of ICU stroke patient in hospital
CN119230130A
Acute myocardial infarction and atrial fibrillation coexistence risk prediction model construction method and application
CN119560162A
Method and system for predicting sepsis occurrence risk of ICU patient
CN119581018A
Construction method and system of intestinal lymphoma and Crohn disease identification model
CN120107161A
Model for predicting severe SA-AKI risk of ICU sepsis patient by combining clinical and immune inflammation variables
CN122201783A
