CN116543902A - An interpretable death risk assessment model, device and establishment method for critically ill children - Google Patents

Abstract

The invention discloses an interpretable severe child death risk assessment model, a device and an establishment method thereof, which concretely comprise the following steps: and establishing an interpretable severe child death risk assessment model based on an extreme gradient lifting (XGBoost) integrated machine learning algorithm fused with an interpretable SHAP method. The characteristic variables are determined by collecting 5 types of clinical data, such as target children demographics, vital signs, disease scores, laboratory indicators, treatment and the like. The evaluation module ranks the feature variables from high to low according to the importance of the input features on the prediction result; and the evaluation module performs severe child death risk evaluation based on input features corresponding to at least some of the features, and calculates the contribution of each input feature to the prediction result as the contribution degree of the risk factors. The interpretable severe child death risk assessment model and the interpretable severe child death risk assessment device are beneficial to more comprehensive and early-stage disease emergency and risk degree assessment of the severe child by pediatricians in different areas and different centers, so that scientific basis is provided for accurate diagnosis and treatment.

Description

A model, device and establishment method for interpreting the risk assessment of death in critically ill children

Technical Field

The present invention relates to the field of medical technology, and more specifically, to an interpretable death risk assessment model for critically ill children, a device and an establishment method thereof.

Background Art

Child mortality is a comprehensive and feasible indicator to measure the level of social development and civilization progress of a country or region, including economy, culture, and health. Critically ill children threaten human health due to their high mortality. Therefore, seeking indicators that can timely and accurately predict the death of critically ill children is of great significance for actively taking effective measures to prevent child deaths and reduce child mortality. At present, clinical risk assessment of critically ill children at home and abroad still relies on disease severity scores such as child mortality risk scores, children's sequential organ failure scores, or single clinical indicators such as blood sugar, blood lactate, blood creatinine, volume overload, or single biomarkers such as urine cystatin C, blood soluble urokinase plasminogen activator receptor, etc. However, these existing child critical assessment and prognosis judgment indicators/scores may be affected by other factors such as age, weight, and gender, or do not have sufficient sensitivity and specificity, and the clinical prediction performance cannot be fully affirmed in multi-center, large-sample clinical studies. In recent years, research on disease prediction models based on electronic medical records has greatly promoted the development of more accurate disease risk assessment models/scores, but many studies are limited by limited sample sets (such as single centers, small samples), making it impossible to guarantee the universality and robustness of the models; or the research is mainly focused on adults, and there are certain pathological and physiological differences between children and adults. Therefore, the use of the same evaluation criteria may cause evaluation bias, and the risk assessment model established for adults is difficult to apply to children.

Summary of the invention

In view of the above problems, the present invention develops a prediction model for critically ill children admitted to the pediatric intensive care unit (PICU) based on a multi-center data set, which can timely evaluate the risk of death during PICU hospitalization, and simultaneously presents the reasoning and analysis of the model to facilitate doctors' understanding, so as to help pediatricians more comprehensively and timely realize the urgency and danger level of the children's potential diseases, thereby providing a scientific basis for the next decision-making and treatment.

The present invention adopts the following technical solution:

A model that can explain the risk of death in critically ill children, including:

(1) Dataset construction module: obtain data of children admitted to the PICU and determine characteristic variables;

(2) Data processing module: cleans and merges the data in the data set, performs sampling and interpolation, and obtains multiple statistical features;

(3) Model construction and evaluation module: The extreme gradient boosting XGBoost model integrated with the SHAP method is used to train the model and optimize the parameters of the processed data set to build an interpretable mortality risk assessment model for critically ill children; multiple statistical features are arranged from high to low according to their importance to the prediction results.

A method for establishing an interpretable critically ill child mortality risk assessment model includes data set construction, data processing, model construction and evaluation. In data set construction, a data set of children admitted to a PICU is obtained to determine characteristic variables, wherein the characteristic variables include general demographics, vital signs, disease severity scores of critically ill children, laboratory tests and treatments. In data processing, data from the data set are cleaned, integrated, sampled and interpolated to obtain multiple statistical features. In model construction and evaluation, the data set is trained and parameterized based on an extreme gradient boosting XGBoost model fused with a SHAP method, and multiple statistical features are arranged from high to low according to their importance to the prediction results.

In the present invention, the characteristic variables include general demographics, vital signs, disease severity scores of critically ill children, laboratory tests and treatment.

In the present invention, the data processing module is: cleaning and merging the data of the data set, including checking data consistency, processing invalid values and missing values, interpolating missing data by median interpolation, and eliminating if the missing ratio is greater than or equal to 20%; excluding independent variables with significant collinearity and little contribution to the dependent variable by multicollinearity test; constructing model input features based on the regularized data, that is, constructing statistical features, including original value, mean, median, quartile, maximum value, minimum value, sum, and obtaining 4 to 42 statistical features. Preferably, the input features are the first 4, 6, 12, 20 or 42 features in its multiple features in descending order of importance.

In the present invention, the statistical characteristics include age, weight, height, body mass index BMI, gender, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, body temperature, the third-generation child mortality risk score, the child sequential organ failure score, the Glasgow Coma Scale, inspired oxygen concentration FiO ₂ , oxygen partial pressure PaO ₂ , carbon dioxide partial pressure PaCO ₂ , oxygenation index PaO ₂ /FiO ₂ , SpO ₂ /FiO ₂ , blood pH, bicarbonate HCO3 ^- , excess base BE, prothrombin time PT, partial thromboplastin time APTT, international normalized ratio INR, antithrombin III, troponin cTn, creatine phosphokinase isoenzyme CKMB, alanine aminotransferase ALT, aspartate aminotransferase AST, total bilirubin Tbil, lactate dehydrogenase LDH, creatinine SCr, urea nitrogen BUN, albumin Alb, glucose Glu, potassium ion K ⁺ , sodium ion Na ⁺ , chloride Cl ^- , calcium ion Ca ²⁺ , lactate, white blood cells WBC, neutrophils Neu, platelets PLT, procalcitonin PCT, C-reactive protein CRP, glomerular filtration rate eGFR, whether mechanical ventilation is performed, whether renal replacement therapy is performed, whether vasoactive drugs are used, vasoactive drug scores, whether furosemide is used, whether hormones are used, whether antibiotics are used, whether meping/Tianon is used, whether vancomycin is used. In the present invention, the evaluation module ranks the features from high to low according to the importance of the input features to the prediction results; and the evaluation module evaluates the risk of death in critically ill children based on the input features corresponding to at least some of the multiple features, and calculates the contribution of each of the input features to the prediction results as the contribution degree of the risk factor. The evaluation module is based on the extreme gradient boosting XGBoost model integrated with the SHAP method; the top 20 features among the multiple features in descending order of importance to the prediction results are: children's sequential organ failure score pSOFA, mechanical ventilation MV, lactate dehydrogenase LDH, C-reactive protein CRP, alanine aminotransferase ALT, the third-generation children's mortality risk score PRISM Ⅲ, maximum carbon dioxide partial pressure PaCO ₂ max, procalcitonin PCT, minimum body temperature Tepmin, platelet PLT, blood glucose Glu, blood creatinine SCr, creatine phosphokinase isoenzyme CKMB, body mass index BMI, blood calcium ion Ca ²⁺ , neutrophil count Neu, lactate Lactate, total bilirubin TBil, international normalized ratio INR, and blood potassium K ⁺ .

The present invention discloses a computer-readable carrier, including a computing program, which is used to execute the above-mentioned interpretable critically ill child death risk assessment model. The present invention discloses an interpretable critically ill child death risk assessment device, including a computing unit, which is used to execute the above-mentioned interpretable critically ill child death risk assessment model. The specific carrier and computer are conventional technologies. In the present invention, the data processing module obtains input features from the demographic information, vital signs, critically ill child disease severity score and laboratory test indicators of the child on the first day of admission to the pediatric intensive care unit (PICU) and the data of treatment during the PICU stay, and inputs them into the assessment module. The model uses the SHAP method it integrates to obtain the risk factor contribution degree assessment of individual critically ill children; wherein, red is used to represent that the factor is currently in an abnormal state and has a harmful effect on the outcome of the child; blue is used to represent that the factor is currently in a normal state and has no harmful effect on the outcome of the child, and the larger the SHAP value, the greater the degree of influence on the outcome.

In the present invention, the research population screens children based on the PICU admission standard and the established inclusion process to obtain a data set of critically ill children; in model construction and evaluation, the model is trained and tuned for critically ill children. When training the model, the data is fused as a large-sample, multi-center training set, wherein 70-90% of the children's data are used for model training and cross-validation to tune the hyperparameters of the prediction model, and the remaining children's data are used for internal verification of model performance.

When verifying the model performance, the eight evaluation indicators are: area under the receiver operating characteristic curve AUC, sensitivity, specificity, accuracy, precision, F1 value, area under the precision-recall curve AUPRC, calibration curve; the one functional indicator is the interpretability functional indicator. Internal validation uses 30-10% of the children's data from the multi-center data set for evaluation; external validation uses critically ill children data that are inconsistent with the training data set and the internal validation data set for evaluation; subgroup analysis divides the multi-center data set into critically ill children aged 2 years or less and older than 2 years for verification; to comprehensively evaluate whether there is bias and the universality and robustness of the model.

The interpretable critically ill child death risk assessment device disclosed in the present invention can execute the above-mentioned interpretable critically ill child death risk assessment model; the model uses the SHAP method integrated therein to obtain the risk factor contribution degree assessment of individual critically ill children; wherein, red is used to represent that the factor is currently in an abnormal state and has a harmful effect on the outcome of the child, and blue is used to represent that the factor is currently in a normal state and has no harmful effect on the outcome of the child, and the larger the SHAP value, the greater the degree of influence on the outcome. The computing unit can be a central processing unit, a single-chip microcomputer, etc.

The advantages of the present invention are:

(1) For critically ill children in the PICU, the probability of adverse outcomes (death) and the contribution of risk factors during the PICU stay can be predicted in a timely and accurate manner, thereby assisting pediatricians in timely intervention and accurate diagnosis and treatment of children;

(2) After training on a multi-center large sample data set, the model performance was evaluated using 8 evaluation indicators and 1 functional indicator, as well as internal validation, external validation, comparison of calibration curves, and subgroup analysis based on age stratification. The model performance was good, universal, and robust, and consistently superior to existing clinical scores;

(3) It can provide important factors and rankings associated with adverse outcomes in critically ill children, helping pediatricians understand the development of the disease;

(4) The number of characteristic variables of the input data can be selected to be 4 to 42 according to the actual application scenario, and the prediction evaluation performance that meets clinical needs can be obtained;

(5) The risk prediction device can automatically output timely risk assessment results and visualized risk reasoning processes for adverse outcomes (death) in critically ill children during their stay in the PICU, and can be incorporated into the hospital's electronic medical record information system to facilitate doctors' operation and use.

The interpretable risk assessment model and device for death in critically ill children of the present invention have been trained with a large sample size and multiple centers, and their performance is consistently better than other compared models and scores. They have shown good universality and robustness in multiple verification and evaluation methods, and can provide analytical reasons while providing the risk probability of adverse outcomes in children. Therefore, the device helps pediatricians obtain more timely and accurate assessments of the urgency and danger level of diseases in critically ill children, helps to make timely and accurate decisions on the treatment of patients, and is suitable for use in more children's medical institutions in different regions and centers.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flowchart of the method for establishing an interpretable mortality risk assessment model for critically ill children.

Figure 2 shows the inclusion and exclusion criteria and screening flow chart of the study population.

Figure 3 shows the receiver operating characteristic (ROC) curve, calibration curve, and precision-recall curve for the internal validation of the mortality risk prediction model.

Figure 4 shows the comparison between the risk prediction model and other machine learning models included simultaneously (internal validation ROC and Precision-Recall Curve).

Figure 5 is a comparison of other machine learning models and clinical disease scores included simultaneously (internal validation ROC and Precision-Recall Curve).

Figure 6 shows the receiver operating characteristic (ROC) curve, calibration curve, and precision-recall curve of the external validation of the mortality risk prediction model.

Figure 7 shows the comparison of the risk prediction model with other machine learning models and clinical disease scores included simultaneously (external validation: AUC comparison).

Figure 8 shows the age subgroup analysis (receiver operating characteristic curve and calibration curve) of the prediction model in internal validation.

Figure 9 shows the ranking of the top 20 important features for assessing disease risk of the prediction model based on the SHAP method and the SHAP value of the explanation feature's impact on the model prediction.

Figure 10 presents the inference interpretable analysis of the prediction model (non-surviving children).

Figure 11 presents the inference interpretable analysis of the prediction model (surviving children).

Figure 12 is a schematic diagram of the application of the PICU risk assessment model for critically ill children based on interpretable machine learning. The data in the figure are schematic results and do not affect the understanding of the technical effects of the present invention by those skilled in the art.

DETAILED DESCRIPTION

The present invention uses a multicenter prospective cohort study, based on the electronic medical records of critically ill children, and adopts a machine learning model that integrates interpretable methods to develop a critically ill children death risk prediction model that has been trained and verified by multiple centers and has universality, robustness and interpretability. The risk factors associated with adverse outcomes during PICU stay and the reasoning process of the model are obtained. Finally, it can be seen that the model is integrated into a device that can automatically and timely assess the death risk of critically ill children, such as a computer. The steps of the present invention are as follows: (1) constructing a large-sample multicenter data set that can support the development of a model with excellent evaluation performance, and constructing a research data set for critically ill children based on clinical diagnostic standards and clinical and literature knowledge; (2) cleaning and organizing the data, including data merging, data sampling, outlier removal, interpolation, and building statistical features, and constructing five types of data, including general information, vital signs, disease severity scores of critically ill children, laboratory tests, and treatment, based on the data collection and change characteristics; (3) training the models based on three machine learning models (random forest RF and extreme gradient boosting XGBoost and support vector machine SVM of ensemble learning model methods), and the present invention is the model with the best performance, and then the performance of the model is evaluated by 8 evaluation indicators (area under the receiver operating characteristic curve AUC, sensitivity, specificity, accuracy, precision, F1 value, area under the precision-recall curve AUPRC, calibration curve) and 1 functional indicator (interpretability). The model was compared with the scores commonly used in pediatric clinical practice (the third-generation pediatric mortality risk score PRISM Ⅲ and the pediatric sequential organ failure assessment pSOFA); (4) the universality and robustness of the prediction model were evaluated by internal validation, external validation, subgroup analysis (critically ill children aged ≤2 years and older than 2 years) and inclusion of some features (42 to 4); (5) risk assessment factors associated with adverse outcomes in critically ill children during PICU stay were obtained based on the interpretable method SHAP. The above process was encapsulated to obtain a fully automatic and timely assessment of the risk of adverse outcomes in children in the pediatric intensive care unit (PICU) that is easy for doctors to understand and operate, helping doctors to be more fully and timely aware of the potential body state of the children, thereby providing a scientific basis for further precise decision-making and treatment.

The model method of the present invention for timely assessing the risk of death of critically ill children specifically includes the following steps:

(1) Dataset construction module

Clinical medical data of critically ill children in PICUs from four representative children's medical centers in China were obtained. Critically ill children were obtained according to the following inclusion and exclusion criteria. Inclusion criteria: ① Age 1 month to 18 years old; ② Weight greater than 4 kg; ③ Meet the criteria for admission to the PICU. Exclusion criteria: ① Children admitted to the PICU for less than 24 hours, including children who died, transferred to other hospitals, or were discharged automatically after giving up treatment within 24 hours of admission to the PICU; ② Suffering from severe congenital developmental malformations. Further, based on clinical diagnostic criteria and clinical and literature knowledge, combined with the clinical characteristics of pediatric patients and the patient data recorded in the electronic medical record data, the features used for the subsequent development of the prediction model were determined. Including: general demographic information (a total of 5 dimensions), vital signs (a total of 5 dimensions), severity scores of critically ill children (a total of 3 dimensions), laboratory tests (a total of 32 dimensions), and treatment interventions (a total of 6 dimensions). Children who died during their stay in the PICU were marked as positive samples, and the rest were marked as negative samples.

(2) Data processing module

For the research population and research data set determined in step 1, the data from the four centers were cleaned and merged, including checking data consistency, handling invalid values and missing values: missing data were interpolated by median interpolation, and if the missing ratio was greater than or equal to 20%, they were eliminated; the independent variables with significant collinearity (Sperman correlation coefficient r>0.6 or collinearity test variance inflation factor VIF>10) and little contribution to the dependent variable were excluded by multicollinearity test. Based on the regularized data, the model input features were constructed, that is, statistical features (original value, mean, median, quartile, maximum value, minimum value, sum) were constructed, and 42 research features (3 demographic information, 2 critical children's disease severity scores, 5 vital signs, 24 laboratory indicators, and 8 treatment information) were obtained.

(3) Model building and evaluation module

This method uses the extreme gradient boosting XGBoost model, an integrated learning method that integrates the SHapley Additive exPlanations (SHAP) method, to timely evaluate the mortality risk of critically ill children; the model with the best performance is the XGBoost model. ① The present invention simultaneously incorporates three machine learning models (extreme gradient boosting XGBoost model, random forest RF model, support vector machine SVM model). The prediction model trains and tunes the model using multi-center data, and conducts internal validation evaluation; then 8 evaluation indicators (AUC, sensitivity, specificity, accuracy, precision, F1 value, AUPRC, calibration curve) and 1 interpretability function indicator are used to evaluate the model performance; further, data of critically ill children from different periods that are inconsistent with the training data set are used for external validation of the model; ② Performance evaluation of the model: The two machine learning models (random forest RF and support vector machine SVM models) and two commonly used clinical scores (the third-generation pediatric mortality risk score PRISM Ⅲ and pediatric sequential organ failure score pSOFA), and compared with the model selected by the present invention (XGBoos model); and compared the model fit by obtaining the calibration curve of the model, and further evaluated the impact of age performance, which is the focus of pediatric clinical attention, on the model prediction performance (severely ill children aged 2 years or less and older than 2 years old); ③ At the same time, the risk factors and rankings related to the death of critically ill children were obtained based on the SHAP method; and the changes in the predictive performance of the model by reducing the inclusion of feature variables were evaluated (all features 42 to 4 features). Finally, the relevant modules of the fully verified model and data processing link were encapsulated to obtain the death risk assessment model and device for critically ill children.

The present invention provides a death risk assessment model and device for PICU critically ill children with an interpretable function, which specifically comprises the following steps: obtaining 3 pieces of general demographic information of the children on the first day of PICU stay, 2 disease severity scores of critically ill children, 5 vital signs, 24 laboratory test indicators and 8 treatment information during PICU stay; the above-mentioned characteristic variables are processed by a data processing module of the device to obtain features that can be directly input into the model, and further calculated by a death risk assessment module and visualized by an interpretable method the model assessment process, i.e., the contribution ratio of important risk factors to the outcomes of the children; finally, the risk of adverse outcomes (death) of critically ill children and the explanation of model reasoning are obtained.

The present invention will be described in detail below with reference to the accompanying drawings. The specific data acquisition, input/output operations and algorithms are conventional technologies.

The invention proposes a risk assessment model that is robust, universal, clinically feasible, understandable and trustworthy, based on high-quality electronic medical record data from multiple centers, to timely and accurately assess and predict the risk of adverse outcomes in critically ill children during their stay in the PICU. The model is developed through comprehensive assessment indicators and external validation to obtain a risk assessment model that is robust, universal, clinically feasible, understandable and trustworthy. Combined with the characteristics of children, the severity of the disease in critically ill children can be fully and automatically assessed in a timely manner, providing a scientific basis for early intervention and accurate diagnosis and treatment.

The present invention utilizes the rich information collected by electronic medical records. These three-dimensional data can characterize the disease development trajectory of the child during his stay in the PICU. The complex nonlinear correlation between data and outcomes is mined through a machine learning model to obtain a prediction model with better performance than the disease scoring used in clinical practice, which can be used as a system. At the same time, since the data comes from multiple centers, a more universal model can be developed, which is an advantage that traditional linear addition clinical scoring cannot have. After multi-center and external validation and subgroup analysis, the model with the best performance is finally encapsulated and can be integrated into the existing hospital electronic medical record information system. The analysis results can be automatically obtained, and visual interpretation and risk factor ranking can be performed to provide a reference basis for pediatricians' treatment evaluation without increasing the workload of pediatric medical staff.

The process proposed in the present invention mainly includes three models: (1) a data set construction module, which obtains a research data set of multi-center critically ill children based on the inclusion and exclusion criteria and characteristic variables of the research population; (2) a data processing module: according to the original research data set obtained in step (1), the data is cleaned, regularized, sampled and interpolated. Further, the statistical features are constructed according to the data characteristics, and 5 types of research characteristic data are obtained; (3) a model construction and evaluation module, based on the data obtained in (2), it is input into the selected machine learning model, the model construction and parameter tuning are completed based on the selected model training set, and the performance evaluation of internal verification is performed. Subsequently, based on the external verification data set and the determined 8 evaluation indicators and 1 functional indicator, as well as the subgroup analysis and the inclusion of some features, the prediction performance of the model is further evaluated to obtain the robustness, universality and generalizability of the model. Then, the model program with the best performance is encapsulated and written into the computing device to obtain a model that can automatically help pediatricians obtain the mortality risk and risk factor ranking of children in a timely manner, so as to assist pediatricians in disease diagnosis and treatment.

The present invention proposes a PICU mortality risk assessment method for critically ill children based on a multi-center electronic medical record dataset. Its prediction performance is consistently better than the baseline model and clinical score, and can provide pediatricians with a more convenient and accurate assessment method for timely assessment of the condition of critically ill children. It is the first time that a risk assessment model has been constructed for critically ill children. It uses a multi-center dataset for model training and has been externally verified. The performance has shown good universality and robustness. At the same time, the method obtains a ranking of risk factors associated with the risk of death in critically ill children, among which the Sequential Organ Failure Score for Children, whether mechanical ventilation is used, lactate dehydrogenase, C-reactive protein, and alanine aminotransferase are very critical and are in the top 5 of the risk factor rankings. It plays an important role in evaluating the outcome of children. Finally, the method can automatically and conveniently assess the risk of adverse outcomes (death) in critically ill children during their stay in the PICU.

Embodiment 1

The present invention proposes a method for interpretable assessment of mortality risk and ranking of risk factors for critically ill children based on electronic medical records. The specific implementation is shown in FIG1 , and the specific steps are as follows.

Dataset construction module. Critically ill children from PICUs of four representative children's medical centers in China were used as the basic population, and the population involved in this study was further obtained according to the screening criteria (screening process) shown in Figure 2 (age 1 month-18 years, weight >4kg; and PICU stay ≥24 hours). In this cohort, a total of 947 critically ill children were included, and 79 (8.3%) died during their stay in the PICU. The general information of surviving and deceased children is compared in Table 1. This study was approved by the Medical Ethics Committees of the four children's hospitals, and the informed consent was signed by the guardians of the children.

Establishment of data set: Clinical medical data of critically ill children in PICU were collected. According to clinical diagnostic criteria and clinical and literature knowledge, combined with the clinical characteristics of pediatric patients and the patient data recorded in the electronic medical record data, the features used to develop the prediction model were determined. As shown in Table 2, it includes 5 demographics (age, weight, height, BMI index, gender), 3 critical illness severity scores for children (third-generation pediatric mortality risk score PRISM Ⅲ, pediatric sequential organ failure assessment pSOFA, Glasgow Coma Scale GCS), 5 vital signs (heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, body temperature), 34 blood oxygen indices and laboratory test indicators (oxygen partial pressure _PaO2 , inspired oxygen concentration _FiO2 , carbon dioxide partial pressure _PaCO2 , oxygenation index _PaO2 / _FiO2 , _SpO2 / _FiO2 , blood pH, bicarbonate HCO3, excess base BE, prothrombin time PT, partial thromboplastin time APTT, international normalized ratio INR, antithrombin III, troponin cTn, creatine phosphokinase isoenzyme CKMB, alanine aminotransferase ALT, aspartate aminotransferase AST, total bilirubin Tbil, lactate dehydrogenase LDH, creatinine SCr, urea nitrogen BUN, albumin Alb, glucose Glu, potassium ion K ⁺ , sodium ion Na ⁺ , chloride Cl ^- , calcium ion Ca ²⁺ , lactate Lactate, white blood cells WBC, neutrophils Neu, platelets PLT, procalcitonin PCT, C-reactive protein CRP, glomerular filtration rate eGFR), and 6 treatment methods (mechanical ventilation, renal replacement therapy, vasoactive drugs, furosemide, hormones, antibiotics). Table 2 also simultaneously presents the missing proportions of all study variables in the four clinical centers.

Age: age, AKI: acute kidney injury, BMI: body mass index, Body weight: weight, GCS: Glasgow Coma Scale; Male: male, MODS: multiple organ dysfunction, PICU: pediatric intensive care unit, PRISM Ⅲ: third-generation pediatric mortality risk score, pSOFA: sequential organ failure assessment for children, Sepsis: sepsis, Shock/DIC: shock/disseminated intravascular coagulation.

Data processing module. The raw data of the research population and the determined research variables obtained through the above process are input into the data processing module to complete the preparation work before model building. (1) Data cleaning, including checking data consistency, determining the unified name of each variable in the four centers, and merging multiple expressions of the same variable; (2) Processing invalid values and missing values. Data interpolation: The conditional median interpolation method was used to interpolate variables with a missing rate of less than 20% in the study population. Since the missing values were clinical laboratory indicators, most of them were due to the judgment of clinicians that no testing was needed, and the variables were non-normally distributed, the conditional median interpolation method was used: after the population was stratified according to age and gender, the median of the variable in the stratum to which the child belonged was used to replace the missing data; (3) Variables with a missing rate of more than 20% were eliminated; (4) Collinearity analysis: For variables with significant collinearity (Sperman correlation coefficient r>0.6 or collinearity test variance inflation factor VIF>10), stepwise regression analysis was used to eliminate independent variables with collinearity; (5) Statistical feature construction: Statistical features were extracted, and the names of the characteristic variables finally obtained are listed in Table 3: including 3 demographic information indicators, 2 critical illness severity scores for children, 5 vital sign indicators, 24 laboratory test indicators, and 8 treatments.

PRISM Ⅲ: the third-generation pediatric mortality risk score PRISM Ⅲ, pSOFA: pediatric sequential organ failure assessment.

Model construction and training: The model construction uses data from 80% of the research population. This study uses the extreme gradient boosting XGBoost model of ensemble learning, inputs the feature variables obtained in the data processing module, and on this basis, 80% of the data set is used for model training and model parameter tuning. The final model operation function and hyperparameter settings are:

model_use=xgboost.XGBClassifier (**params)

params={booster=gbtree,

eta=0.1, gamma=0.001,

max_depth=6, min_child_weight=1,

colsample_bytree=0.5,

objective=binary:logistic

subsample=0.85,

nrounds=100,

seed=0, silent=0,

watchlist=watchlist,

verbose=1,

print_every_n=100,

early_stopping_rounds=200,

num_boost_round=10}

explainer=shap.TreeExplainer(model_use)

20% of the study population was used for internal validation of the model. In order to compare the performance of the models later, three machine learning models (XGBoost, random forest RF and support vector machine SVM models) were trained simultaneously. Figure 3 presents the internal validation results of the mortality risk prediction model. ROC curve, precision-recall curve and calibration curve of the prediction model. Figures 4 and 5 present the internal validation results of the risk prediction model of the present invention compared with other machine learning models and clinical disease scores included simultaneously: the extreme gradient boosting model (XGBoost, the present invention) performs consistently better than the two machine learning models (random forest RF and SVM models) and clinical disease scores (the third-generation pediatric mortality risk score PRISM Ⅲ and the pediatric sequential organ failure score pSOFA). AUC: XGBoost (0.935), RF (0.894), SVM (0.912), clinical disease score (0.845).

The performance of the model was verified by external validation and subgroup analysis. Severely ill children at different periods that were inconsistent with the training data set were used as the evaluation population to establish an external validation data set. The selected model of the present invention was compared with the two machine learning models mentioned above (random forest RF and SVM models) and clinical disease scores (third-generation pediatric mortality risk score PRISM Ⅲ and pediatric sequential organ failure score pSOFA). Eight evaluation indicators were selected for quantitative and qualitative evaluation of the performance of the model and other comparative models/scores. Figure 6 presents the external validation results of the prediction model selected in this study. Prediction model ROC and precision-recall curve and calibration curve. The performance of the calibration curve of the external validation shows that the results of the model are well aligned with the y=x curve. Figure 7 presents the external validation results of the risk prediction model of the present invention compared with other machine learning models and clinical scores included simultaneously. The XGBoost model is significantly better than the clinical score. AUC: XGBoost (0.940), RF (0.951), SVM (0.911), PRISM III and pSOFA (0.892). Figure 8 shows the bias of the model in two age strata (severely ill children aged 2 years or less and older than 2 years). In the two age strata, the model prediction performance is slightly different. Table 4 shows the detailed performance of the seven indicators of internal and external validation of the prediction model (AUC, specificity, sensitivity, accuracy, precision, F1 value, AUPRC). Table 5 shows the prediction performance comparison of the prediction model of the present invention with the other two machine learning models and clinical scores.

Evaluation module, based on the SHAP method, obtains the ranking of risk factors of the prediction model in assessing disease risk. Figure 9 presents the top 20 important features for assessing disease risk based on SHAP values. The ranking of risk factors is: Pediatric Sequential Organ Failure Score pSOFA, Mechanical Ventilation MV, Lactate Dehydrogenase LDH, C-Reactive Protein CRP, Alanine Aminotransferase ALT, Third Generation Childhood Mortality Risk Score PRISM Ⅲ, Maximum Carbon Dioxide Partial Pressure PaCO ₂ max, Procalcitonin PCT, Minimum Body Temperature Tepmin, Platelet PLT, Blood Glucose Glu, Blood Creatinine SCr, Creatine Phosphokinase Isoenzyme CKMB, Body Mass Index BMI, Blood Calcium Ion Ca ²⁺ , Neutrophil Count Neu, Lactate Lactate, Total Bilirubin TBil, International Normalized Ratio INR, Blood Potassium Ion K ⁺ . Figures 10 and 11 present the analysis reasons for the disease severity assessment of 2 children (surviving and non-surviving children) based on the interpretable prediction model, presenting the risk factors and protective factors and their respective proportions in the assessment of the children's outcomes. Table 7 shows the results of the prediction performance of the model based on the above-mentioned feature rankings, including the top 42 (all), top 20, top 12, top 6 and top 4 important features. It can be seen that the performance has slightly decreased, but the performance of the model is still better than the commonly used clinical scores in most of the various evaluation methods.

AUC: area under the receiver operating characteristic curve, AUPRC: area under the precision-recall curve.

AUC: area under the receiver operating characteristic curve, AUPRC: area under the precision-recall curve; PRISM III: the third-generation pediatric mortality risk score PRISM III, pSOFA: sequential organ failure assessment for children.

Table 6 Feature ranking of prediction model based on SHAP method

Embodiment 2

The present invention discloses an interpretable critically ill child death risk assessment device, which can execute the above-mentioned interpretable critically ill child death risk assessment model; the SHAP method fused by the model obtains the risk factor contribution degree assessment of a single critically ill child; wherein, red is used to represent that the factor is currently in an abnormal state and has a harmful effect on the outcome of the child, and blue is used to represent that the factor is currently in a normal state and has no harmful effect on the outcome of the child, and the greater the SHAP value, the greater the degree of influence on the outcome. The computing unit can be a central processing unit, a single-chip microcomputer, etc. As shown in Figure 12, the above-mentioned data processing process, prediction model, and interpretable function are program-encapsulated to form a device that can automatically clean data, calculate, evaluate, and give analysis reasons; the interpretable critically ill child death risk assessment model of the present invention is written into a computing device as a program, such as a computer or a mobile phone, specifically a conventional technology. In clinical application, doctors can obtain risk assessment results based on the data (features) of specific patients, including the importance of these features to the prediction results.

Unless otherwise defined, all technical and/or scientific terms used in this application have the same meaning as those generally understood by those of ordinary skill in the art to which the present invention relates. The devices, methods and embodiments mentioned in this application are illustrative only and not restrictive. Although the present invention has been described in conjunction with specific embodiments, those skilled in the art may make appropriate substitutions, modifications and changes under the inventive spirit of this application, and such substitutions, modifications and changes still fall within the scope of protection of this application.

Claims (10)

1. An interpretable severe child death risk assessment model, characterized in that it comprises: (1) Data set construction module: obtain the data of children admitted to the PICU, and determine the characteristic variables; (2) Data processing module: clean up and merge the data in the data set, sample and interpolate, and obtain multiple statistical features; (3) Model construction and evaluation module: use the extreme gradient boosting XGBoost model that incorporates the SHAP method, conduct model training and parameter tuning on the processed data set, and build an interpretable critically ill child death risk assessment model; according to the prediction results The importance of multiple statistical features are ranked from high to low. 2. The explainable critically ill child death risk assessment model according to claim 1, wherein the characteristic variables include general data demographics, vital signs, critically ill child disease severity scores, laboratory tests and treatment. 3. The interpretable death risk assessment model for critically ill children according to claim 1, wherein the data processing module is: cleaning and merging the data of the data set, including checking data consistency, processing invalid values and missing values, through The median imputation method was used to interpolate missing data, and if the missing ratio was greater than or equal to 20%, it was eliminated; multicollinearity test was used to exclude independent variables with significant collinearity and little contribution to the dependent variable; based on the regularized data, the The construction of model input features, that is, the construction of statistical features, including original value, mean, median, quartile, maximum value, minimum value, and sum, obtains 4 to 42 statistical features. 4. According to claim 3, the interpretable death risk assessment model for critically ill children is characterized in that the statistical features include age, body mass, length, body mass index, gender, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, body temperature, The third generation child mortality risk score, child sequential organ failure score, Glasgow coma score, inspired oxygen concentration, oxygen partial pressure, carbon dioxide partial pressure, oxygenation index PaO ₂ /FiO ₂ , oxygenation index SpO ₂ /FiO ₂ , blood pH, bicarbonate, residual base, prothrombin time, partial thromboplastin time, international normalized ratio, antiprothrombin III, troponin, creatine phosphokinase isoenzyme, alanine aminotransferase, asparagine amino acid transaminase, total bilirubin, lactate dehydrogenase, creatinine, blood urea nitrogen, albumin, glucose, potassium ions, sodium ions, chloride, calcium ions, lactic acid, white blood cells, neutrophils, platelets, calcitonin Original, C-reactive protein, glomerular filtration rate, whether to use mechanical ventilation, whether to use renal replacement therapy, whether to use vasoactive drugs, vasoactive drug score, whether to use furosemide, whether to use hormones, whether to use antibiotics, Whether to use Mepine/Taineng, whether to use multiple types of vancomycin. 5. According to claim 4, the explainable severe child death risk assessment model is characterized in that, the evaluation module is based on the extreme gradient lifting XGBoost model that incorporates the SHAP method; among the multiple features, according to its importance to the predicted results, it is composed of The top 20 statistical features from high to low are: Pediatric Sequential Organ Failure Score, Mechanical Ventilation, Lactate Dehydrogenase, C-Reactive Protein, Alanine Aminotransferase, Third Generation Child Mortality Risk Score, Highest partial pressure of carbon dioxide , procalcitonin, minimum body temperature, platelets, blood glucose, serum creatinine, creatine phosphokinase isoenzyme, body mass index, serum calcium ion, neutrophil count, lactic acid, total bilirubin, international normalized ratio, serum potassium ion. 6. A method for establishing an interpretable death risk assessment model for critically ill children, including data set construction, data processing, model construction and evaluation; characterized in that, in the data set construction, the data set of children admitted to the PICU is obtained and determined Feature variables, the feature variables include general demographics, vital signs, critically ill children disease severity score, laboratory examination and treatment; in data processing, the data from the data set are cleaned and integrated, sampled and interpolated to obtain multiple In the model construction and evaluation, the data set is based on the training and parameter tuning of the extreme gradient boosting XGBoost model combined with the SHAP method, and the multiple statistical features are ranked from high to low according to their importance to the prediction results. arrangement. 7. The method for establishing an interpretable critically ill child death risk assessment model according to claim 6, wherein the research population is based on the PICU income standard and the established inclusion process to screen children and obtain a data set for critically ill children; In model construction and evaluation, the model is trained and tuned for critically ill children. When training the model, the data is fused as a large-sample, multi-center training set, of which 70-90% of the children's data is used for the model. The hyperparameters of the predictive model were trained and tuned using cross-validation, and the remaining patient data were used for internal validation of model performance. 8. The application of the explainable critically ill child death risk assessment model described in claim 1 in explainable critically ill child death risk assessment or in the preparation of an explainable severe child death risk assessment device. 9. A computer-readable carrier, comprising a calculation program, characterized in that the calculation program is used to implement the interpretable death risk assessment model for critically ill children according to claim 1. 10. An interpretable death risk assessment device for critically ill children, comprising a calculation unit, characterized in that the calculation unit is used to implement the explainable death risk assessment model for critically ill children according to claim 1.