CN110827993A - Early death risk assessment model establishing method and device based on ensemble learning - Google Patents

Abstract

The present application proposes a method and device for establishing an early death risk assessment model based on ensemble learning, which is used to assess the death risk of a patient during hospitalization based on the diagnosis and treatment data of an elderly patient with multiple organ failure on the first day of entering an intensive care unit; the method includes: Data set construction, data processing, model construction and evaluation; by obtaining 3 demographic information, 5 vital sign information, 5 laboratory test indicators and 2 clinical indicators on the first day of the patient's stay in the intensive care unit; It is input into the risk assessment device, and after internal data preprocessing, feature calculation and model operation, it can finally predict the risk of adverse outcomes in the patient's hospital at an early stage, and assist doctors in early intervention and treatment of patients.

Description

Method and device for establishing early death risk assessment model based on ensemble learning

technical field

The invention relates to a death risk assessment method and device, in particular to a method and device for establishing an early death risk assessment model for elderly multiple organ failure based on integrated learning.

Background technique

Multiple organs dysfunction syndrome (MODS) is characterized by the occurrence of progressive physiological dysfunction of two or more organs (organ systems) triggered by a certain incentive. Multiple organ failure in the elderly (MODSE) is a clinical syndrome with insidious onset, multiple etiologies, complex pathogenesis, and easily overlooked by clinicians. It is an important cause of death in elderly critically ill patients. According to relevant literature reports, if three or more organ failure occurs, the mortality rate is 50% to 100%. At the same time, its treatment and nursing costs are very huge. According to statistics, the average medical cost of each patient is about 22,000 US dollars, and the national annual total expenditure is about 16.7 billion US dollars, which brings a heavy burden to families, hospitals and society. Early diagnosis and timely intervention are important ways and methods to reduce its mortality.

SOFA (Sepsis Related Organ Failure Assessment) is a score commonly used to evaluate systemic infection-related organ failure, including respiratory, neurological, cardiovascular, liver, and coagulation systems. The modified multiple organ dysfunction score (MODS) was used to evaluate the daily changes of organ failure in patients admitted to the intensive care unit (ICU). The Simplified Acute Physiology Score (SAPS) is often used for critical illness severity. Acute physiology and chronic health evaluation scoring system (APACHE-IV) is currently the most widely used and authoritative critical disease evaluation system in ICU. However, the above scoring features are usually modeled by logistic regression and do not take into account the association between organ functions. It is assumed that there is a linear superposition relationship between predictors and outcomes, and the complexity of real disease deterioration and evolution is far from being expressed and quantified by a simple linear model; the weights/coefficients assigned to each sub-score are highly subjective, that is, Selecting and assigning weights to the corresponding variables based on the knowledge of the mortality prediction correlation of the expert group has not been verified by a large number of experimental data; the elderly have many underlying diseases, complications, organ aging and functional decline, etc. , the evaluation criteria for adults do not apply to the elderly. Few of these scores provide a more detailed assessment scheme for elderly patients.

Electronic healthcare records (EHR) are the systematic collection of health information of patients and populations in a digital format. It contains patient demographics, medical history, medications and allergies, laboratory test results, vital signs and billing details. In recent years, with the popularization and coverage of EHR in medical office affairs and the development and application of big data mining technology, it is no longer a good vision to use EHR to improve medical quality, improve medical safety, and optimize medical process. EHR-based predictive analysis of disease severity/adverse outcomes has now become a research hotspot in academia and industry. In 2014, the non-parametric, integrated learning algorithm Super ICULearner developed by Pirracchio R et al. can more accurately assess the mortality risk of patients in the ICU at an early stage. The AUC of the model is as high as 0.88. In 2015, Henry, K.E et al. developed a "Targeted Real-Time Early Warning Score" (TREWScore), which predicted septic shock 28.2 hours ahead (on average) using a Cox proportional hazards regression model with an AUC of 0.83 and a sensitivity of 0.85. The 'Previse' automatic diagnosis algorithm developed by Dascena in 2017 can predict acute kidney injury a day before patients reach clinical diagnostic criteria, providing clinicians with sufficient time to intervene and prevent long-term damage. The algorithm predicts acute kidney injury 48 hours in advance 84% accuracy.

XGBoost (eXtreme Gradient Boosting) is a boosting algorithm whose idea is to integrate many weak classifiers (tree models) together to form a strong classifier (boosted tree model). The model has high accuracy in solving prediction problems, and also has the advantages of fast running speed, preventing overfitting and modeling sparse data. Therefore, it is widely used in data competition, scientific research and teaching, and industrial applications.

SUMMARY OF THE INVENTION

The purpose of this application is to propose a method and device for establishing an early death risk assessment model based on ensemble learning, which can use the diagnosis and treatment data of elderly patients with multiple organ failure to enter the intensive care unit on the first day to assess the death risk of patients during hospitalization.

The method for establishing an early death risk assessment model based on ensemble learning of the present application, the early death risk assessment model is used to assess the death risk of the patient during hospitalization based on the diagnosis and treatment data of the elderly patient with multiple organ failure on the first day of entering the intensive care unit; The evaluation model is an XGBoost model; the method includes:

The step of constructing the dataset; using the dataset building module, identifying the patients with multiple organ failure in the first database according to the clinical diagnosis definition, determining the study population and four types of characteristics included in the study, and forming the first dataset;

The steps of data processing; using the data extraction and processing unit of the data processing module to perform data extraction and processing in the first data set to form a second data set, the second data set is divided into a training set, a verification set and a test set; using the data The feature construction unit of the processing module performs feature construction on the data to be input into the risk assessment model, so that it meets the data input requirements of the risk assessment model; uses the patient outcome labeling unit of the data processing module to label patient outcomes;

Model construction and evaluation steps; use the XGBoost model as an early model for training, that is, through the early stop mechanism, use AUC as a performance evaluation indicator, set the number of early stop rounds, and test the performance of the model on the validation set; If the performance of the continuous predetermined number of iterations is no longer improved, the training is terminated to obtain the optimal parameters for model training; and the test set data is used to evaluate the prediction performance of the trained model to prevent the occurrence of model overfitting;

The final model was used as an early mortality risk assessment model.

Preferably, the four categories of characteristics identified for inclusion in the study are:

Demographic characteristics, including: type of admission, age, BMI, ethnicity, type of ICU admission, gender, height, weight;

Vital signs, including: central venous pressure, diastolic blood pressure, heart rate, mean arterial pressure, respiratory rate, systolic blood pressure, shock index, blood oxygen saturation, body temperature;

Laboratory findings including: albumin, alkaline phosphatase, alanine aminotransferase, aspartate aminotransferase, base excess, bicarbonate, bilirubin, B-type natriuretic peptide, blood urea nitrogen, chloride, creatinine, fiber Proteinogen, inspired oxygen concentration, blood glucose, hematocrit, hemoglobin, international normalized ratio, lactate, lymphocytes, magnesium, neutrophils, arterial blood carbon dioxide partial pressure, oxygenation index, blood oxygen partial pressure, ph value, Platelet, potassium, prothrombin time, thromboplastin time, serum sodium, troponin, white blood cell count;

Clinical characteristics, including: Glasgow score, systemic inflammatory response syndrome score, neurological and respiratory score of systemic infection-related organ failure score, whether mechanical ventilation, continuous renal replacement therapy, total urine output, Norepinephrine use rate.

Preferably, the data extraction and processing performed by the data extraction and processing unit of the data processing module includes: data cleaning, data sampling, and data interpolation;

The feature construction performed by the feature construction unit of the data processing module includes: the original value of the demographic characteristics; the maximum value, minimum value and mean value of the vital sign characteristics; the maximum value, minimum value, original value, clinical value of the laboratory test characteristics Maximum, minimum, mean, sum, original value of the feature.

Preferably, in the performance evaluation, the performance of the model is evaluated through internal verification and external verification, and the performance metrics are AUC, specificity, sensitivity, accuracy, F1 value, robustness, universality, performance The baseline model and clinical scores were evaluated against the baseline; the final model was able to obtain a ranking of risk factors associated with death based on a feature ranking function.

Preferably, the benchmark model is selected from: logistic regression LR, support vector machine SVM, neural network NN, random forest RT, naive Bayesian NB model;

The clinical score is selected from: Oxford Acute Illness Severity Score (OASIS), Acute Physiology and Chronic Health Score (APACHE-IV), Acute Physiological Assessment Score (APSIII), Multiple Organ Dysfunction Syndrome Score (MODS), Systemic Infection-Associated Organ Failure Score (SOFA), Simplified Acute Physiological Score (SAPS), Modified Early Warning Score (MEWS), Systemic Inflammatory Response Syndrome Score (SIRS), Rapid Sequential Organ Failure Score (qSOFA), Charlesson Combination Charlson Comorbidity score.

Preferably, robustness and generalizability are evaluated through the following three aspects:

Internal validation: through the validation set and test set of the second data set, compare the changes in AUC, specificity, sensitivity, accuracy, and F1 value to obtain model parameters with optimal performance and prevent overfitting;

External validation: The second database data set corresponding to the data set is obtained from the multi-center large sample data set in the same data processing method, and the changes in AUC, specificity, sensitivity, accuracy, and F1 value are compared and compared with the internal data set. Compare the verification results;

Reduce input features: By reducing the number of features input to the final model, evaluate the AUC changes of the final model in internal and external validation compared to the input of all features, and compare with clinical scores.

Preferably, the diagnosis and treatment data of the elderly patient with multiple organ failure on the first day of entering the intensive care unit include:

Demographic characteristics: age, weight, BMI;

Vital signs: minimum systolic blood pressure, mean systolic blood pressure, maximum shock index, minimum shock index, mean respiratory rate, minimum blood oxygen saturation, and maximum body temperature;

Laboratory test indicators: minimum blood urea nitrogen, minimum blood sugar, maximum chloride, maximum partial pressure of oxygen, and minimum white blood cell count;

Clinical features: sum of urine output, minimum urine output, maximum Glasgow score, mean Glasgow score.

The integrated learning-based early death risk assessment device of the present application evaluates the death risk of the patient during hospitalization based on the diagnosis and treatment data of the elderly patient with multiple organ failure on the first day of entering the intensive care unit; it includes: a data preprocessing module, a feature calculation module module, model operation module;

The data preprocessing module is realized by a computer; the data preprocessing module is used for extracting and processing the diagnosis and treatment data to obtain preprocessing data;

The feature calculation module is realized by a computer; the feature calculation module is used for feature construction on the preprocessed data to form data suitable for inputting the model calculation module;

The model calculation module is realized by a computer; the model calculation module includes an early death risk assessment model, and the early death risk assessment model is obtained by using the data set in the database for training, tuning and evaluation based on the XGBoost model, and its performance After the effective verification and evaluation of the multi-center large sample data set; the data constructed by the feature is input into the model operation module, and the model operation module outputs the risk prediction results of patients during hospitalization.

Preferably, the model included in the model operation module is an early death risk assessment model established by the method for establishing an early death risk assessment model based on ensemble learning according to any one of claims 1-7.

Preferably, the extraction and processing by the data preprocessing module include cleaning, sampling and interpolation of the diagnosis and treatment data.

The method and device for establishing an early death risk assessment model based on integrated learning of the present application,

(1) The probability of in-hospital adverse outcomes of MODSE patients can be predicted early, so as to assist doctors in early intervention and treatment of patients;

(2) Validated on both internal and external (large-sample, multi-center) datasets, the model performs well and outperforms the baseline model and clinical existing scores. At the same time, reducing the number of model input parameters to 20 still has good prediction performance. The prediction model can be easily deployed in the hospital information system;

(3) It can provide important factors associated with adverse outcomes to help doctors understand the development of the disease;

(4) Only the clinical information of 15 kinds of patients is input, and the risk prediction device can automatically output the early risk assessment results of in-hospital adverse outcomes (death) of patients.

Description of drawings

Fig. 1 is the execution flow of the early death risk assessment model establishment method based on ensemble learning of the present application;

Figure 2 shows the inclusion process of MIMIC-III middle-aged and elderly patients with multiple organ failure;

Figure 3 is a visualization of the disease development trajectories during the ICU stay of MODSE dead and surviving patients;

Figure 4 is the ROC curve of the early prediction patient mortality risk model;

Figure 5 shows the performance comparison between the prediction model and the baseline model;

Figure 6 shows the performance comparison between the prediction model and the commonly used clinical score;

Figure 7 shows the importance ranking of the top 30 features of the prediction model;

Figure 8 is the ROC curve of the prediction model using the top 20 important features;

Figure 9 shows the performance comparison of the prediction model using the top 20 important features and the clinical score;

Figure 10 shows the inclusion process of the prediction model in the eICU multicenter database for MODSE;

Figure 11 shows the ROC curve of the prediction model validated in eICU;

Figure 12 shows the comparison of the performance of the prediction model in the eICU and the commonly used clinical score;

Figure 13 shows the ROC curve of the prediction model with the top 20 features selected in eICU;

Figure 14 shows the comparison between the validation and clinical scoring performance of the prediction model with the top 20 features selected in the eICU;

Figure 15 is a schematic structural diagram of an early death risk assessment device for elderly multiple organ failure based on integrated learning;

Figure 16 is a gradient boosted decision tree for the prediction model.

Detailed ways

The method and device for establishing an early death risk assessment model based on ensemble learning of the present application will be described in detail below with reference to the accompanying drawings.

The development of a disease/outcome prediction model and device based on electronic health records proposed by the present invention is mainly used for early prediction of the probability/risk of adverse outcomes in elderly patients with multiple organ failure during hospitalization. The implemented risk assessment model can automatically assess the disease severity of elderly patients with multiple organ failure early and help doctors to intervene and treat patients at risk of deterioration as soon as possible. The invention utilizes the large sample data accumulated in the electronic health records for many years, can develop the model quickly, effectively and at low cost, and effectively solves the problems of time-consuming, labor-intensive and huge cost of clinical randomized controlled research; The verification of the central intensive care data set, its universality and robustness have been effectively verified; and the method can still maintain good prediction performance with fewer input parameters; the method is finally encapsulated, which can automatically calculate the incidence of patients Risk (probability) of adverse outcomes.

The method proposed in the present invention generally includes three modules: (1) a data set building module; (2) a data processing module: extracting, processing and characterizing data according to the research group determined in step (1) and the characteristics of the included research Construction; (3) Model construction and evaluation: According to the data set obtained in step (2), the learning/training and verification of the model are performed. Among them, step (1) mainly uses the criteria for clinical assessment of whether patients have multiple organ failure to mark MODSE patients in the MIMIC-III database; The III database obtains the required information. In order to facilitate the model calculation and learning to enrich the information, it is necessary to extract statistical features; step (3) mainly conducts model training, optimization and internal verification according to the data obtained in step (2), and further uses the eICU database to Models are externally validated.

The electronic health record-based early death risk assessment method for multiple organ failure in the elderly proposed in the present invention has better predictive performance than baseline models and commonly used clinical scores, and provides a more accurate method for assessing the deterioration risk of patients; its universality and robustness have been validated on the largest sample dataset to date (to our knowledge); and with reduced input to the model, the model performs well and outperforms clinical scores, and the model's generalizability and robustness are differentiated At the same time, the method can obtain the ranking of risk factors associated with disease deterioration, which helps doctors to have a deeper understanding of the development of MODSE disease; finally, the method has built-in parallel computing, which can automatically and quickly obtain the patient's hospitalization period. risk of early mortality.

The specific implementation of an electronic health record-based early death risk assessment method for multiple organ failure in the elderly proposed by the present invention is shown in Figure 1, including the following steps:

One, the data set building module process in the present invention is as follows:

First, MODSE patients in the MIMIC-III database were calibrated according to the clinical score of SOFA ≥ 2 (and at least two systemic abnormalities) for clinical diagnosis of multiple organ failure and the international definition of the elderly (age ≥ 65 years); then, according to Fig. The MODSE patient inclusion process in 2 identified the population for the study using this method from the database. The specific inclusion criteria are: ① age ≥ 65 years old, ② ICU length of stay ≥ 24 hours, ③ first hospitalization and ICU admission for the first time, ④ MODS (i.e. SOFA ≥ 2 points, and at least two organ systems) occurred during hospitalization failure occurred), ⑤ measured at least once within the first day in the ICU: heart rate (HR), respiratory rate (RR), mean arterial pressure (MAP), Glasgow score (GCS), body temperature (T), blood oxygen saturation (SpO ₂ ). A total of 15,804 patients (2,353 patients died) were finally included through the conditions ① to ⑤; finally, the characteristics of the study needed to be determined. We combined the experience and knowledge of clinicians and some literature to determine a total of 4 types of data to be included, namely: demographics , vital signs (added shock index shock index, namely pulse/systolic blood pressure), laboratory tests, clinical (score, medical intervention, output), the specific content will be described in detail in (2). In addition, we selected 7 important indicators of clinical interest to visualize the disease progression trajectories of MODSE in-hospital deceased and surviving patients, as shown in Figure 3.

Two, the data processing module process in the present invention is as follows:

First, based on the included population and study characteristics confirmed in (1), write Postgresql statements from MIMIC-III data;

All the relevant information of the patient in the first 24 hours after entering the ICU is extracted from the database, which is divided into the following three steps: ① Standardize the format of the data (for example: the itemid of glucose is 50809 or 50931, the itemid of heart rate is 211 or 220045), and further It is necessary to clean the information that contains both strings and numerical values, and remove outliers/outliers according to the physiological limit range of each indicator given by the doctor; Signs and clinical indicators are down-sampled (if there are multiple values within one hour, the average value is taken, among which the urine volume is special, and the sum within one hour is used); ③ The down-sampled data is further interpolated. When the missing proportion of the entire study population was ≤30%, the population median was imputed. If the missing ratio is >30%, insert the mean value of the population, and add the corresponding marker index of the feature (ie 0/1 label, representing untested/measured);

Next, feature construction is performed on the sorted data, as shown in the observation window in Figure 3 (the first 24h of ICU admission). The information in the observation window of all patients is extracted, as follows: ①Demographic characteristics (8): keep the original value; ②Vital sign statistical characteristics (25): maximum value, minimum value, mean value; ③Laboratory Check statistical features: maximum, minimum and label flag (eg: lactate_flag is lactate measurement (0/1)); clinical (16): maximum, minimum, mean, sum (urine volume) and label flag (eg : ventilation_flag is whether to perform mechanical ventilation (0/1)). Table 1 is a summary table of characteristic statistics;

Finally, in order to facilitate the training of the model, the outcome of the patient is annotated (0: survived, 1: died).

Table 1 Included features and outcome annotations in the prediction model

3. The process of model construction and evaluation module in the present invention is as follows:

First, based on the sorted data obtained in (2), it is input into our selected model XGBoost. XGBoost is an ensemble model, which is fast in operation and has a certain degree of interpretability. Divide the dataset into 80% training set and validation set and 20% test set, of which 80% training set and validation set will be further divided into 80% training set and 20% validation set, using cross-validation method to optimize Hyperparameter settings of the XGBoost model;

Next is the performance evaluation of the prediction model XGBoost. The evaluation indicators used in this method are AUC, specificity, sensitivity, accuracy, F1 value and robustness; the reference standard used is the baseline model (LR, NN, SVM, RF, NB ) and commonly used clinical scores to assess the severity of critically ill patients (OASIS, APACHE-IV, MODS, SOFA, SAPS, MEWS, SIRS, qSOFA, Charlson Comorbidity); The model was trained and evaluated based on 20% of the test set) and external validation (i.e. the predictive model developed on the MIMIC-III dataset was evaluated for performance across the entire study population in the eICU dataset). The following is a more detailed explanation of this process by combining the forecast results:

1) Interpretation of the internal validation results of the XGBoost prediction model

Figure 4 shows the ROC curve of the XGBoost prediction model (using all 107 features), and the AUC is 0.866. Figure 5 is the ROC curve graph comparing the prediction performance of this method and the five baseline models, and XGBoost performs the best. Table 2 shows the specific indicators of performance comparison. The AUC of XGBoost is 0.866, the sensitivity is 0.774, the specificity is 0.808, the accuracy is 0.877, and the F1 value is 0.680. Figure 6 is a ROC curve graph comparing the performance of this method with 10 commonly used clinical scores, and the performance of XGBoost is also consistently the best. Table 3 shows the specific indicators compared with clinical performance. The XGBoost prediction model can provide the model feature importance ranking, which provides the model with the function of interpretability, that is, the feature importance ranking provides clinicians with indicators that they should focus on in the early stage of deterioration in MODSE patients. Figure 7 shows the importance ranking of the top 30 features, including the maximum value of Glasgow score (gcs_max), the mean value of respiratory rate (rr_mean), the mean value of Glasgow score (gcs_mean), the mean value of systolic blood pressure (sbp_mean), the mean value of total urine Volume (uo_sum), body weight (weight), minimum value of blood urea nitrogen (bun_min), age (age), mean value of urine output (uo_mean), and maximum value of shock index (si_max) ranked in the top 10. Table 3 lists the names and specific importance values of the top 30 important features. In addition, in order to verify whether the model can be applied to different medical resource allocation institutions, we limit the number of input features of the model to 20 (@20). Figure 8 shows the ROC curve of this method, and the AUC is 0.857; Figure 9 shows the reduction ROC curves of the performance of the input feature model compared to 2 well-performing clinical scores (AUC _OASIS = 0.752, AUC _APACHE-IV = 0.704), although the prediction performance of our method decreases slightly compared to the input full features, it is still far Better performance than clinical scoring.

Table 2 Performance comparison between XGBoost prediction model and other baseline models

model name AUC Sensitivity specificity accuracy F1 value XGBoost 0.866 0.774 0.808 0.877 0.680 NN 0.848 0.802 0.743 0.867 0.601 LR 0.847 0.841 0.703 0.871 0.674 SVM 0.834 0.763 0.773 0.853 0.460 RF 0.795 0.634 0.819 0.859 0.597 NB 0.778 0.696 0.742 0.819 0.664

Table 3 Performance comparison between the XGBoost prediction model and the commonly used clinical scores

Table 4 The top 30 important features of the XGBoost prediction model

2) Interpretation of the external validation results of the XGBoost prediction model

The content of this part is used to verify the universality and robustness of the MODSE early mortality risk assessment model developed based on the MIMIC-III data set through the multi-center large sample data set eICU. The first is the construction process of the dataset module. It is necessary to obtain the population included in the eICU dataset through the same process as the study population obtained in MIMIC-III. The specific inclusion process is shown in Figure 10. The final included population is 34,523 MODSE patients, of which 3,966 died. . Consistent with the MIMIC-III dataset, the same research characteristics are also included; followed by the data processing module, which is consistent with the process in the MIMIC-III dataset. Finally, the external validation of the model, our method is validated in the eICU dataset, the ROC curve of the model is shown in Figure 11, where the AUC is 0.837, which shows that the universality/robustness of the model has been well validated. Figure 12 shows the ROC curve comparing the predictive performance of this method with the clinical score (OASIS, APACHE-IV). Our method is still far superior to the clinical score (AUC _OASIS = 0.746, AUC _APACHE-IV = 0.742). We still select some input features (top 20 important features) to evaluate the generality/robustness of the model. Figure 13 shows the results of external validation of the method of inputting model features in the eICU dataset, with an AUC of 0.828. Figure 14 is a ROC curve versus clinical score versus performance. Combining Figures 12 and 14, it can be seen that although the performance of the model model is slightly modeled, the performance of the model is still good.

Table 5 shows a summary of the performance comparisons of the prediction model with full input, the prediction model with partial input (@20), and 10 clinical scores.

Table 5 Validation and clinical score comparison of the prediction model in the eICU dataset

name AUC Sensitivity specificity XGBoost_all_features 0.837 0.763 0.752 XGBoost_20_features 0.828 0.733 0.769 MODS 0.763 0.719 0.667 OASIS 0.746 0.673 0.694 APACHE-IV 0.742 0.635 0.751 APSIII 0.736 0.663 0.716 SOFA 0.731 0.684 0.673 MEWS 0.719 0.614 0.712 SAPS 0.693 0.569 0.717 SIRS 0.670 0.734 0.531 qSOFA 0.675 0.597 0.710 Charlson Comorbidity 0.553 0.5229 0.569

The fully automatic risk assessment device proposed in the present invention is shown in Figure 15. We take into account the prediction performance (specificity, sensitivity, accuracy, etc.) and applicability of the model, and select the top 20 features with feature importance (see Table 4) as the input information of the final prediction model. The hyperparameter settings of the model is {'base_score':0.5,'booster':'gbtree','colsample_bylevel':1,'colsample_bytree':1,'gamma':0,'learning_rate':0.1,'max_delta_step':0,'max_depth': 3,'min_child_weight':1,'missing':None,'n_estimators':100,'n_jobs':-1,'nthread':None,'objective':'binary:logistic','random_state':0,' reg_alpha':0,'reg_lambda':1,'scale_pos_weight':1,'seed':None,'silent':1,'subsample':1,'eval_metric':'auc','tree_method':'approx' }.

In order to understand the model more intuitively, the gradient decision lifting tree of the prediction model is drawn, as shown in Figure 16.

Industrial use possibility

The method and device for establishing an early mortality risk assessment model based on ensemble learning of the present application fully automatically encapsulates data preprocessing, feature calculation and model calculation, so that the required diagnosis and treatment information of a single patient/multiple patients can be received / data, it is possible to automatically and quickly obtain whether patients will experience adverse outcomes during hospitalization. Helps physicians conduct a more comprehensive assessment of the severity of a patient's disease, leading to early care and treatment of high-risk patients.

Claims (10)

1. An early death risk assessment model building method based on ensemble learning, wherein the early death risk assessment model is used for assessing the death risk of an aged multi-organ failure patient during hospitalization based on the diagnosis and treatment data of the first day when the aged multi-organ failure patient enters an intensive care unit; the evaluation model is an XGboost model; the method comprises the following steps:

a step of constructing a data set; identifying, with a dataset construction module, multiple organ failure patients in a first database according to a clinical diagnosis definition, determining four types of characteristics for study population and study inclusion, forming a first dataset;

a step of data processing; performing data extraction and processing in the first data set by using a data extraction and processing unit of the data processing module to form a second data set, wherein the second data set is divided into a training set, a verification set and a test set; performing feature construction on data to be input into the risk assessment model by using a feature construction unit of the data processing module so as to enable the data to meet the data input requirement of the risk assessment model; marking the patient outcome by using a patient outcome marking unit of the data processing module;

constructing and evaluating a model; training by using an XGboost model as an early prediction model, setting the number of early stop rounds by using an early stop mechanism and AUC as a performance evaluation index, and testing the performance of the model on a verification set; terminating the training when the iteration performance of the model represented on the verification set for continuous preset times is not improved any more, and obtaining the optimal parameters of the model training; evaluating the prediction performance of the trained model by adopting the test set data so as to prevent the model from over-fitting;

the final model was used as an early mortality risk assessment model.

2. The ensemble learning-based early mortality risk assessment model building method of claim 1, wherein:

four types of features included in the identified studies are:

demographic characteristics, comprising: admission type, age, BMI index, race, type of intensive care unit admitted, sex, height, weight;

vital signs features, comprising: central venous pressure, diastolic pressure, heart rate, mean arterial pressure, respiratory rate, systolic pressure, shock index, blood oxygen saturation, body temperature;

a laboratory examination feature, comprising: albumin, alkaline phosphatase, glutamic-pyruvic transaminase, glutamic-oxalacetic transaminase, residual alkali, bicarbonate, bilirubin, B-type natriuretic peptide, blood urea nitrogen, chloride, creatinine, fibrinogen, inhaled oxygen concentration, blood glucose, hematocrit, hemoglobin, international normalized ratio, lactic acid, lymphocytes, magnesium, neutrophils, arterial blood partial pressure of carbon dioxide, oxygenation index, blood oxygen partial pressure, ph value, platelets, potassium, prothrombin time, thromboplastin time, blood sodium, troponin, white blood cell count;

a clinical profile, comprising: glasgow score, systemic inflammatory response syndrome score, nervous system and respiratory system score for systemic infection-related organ failure score, whether mechanical ventilation was performed, whether continuous renal replacement therapy was performed, total urine output, norepinephrine use rate.

3. The ensemble learning-based early mortality risk assessment model building method according to claim 2, wherein: the data processing step comprises:

the data extraction and processing performed by the data extraction and processing unit of the data processing module comprises: data cleaning, data sampling and data interpolation;

the feature construction by the feature construction unit of the data processing module includes: raw values of demographic characteristics; maximum value, minimum value and mean value of vital sign features; maximum, minimum, raw value for a laboratory test feature, maximum, minimum, mean, sum, raw value for a clinical feature.

4. The ensemble learning-based early mortality risk assessment model building method according to claim 3, wherein:

during performance evaluation, the performance of the model is evaluated through an internal verification mode and an external verification mode, the measurement indexes of the performance are AUC, specificity, sensitivity, accuracy, F1 value, robustness and universality, and the reference standard of the performance evaluation is a reference model and clinical scores; the final model may obtain a mortality-associated risk factor ranking based on a feature ranking function.

5. The ensemble learning-based early mortality risk assessment model building method of claim 4, wherein:

the reference model is selected from: a logistic regression LR, a support vector machine SVM, a neural network NN, a random forest RT and a naive Bayes NB model;

the clinical score is selected from: oxford acute disease severity score (OASIS), acute physiology and chronic health score (APACHE-IV), acute physiology assessment score (apsiiii), multiple organ dysfunction syndrome score (MODS), systemic infection-related organ failure Score (SOFA), Simplified Acute Physiology Score (SAPS), improved early warning score (MEWS), systemic inflammatory response syndrome score (SIRS), rapid sequential organ failure score (qSOFA), Charlson syndrome index score (Charlson organization).

6. The ensemble learning-based early mortality risk assessment model building method of claim 4, wherein:

robustness and universality are evaluated by the following three aspects:

internal verification: comparing the change conditions of AUC, specificity, sensitivity, accuracy and F1 values through the verification set and the test set of the second data set to obtain the model parameters with optimal performance and prevent overfitting;

external verification: acquiring a second database data set corresponding to the data set by using a multi-center large sample data set in the same data processing mode, comparing the change conditions of AUC, specificity, sensitivity, accuracy and F1 values, and comparing the change conditions with an internal verification result;

reducing input features: by reducing the number of features input into the final model, the AUC change of the final model is evaluated for internal and external validation compared to the input of all features, and compared to the clinical score.

7. The ensemble learning-based early mortality risk assessment model building method of claim 1, wherein:

the diagnosis and treatment data of the aged multi-organ failure patient entering an intensive care unit for the first day comprise:

demographic characteristics: age, weight, BMI;

vital sign characteristics: the blood oxygen saturation control method comprises the following steps of (1) systolic pressure minimum value, systolic pressure mean value, shock index maximum value, shock index minimum value, respiratory rate mean value, blood oxygen saturation minimum value and body temperature maximum value;

laboratory examination indexes: blood urea nitrogen minimum, blood glucose minimum, chloride maximum, blood oxygen partial pressure maximum, white blood cell count minimum;

the clinical characteristics are as follows: total urine volume, minimum urine volume, maximum glasgow score, mean glasgow score.

8. An integrated learning-based early death risk assessment device for assessing the death risk of an elderly multi-organ failure patient during hospitalization based on first day clinical data of the patient entering an intensive care unit; it includes: the system comprises a data preprocessing module, a characteristic calculation module and a model operation module;

the data preprocessing module is realized by a computer; the data preprocessing module is used for extracting and processing the diagnosis and treatment data to obtain preprocessed data;

the characteristic calculation module is realized by a computer; the characteristic calculation module is used for carrying out characteristic construction on the preprocessed data to form data suitable for being input into the model operation module;

the model operation module is realized by a computer; the model operation module comprises an early death risk assessment model, the early death risk assessment model is obtained by training, tuning and assessing by utilizing a data set in a database on the basis of an XGboost model, and the performance of the early death risk assessment model is effectively verified and assessed by a multi-center large sample data set; the data constructed through the characteristics is input into a model operation module, and the model operation module outputs a risk prediction result of the patient during the hospitalization period.

9. The ensemble learning-based early mortality risk assessment apparatus according to claim 8, wherein:

the model operation module comprises a model which is an early death risk assessment model established by the ensemble learning-based early death risk assessment model establishing method of any one of claims 1-7.

10. The ensemble learning-based early mortality risk assessment apparatus according to claim 8, wherein:

the extraction and processing comprises cleaning, sampling and interpolation of the diagnosis and treatment data.

Images