CN110051324B

CN110051324B - Method and system for predicting death rate of acute respiratory distress syndrome

Info

Publication number: CN110051324B
Application number: CN201910194628.2A
Authority: CN
Inventors: 黄炳升; 梁栋; 刘勇; 邹儒诗; 黄树华; 余夏夏
Original assignee: Shenzhen University; Shenzhen Hospital of Southern Medical University
Current assignee: Shenzhen University; Shenzhen Hospital of Southern Medical University
Priority date: 2019-03-14
Filing date: 2019-03-14
Publication date: 2022-06-10
Anticipated expiration: 2039-03-14
Also published as: CN110051324A

Abstract

The invention discloses a method and a system for predicting the death rate of acute respiratory distress syndrome, wherein the method comprises the following steps: acquiring medical data of a patient suffering from acute respiratory distress syndrome; performing model training by adopting a machine learning method according to the acquired medical data to obtain an acute respiratory distress syndrome mortality prediction model; and predicting the object to be predicted by adopting an acute respiratory distress syndrome mortality prediction model to obtain a prediction result of the acute respiratory distress syndrome mortality. The method trains an acute respiratory distress syndrome mortality prediction model through a machine learning method, then predicts the mortality of the ARDS patient by adopting the acute respiratory distress syndrome mortality prediction model, applies the machine learning to the mortality prediction of the ARDS patient, can accurately and objectively predict the mortality of the ARDS patient through the model trained through the machine learning, provides more effective and feasible prediction information for clinicians, and can be widely applied to the field of medical data mining.

Description

Method and system for predicting death rate of acute respiratory distress syndrome

Technical Field

The invention relates to the field of medical data mining, in particular to a method and a system for predicting the death rate of acute respiratory distress syndrome.

Background

Acute Respiratory Distress Syndrome (ARDS) is a common critical condition, which refers to Acute diffuse lung injury associated with exposure to risk factors, often accompanied by increased pulmonary vascular permeability and decreased gas-containing lung tissue due to lung inflammation. Clinically, various critically ill patients have the potential risk of developing ARDS, and the hospitalization mortality rate after the ARDS is between 34.9 and 46.1 percent, which seriously threatens the lives of critically ill patients and influences the quality of life of the critically ill patients. Therefore, a mortality prediction model of the ARDS patient is established, so that the disease severity can be distinguished, and different treatment strategies can be determined; meanwhile, the influence factors of various variables on the death rate of the patient can be explored by utilizing a prediction model, and the method has positive significance for improving the survival rate of the patient.

Currently available methods for predicting survival (survival of survival) in ICU patients are based on traditional analytical methods, including APACHE II, OSI, OI, and LIS analyses. These conventional analysis methods usually collect data at one or more medical centers, then obtain relevant variables based on the experience of disease experts and statistical methods (most commonly logistic regression), and finally construct and verify predictive models from the obtained variables. However, such methods have the following problems: (1) the variables obtained from expert experience or statistical analysis can have subjectivity and data bias; (2) the factors influencing the occurrence and development of the ARDS are very complex, and the statistical analysis is difficult to be carried out by combining multidimensional variables; (3) these methods are not specifically designed for ARDS, and no effective scoring model currently exists that is suitable for predicting mortality in ARDS patients.

Disclosure of Invention

In order to solve the above technical problems, the present invention aims to: provides an objective and accurate method and system for predicting the death rate of the acute respiratory distress syndrome.

The technical scheme adopted by the invention is as follows:

a method for predicting mortality of acute respiratory distress syndrome, comprising the steps of:

acquiring medical data of a patient suffering from acute respiratory distress syndrome;

performing model training by adopting a machine learning method according to the acquired medical data to obtain an acute respiratory distress syndrome mortality prediction model;

and predicting the object to be predicted by adopting an acute respiratory distress syndrome mortality prediction model to obtain a prediction result of the acute respiratory distress syndrome mortality.

Further, the step of acquiring medical data of the patient with acute respiratory distress syndrome specifically includes:

medical data for patients with acute respiratory distress syndrome are downloaded from the MIMIC-III database.

Further, the step of performing model training by using a machine learning method according to the acquired medical data to obtain an acute respiratory distress syndrome mortality prediction model specifically includes:

preprocessing the acquired medical data, wherein the preprocessing comprises sample screening and feature extraction;

And performing model training by adopting a machine learning method according to the preprocessed data to obtain an acute respiratory distress syndrome mortality prediction model, wherein the acute respiratory distress syndrome mortality prediction model comprises a hospitalization mortality prediction model, a 30-day mortality prediction model and an annual mortality prediction model.

Further, the step of preprocessing the acquired medical data specifically includes:

performing sample screening on the obtained medical data according to inclusion criteria and exclusion criteria to obtain a screened sample, wherein the inclusion criteria comprise patients who live in an intensive care unit at age of 18 years or more and are diagnosed as acute respiratory distress syndrome by Berlin criteria, and the exclusion criteria comprise any one of data with incomplete data records in a MIMIC-III database, patients with age of less than 18 years, patients adopting palliative therapy and patients with ICU recording time of less than 48 hours;

and extracting variable characteristics for modeling of each sample from the screened samples.

Further, the step of preprocessing the acquired medical data further specifically includes the following steps:

multiple interpolation is performed on missing data in the acquired medical data.

Further, the step of performing model training by using a machine learning method according to the preprocessed data to obtain a model for predicting the death rate of the acute respiratory distress syndrome specifically comprises the following steps:

classifying the preprocessed data according to the survival days of the patient to respectively obtain a positive group and a negative group of 3 mortality prediction models, wherein the 3 mortality prediction models comprise a hospitalization mortality prediction model, a 30-day mortality prediction model and an annual mortality prediction model;

respectively carrying out intergroup analysis on the positive groups and the negative groups of the 3 mortality prediction models, and screening out features with obvious intergroup difference;

and establishing an acute respiratory distress syndrome mortality prediction model by adopting a random forest algorithm according to the characteristics of remarkable difference among groups.

Further, the positive and negative groups of the 3 mortality prediction models are specifically: the positive group of the in-patient mortality prediction model is the data of patients who died in-patient, and the negative group of the in-patient mortality prediction model is the data of patients who survived in-patient; the positive group of the 30-day mortality prediction model is the data of patients who died within 30 days after hospitalization, and the negative group of the 30-day mortality prediction model is the data of patients who survived within 30 days after hospitalization; the positive group of the one-year mortality prediction model is data of patients who died within one year after hospitalization, and the negative group of the one-year mortality prediction model is data of patients who survived within one year after hospitalization.

Further, the step of establishing an acute respiratory distress syndrome mortality prediction model by adopting a random forest algorithm according to the characteristics of obvious difference among groups specifically comprises the following steps:

dividing the characteristics with obvious difference among groups into a first training set and a test set by adopting a K-fold cross verification method;

dividing the first training set into a second training set and a verification set by adopting a K-fold cross verification method;

finding out the optimal model parameters by adopting a grid optimization method according to the second training set and the verification set, and further constructing a plurality of acute respiratory distress syndrome mortality prediction models by adopting a random forest algorithm according to the optimal model parameters;

testing the test set by adopting each acute respiratory distress syndrome mortality prediction model respectively to obtain prediction results of each fold;

averaging the prediction results of all folds to obtain the corresponding prediction performance result of each acute respiratory distress syndrome mortality prediction model;

and according to the obtained predicted performance results, selecting a model with the best predicted performance result from various acute respiratory distress syndrome mortality prediction models as a final acute respiratory distress syndrome mortality prediction model.

The technical scheme adopted by the other aspect of the invention is as follows:

A system for acute respiratory distress syndrome mortality prediction, comprising the following modules:

the acquisition module is used for acquiring medical data of patients suffering from acute respiratory distress syndrome;

the training module is used for carrying out model training by adopting a machine learning method according to the acquired medical data to obtain an acute respiratory distress syndrome mortality prediction model;

and the prediction module is used for predicting the object to be predicted by adopting the acute respiratory distress syndrome mortality prediction model to obtain the prediction result of the acute respiratory distress syndrome mortality.

an acute respiratory distress syndrome mortality prediction system, comprising:

at least one processor;

at least one memory for storing at least one program;

when executed by the at least one processor, cause the at least one processor to implement an acute respiratory distress syndrome mortality prediction method according to the present invention.

The invention has the beneficial effects that: according to the method and the system for predicting the death rate of the acute respiratory distress syndrome, after medical data of patients with the acute respiratory distress syndrome are obtained, the death rate prediction model of the acute respiratory distress syndrome is trained through a machine learning method, finally, the death rate of the patients with the ARDS is predicted by adopting the death rate prediction model of the acute respiratory distress syndrome, the machine learning is applied to the mortality rate prediction of the patients with the ARDS, the death rate of the patients with the ARDS can be accurately and objectively predicted through the model trained through the machine learning, and more effective and feasible prediction information is provided for clinicians to serve as reference.

Drawings

FIG. 1 is a flowchart of a method for predicting mortality from acute respiratory distress syndrome according to an embodiment of the present invention;

FIG. 2 is a flow chart of data screening according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a random forest algorithm;

FIG. 4 is a flowchart illustrating the embodiment of the present invention for training a mortality prediction model using a machine learning method;

FIG. 5 is a flow chart of the training of the final three mortality prediction models and their application to clinical new data prediction in accordance with an embodiment of the present invention.

Detailed Description

The present invention will be described with reference to the following terms:

EHR: electronic health record, personal electronic health record.

Berlin standard: berlin definition, ARDS 2011 defines a general standard for diagnosing ARDS by the working group.

PEEP: positive end expiratory pressure, positive end expiratory pressure.

And (3) APPS: plateau pressure, airway plateau pressure.

CPAP: continuous Positive air Pressure.

SAPS: simplified Acute Physiology Score.

LIS: the Lung Injury Score, Lung Injury Score.

APACHE: acute Physiology and Chronic Health Evaluation, Acute physiological and Chronic Health score.

OI: oxygenation index, calculation method: [ PaO2/FiO2 ].

OSI: oxygen Saturation Index, calculation method: [ FiO2/SpO2 ].

ROC: receiver Operating characteristics Curve.

AUROC: area Under the Receiver Operating Characteristic Curve.

RF: random forest, a classification algorithm for Machine Learning (ML).

Death in hospital: in-hospital mortalities, patients with ARDS died during hospitalization, based on the time of admission information recording.

Death occurred in 30 days: 30-day mortality, which means death within 30 days after admission to the ARDS patient.

Death occurred in one year: 1-year mortality, which means death within one year after admission to the hospital.

The invention will be further explained and explained with reference to the drawings and the embodiments in the description.

Referring to fig. 1, an embodiment of the present invention provides a method for predicting acute respiratory distress syndrome mortality, including the following steps:

In particular, the medical data of the acute respiratory distress syndrome patient may be medical data such as demographic data, vital signs monitoring, etc. downloaded from a public database (e.g., MIMIC-III database, etc.). The subject to be predicted is a new ARDS patient.

Machine learning is a branch of artificial intelligence, where a computer automatically performs data analysis by designing a special algorithm to master rules (i.e., "learning"), and makes a judgment or prediction on unknown data using the rules. The machine learning method can analyze and master rules through continuous learning, can easily complete information processing, and has incomparable advantages in the aspect of analyzing large data volume and high variable dimensionality compared with a statistical analysis method. The machine learning method comprises a random forest algorithm, a support vector machine algorithm, a deep learning algorithm and the like.

The embodiment applies a machine learning method, finds the mortality rule of the ARDS patient according to the existing medical data of the ARDS patient to obtain a mortality prediction model of the ARDS patient, and enables the prediction model to automatically predict the mortality of the ARDS patient according to the previously learned rule when new data (namely, an object to be predicted) exist next time.

In the embodiment, the machine learning training model is used, more and more complete clinical data are used as input, the influence of expert experience or statistical analysis is small, data acquisition deviation can be avoided, more hidden information can be obtained, and therefore the accuracy of prediction is improved; in addition, machine learning can be trained on data of different specifications or scales, and new variables with or without predictive value can be found, so that additional inspiration is provided for clinical treatment.

Further as a preferred embodiment, the step of acquiring medical data of patients with acute respiratory distress syndrome specifically comprises:

Specifically, MIMIC-III is a database for intensive care patients, which is commonly established by MIT's physical laboratory, bidect (beth Israel deacetones medical center), and philips corporation, and currently contains 53423 adult patient medical data (of which there are 3186 cases of ARDS patient data), including basic demographic data, vital signs monitoring, and the like.

The embodiment utilizes the structured clinical data in MIMIC-III for training, effectively avoids the fuzzy clinical definition and the data acquisition deviation, and improves the accuracy of ARDS mortality prediction.

Further as a preferred embodiment, the step of performing model training by using a machine learning method according to the acquired medical data to obtain a model for predicting the mortality of acute respiratory distress syndrome specifically includes:

In particular, the sample screening is to screen out samples which meet the current ARDS diagnosis standard and meet the requirements of subsequent model construction, and the recorded content can be realized in hospital ICU (ensuring that the model can be practically applied). The feature extraction is to extract features for model training from a sample.

Further as a preferred embodiment, the step of preprocessing the acquired medical data specifically includes:

Specifically, the basis for the diagnosis of acute respiratory distress syndrome by Berlin criteria is: 1) acute attack; 2) OI (PaO2/FiO2) <300mmHg when PEEP (or CPAP) is not less than 5cm H2O; 3) bilateral immersion imaging of the breast image; 4) respiratory failure cannot be fully explained with heart failure.

Patients on palliative therapy refer to patients who have not received active treatment.

The extracted variable characteristics comprise data which can be directly read, such as age, sex, APACHE II score and the like of the patient, and variable characteristics, such as mechanical ventilation time, physiological information of the patient and the like.

Further, as a preferred embodiment, the step of preprocessing the acquired medical data further specifically includes the following steps:

In particular, multiple interpolation is a method of adjusting missing data by filling each missing data value with a series of possible data sets, then analyzing the multiple interpolated data sets using a full data standard, and finally generalizing the analysis results. The embodiment can use SPSS software to complete multiple interpolation operations to supplement the ARDS data with individual data missing.

Further as a preferred embodiment, the step of performing model training by using a machine learning method according to the preprocessed data to obtain an acute respiratory distress syndrome mortality prediction model specifically includes:

respectively carrying out analysis between groups on the positive group and the negative group of the 3 mortality prediction models, and screening out characteristics with obvious difference between the groups;

and establishing an acute respiratory distress syndrome mortality prediction model by adopting a random forest algorithm according to the characteristic of obvious difference among groups.

In particular, intergroup analysis may be performed using SPSS software. In the embodiment, the features with obvious differences among different types of samples are found out as input through the inter-group analysis of the sample features, so that the data redundancy can be reduced, the model calculation amount can be reduced, and more meaningful features can be found out. In this embodiment, the positive group and the negative group in the three prediction models are analyzed between groups, specifically: for continuous variables conforming to normal distribution, t-test (Student t-test) analysis is adopted, and for continuous variables not normally distributed, nonparametric test (Mann-Whitney U test) is adopted; discrete variables were analyzed by Chi2test or Fisher's exact test. The tested variable characteristics with the P value less than 0.05 can be considered as obvious difference among groups, and the characteristics are kept; for the P value of more than 0.05, the difference of the characteristic between the positive group and the negative group is not obvious, the classification influence on the prediction model is small, and the characteristic can be deleted.

The method is characterized in that a random forest is a classifier which trains a plurality of decision trees by using samples and predicts sample results, a top-down recursion method is adopted in the training process of the decision trees, the basic idea is to construct a tree with the fastest descending entropy value by taking information entropy as measurement until the entropy value of leaf nodes is zero, and at the moment, the samples of each leaf node belong to the same category. When a new sample is input, each decision tree in the random forest judges voting respectively, and the decision tree with the largest number of votes is used as a final classification result. The random forest has better anti-noise capability and is not easy to over-fit through an integrated learning and majority voting mechanism of the decision tree, and the death rate of the ARDS patient can be well predicted.

Further as a preferred embodiment, the positive and negative groups of the 3 mortality prediction models are specifically: the positive group of the in-patient mortality prediction model is the data of patients who died in-patient, and the negative group of the in-patient mortality prediction model is the data of patients who survived in-patient; the positive group of the 30-day mortality prediction model is the data of patients who died within 30 days after hospitalization, and the negative group of the 30-day mortality prediction model is the data of patients who survived within 30 days after hospitalization; the positive group of the one-year mortality prediction model is data of patients who died within one year after hospitalization, and the negative group of the one-year mortality prediction model is data of patients who survived within one year after hospitalization.

Further as a preferred embodiment, the step of establishing a model for predicting the mortality of acute respiratory distress syndrome by using a random forest algorithm according to the features of significant difference between groups specifically comprises:

Specifically, more than one acute respiratory distress syndrome mortality prediction model constructed by adopting a random forest algorithm can be adopted, so that model screening can be carried out according to the prediction performance results of all prediction models.

When each pass of the K-pass cross-validation method uses a training set to construct a model, a grid optimization method is used to find the optimal model parameters (the parameters of RF include the number of decision trees n _ estimators, branching criteria, minimum leaf sample number min _ sample _ leaf, etc.). In this embodiment, the first training set is divided into the second training set and the verification set again, the effect of each group of parameters is tested in a loop, a group of parameters with the best AUROC effect is selected to construct a random forest classifier model (i.e., an acute respiratory distress syndrome mortality prediction model), and the test set is tested to obtain a local prediction result; then, averaging the results of K-folding to obtain the predicted performance results of each model; and finally, selecting the model with the best predicted performance result as a final mortality prediction model.

In order to improve the accuracy of the mortality prediction of ARDS patients and provide real-time and feasible prediction information for clinicians, the present embodiment proposes a scheme for acute respiratory distress syndrome mortality prediction. The following describes a specific implementation flow and a usage flow of the scheme.

(I) concrete implementation procedure

The specific implementation process of the scheme of the specific embodiment can be divided into three steps: (1) collecting and preprocessing data; (2) establishing and testing a prediction model; (3) and evaluating and comparing the results.

1. Data collection and data pre-processing

1.1 data Source

This embodiment downloads 3186 cases of ARDS patient data from the public database MIMIC-III (medical Information Mart for Intelligent Care). MIMIC-III is a database for intensive care patients commonly built by MIT's physical laboratory, bidect (beth israel deacons Medical center), and philips corporation, and currently contains 53423 adult patient Medical data, including basic demographic data, vital signs monitoring, etc.

1.2 data preprocessing

This embodiment screens the ARDS patient data according to inclusion and exclusion criteria shown in fig. 2 to select a sample that meets the current ARDS diagnostic criteria and meets the requirements for subsequent model construction, and the record content can be implemented in the ICU of the cooperative hospital (ensuring that the model can be practically used).

The inclusion criteria specifically include the following two requirements (both requirements need to be met to include the data):

1) Hospitalized intensive care unit patients older than 18 years (including 18 years);

2) the patients are diagnosed with ARDS by Berlin standard, and the diagnosis standard is as follows: a) acute attack; b) OI (PaO2/FiO2) <300mmHg when PEEP (or CPAP) is greater than or equal to 5cmH 2O; c) bilateral immersion imaging of the breast image; d) respiratory failure cannot be fully explained with heart failure.

The exclusion criteria are specifically any of the following four requirements (as long as 1 requirement is not satisfied to exclude the data):

1) incomplete data records, such as patients receiving non-invasive ventilation (without mechanical ventilation data), etc.;

2) the age is less than 18 years old;

3) palliative therapy, no positive treatment;

4) ICU recording time was less than 48 hours.

After screening out the unsatisfactory patient data, 475 patients were left for the establishment of the predictive model. In order to ensure that the variables used for modeling have significance, the present embodiment extracts clinical data clinically related to ARDS under the confirmation of the clinician, and has 101 variable information, including directly readable data such as age, sex, APACHE II score, etc. of the patient, and variables such as mechanical ventilation time, patient physiological information, etc. as input variables for modeling.

The mechanical ventilation time (length of mechanical ventilation) in these 101 variables is the duration of the first mechanical ventilation after the patient has been diagnosed with ARDS. The days without mechanical ventilation (days free of mechanical ventilation) is the number of days in the ICU record that the patient did not receive mechanical ventilation (in any form), and is considered 0 if the patient died within 24 hours after extubation. Physiological information (physiologic information) is physiological data recorded prior to the onset of ARDS. The ventilation environment (Ventilator settings) depends on the instrument settings for the first 24 hours of mechanical ventilation the patient receives. After 24 hours of ARDS, the PaO2/FiO2 values were calculated as oxygenation index OI under standard aeration conditions (FiO2 ≧ 0.5; PEEP ≧ 5cm H2O).

In the present embodiment, a method of Multiple interpolation (Multiple interpolation) is used to supplement missing data. Multiple interpolation is a method of adjusting missing data by filling each missing data value with a series of possible data sets, then analyzing the multiple interpolated data sets using a full data standard, and finally generalizing the analysis results. The present embodiment may use SPSS software to perform multiple interpolation operations.

The above steps are the data preprocessing flow of this embodiment, and finally 101 variable characteristics of each patient can be obtained for establishing the prediction model.

2. Predictive model building and testing

2.1 sample Classification

Data can be divided into three categories according to the number of patient survival days, and three prediction models (a hospitalized mortality prediction model, a 30-day mortality prediction model, and an annual mortality prediction model) are used for prediction respectively:

model 1 (hospitalized mortality prediction model): survival during hospitalization, where the former was positive and the latter was negative;

model 2 (30-day mortality prediction model): death within 30 days after hospitalization vs.30 days later, survival, the former is positive, the latter is negative;

Model 3 (annual mortality prediction model): death within one year after hospitalization vs. survival after one year, the former was positive and the latter was negative.

The feature screening and classifier training (i.e., model training) of the three prediction models is basically consistent with the flow of testing and result evaluation.

2.2 feature screening

In the specific embodiment, SPSS software is used for performing inter-group analysis on the features, and the features with obvious differences among different types of samples are found out to be used as input, so that data redundancy can be reduced, model calculation can be reduced, and more meaningful features can be found out. In this embodiment, the positive group and the negative group in the three classified prediction models are analyzed between groups, specifically: for continuous variables conforming to normal distribution, t-test (Student t-test) analysis is adopted, and for continuous variables not normally distributed, nonparametric test (Mann-Whitney U test) is adopted; discrete variables were analyzed by Chi2test or Fisher's exact test. The tested variable characteristics with the P value less than 0.05 can be considered as obvious difference among groups, and the characteristics are kept; for the P value of more than 0.05, the difference of the characteristic between the positive group and the negative group is not obvious, the influence on the classification of the prediction model is small, and the characteristic can be deleted.

2.3 model training and testing

This embodiment uses Random Forest (RF) to build predictive models for hospitalization mortality, 30-day mortality, and one-year mortality of ARDS patients, respectively. The RF is a machine learning algorithm, and can randomly generate a plurality of Decision trees (Decision trees), each Decision tree is a classifier, input data can be predicted through a series of decisions, tags are allocated, and finally, an output result of the RF is generated through "voting" of the Decision trees. In the embodiment, the RF algorithm is realized by adopting a scimit-spare Python library toolkit, as shown in FIG. 3. Fig. 3 includes several decision trees, each tree gives its own prediction result for each sample x, and each tree "votes" to determine the final result y.

The establishment of the prediction model mainly comprises two steps: training (training) and testing (testing).

In order to ensure the reliability and stability of the prediction model, in the present embodiment, an eight-fold cross-validation method (8-folds cross-validation) is used to evaluate the prediction effect of the model, and the overall flow is shown in fig. 4, and specifically includes:

the data are divided into 8 parts with similar category proportion, 7 parts of the data are used as training set training models for each time, and 1 part of the data are used for testing the model effect (namely, a test set). The test set and the training set of each fold training are different, and the total number of loops is 8 (such as Loop 2 in fig. 4), and the final result of the model is obtained on the basis of the eight-fold cross validation. The feature screening of the process is performed separately in each training set of the machine learning cross-validation process, so that different features may be used in each fold.

When the prediction modeling is constructed by using the training set at each discount, a grid optimization method can be used to find the optimal model parameters (the parameters of the RF include the number of decision trees n _ estimators, branching criteria, minimum leaf sample number min _ sample _ leaf, etc.). Therefore, in this embodiment, the training set is divided into two parts, i.e., the training set and the verification set, the effect of each group of parameters (e.g., Loop 1 in fig. 4) is tested in a Loop, a group of parameters with the best AUROC effect is selected to construct each classifier model (i.e., prediction model), and then the test set is tested to obtain the local-turn prediction result.

And averaging the 8-fold results to obtain the prediction performance results of each prediction model.

And finally, selecting the model with the best prediction performance result from all the prediction models as a final prediction model. Since the present embodiment divides the data into three classes and trains 3 different models, each class will get a final prediction result.

3. Evaluation and comparison of results

3.1 Classification result (i.e., model prediction result) evaluation

The evaluation criterion of the classification result (i.e., the model prediction result) in this embodiment is AUROC. The horizontal axis of the ROC curve is a False Positive Rate (FPR), the vertical axis is a True Positive Rate (TPR), and the points on the ROC curve are determined by TPR and FPR expressed under different classification thresholds according to the probability output of the sample (when the output probability is greater than or equal to the set threshold, the sample is predicted to be Positive, otherwise, the sample is negative). As the classification threshold is gradually decreased, more and more samples are predicted to be positive, but these positives are also doped with true negative samples, i.e., TPR and FPR are increased simultaneously. When the threshold is the maximum, the corresponding ROC curve coordinate point is (0,0), when the threshold is the minimum, the corresponding coordinate point is (1,1), and the ideal target is TPR ═ 1, FPR ═ 0, and the corresponding ROC curve coordinate point is (0,1), so this embodiment selects the probability value represented by the point on the ROC curve closest to the coordinate point (0,1) in the model training as the classification threshold.

AUROC is the area under the ROC curve and can be used to evaluate classifier performance. Randomly selecting a positive sample and a negative sample, inputting the positive sample and the negative sample into a prediction model, outputting the prediction probabilities of the two samples, and arranging the positive sample in front of the negative sample from large to small to obtain an AUC (namely the probability that the output probability of the positive sample is greater than that of the negative sample).

Meanwhile, according to the optimal classification threshold, the accuracy, sensitivity and specificity of classification can be obtained, and the calculation formula is as follows:

the accuracy is as follows: accuracy ═ (TP + TN)/(TP + TN + FP + FN)

Sensitivity: sensitivity TP/(TP + FN)

Specificity: specificity TN/(TN + FP)

Wherein, TP: true Positive, i.e., actually Positive, samples predicted to be Positive.

FP: false Positive, i.e., actually negative, sample predicted to be Positive.

TN: true Negative, i.e., actually Negative, samples predicted to be Negative.

FN: false Negative, i.e., actually positive, samples predicted to be Negative.

The mortality prediction model was trained using an RF-based machine learning approach and tested to the following results:

1) AUROC value of hospitalized mortality prediction model is 0.854 (95% confidence interval 0.835-0.874, p < 0.001);

2) AUROC value of 30-day mortality prediction model is 0.817 (95% confidence interval is 0.796-0.839, p < 0.001);

3) AUROC value of one-year mortality prediction model is 0.817 (95% confidence interval is 0.800-0.834, p < 0.001).

3.2 method comparison

In order to compare the prediction effects of the present invention and the existing methods, the present embodiment also reproduces the models mentioned in the existing research about the ARDS mortality prediction, and predicts the data of the same batch. The existing prediction models include SAPS II, OI, OSI, and APPS, with the results appearing as AUROC, and compared to the predicted results for RF, respectively.

SPSS software is used for statistical processing during prediction. The normal distribution-compliant metrology data is expressed as mean ± standard deviation (mean ± std); the measurement data which do not conform to the normal distribution are expressed by median (quartile); the count data is expressed in terms of fractions or percentages. And (3) taking the age, the gender and the APACHE II score as control variables, analyzing the relation between each factor and the mortality by adopting a multivariate logistic regression method, and drawing an ROC curve.

The results of the above prediction methods are shown in table 1 below:

TABLE 1

It is clear from table 1 that the present invention uses an RF-based machine learning approach to predict ARDS patient mortality generally better than existing approaches.

4. Application of predictive model

From the above results, it can be seen that the prediction model trained by the machine learning method can effectively predict the death of ARDS in hospital, 30 days and one year, and the prediction accuracy of the method of the present invention is significantly improved compared to other existing prediction methods. The result shows that the model design scheme and procedure of the present embodiment are feasible, so that the present embodiment can utilize all existing ARDS data, train a final model according to the scheme and procedure, and predict new data to implement clinical application, and the specific procedure is as shown in fig. 5, and by grouping, feature screening, and parameter optimizing existing data, three prediction models can be obtained, namely model 1 (for predicting the probability of ARDS patient hospitalization death), model 2 (predicting the probability of death within 30 days), and model 3 (predicting the probability of death within one year).

When new data needs to be predicted, feature screening is firstly carried out on the new data (according to three screening strategies during training), the same features are selected and then are respectively input into corresponding models, and further prediction on the new data is achieved.

Corresponding to the method in fig. 1, the embodiment of the invention also provides an acute respiratory distress syndrome mortality prediction system, which comprises the following modules:

the acquisition module is used for acquiring medical data of a patient suffering from acute respiratory distress syndrome;

The contents in the above method embodiments are all applicable to the present system embodiment, the functions specifically implemented by the present system embodiment are the same as those in the above method embodiment, and the beneficial effects achieved by the present system embodiment are also the same as those achieved by the above method embodiment.

In correspondence with the method of fig. 1, an embodiment of the present invention further provides an acute respiratory distress syndrome mortality prediction system, including:

at least one processor;

at least one memory for storing at least one program;

when executed by the at least one processor, cause the at least one processor to implement the method for acute respiratory distress syndrome mortality prediction according to the present invention.

While the preferred embodiments of the present invention have been illustrated and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A system for predicting acute respiratory distress syndrome mortality, comprising: the system comprises the following modules:

the pretreatment module is used for pretreating the acquired medical data, and the pretreatment comprises sample screening and feature extraction;

the classification module is used for classifying the preprocessed data according to the survival days of the patient to respectively obtain a positive group and a negative group of 3 mortality prediction models, wherein the 3 mortality prediction models comprise an inpatient mortality prediction model, a 30-day mortality prediction model and an annual mortality prediction model;

The intergroup analysis module is used for respectively carrying out intergroup analysis on the positive groups and the negative groups of the 3 mortality prediction models and screening out characteristics with obvious differences among the groups;

the model establishing module is used for establishing an acute respiratory distress syndrome mortality predicting model by adopting a random forest algorithm according to the characteristics of obvious difference among groups;

2. The acute respiratory distress syndrome mortality prediction system of claim 1, wherein: the acquisition module comprises a download module used for downloading medical data of patients with acute respiratory distress syndrome from the MIMIC-III database.

3. The acute respiratory distress syndrome mortality prediction system of claim 1, wherein: the preprocessing module comprises:

the sample screening module is used for carrying out sample screening on the acquired medical data according to inclusion criteria and exclusion criteria to obtain a screened sample, wherein the inclusion criteria comprise patients who live in an intensive care unit at age of more than or equal to 18 years and are diagnosed as acute respiratory distress syndrome through Berlin criteria, and the exclusion criteria comprise any one of data with incomplete data records in an MIMIC-III database, patients with age of less than 18 years, patients adopting palliative therapy and patients with ICU recording time of less than 48 hours;

And the feature extraction module is used for extracting variable features used for modeling of each sample from the screened samples.

4. The acute respiratory distress syndrome mortality prediction system of claim 3, wherein: the preprocessing module further comprises a multiple interpolation module, and the multiple interpolation module is used for performing multiple interpolation on missing data in the acquired medical data.

5. The acute respiratory distress syndrome mortality prediction system of claim 1, wherein: the positive and negative groups of the 3 mortality prediction models were specifically: the positive group of the in-patient mortality prediction model is the data of patients who died in-patient, and the negative group of the in-patient mortality prediction model is the data of patients who survived in-patient; the positive group of the 30-day mortality prediction model is the data of patients who died within 30 days after hospitalization, and the negative group of the 30-day mortality prediction model is the data of patients who survived within 30 days after hospitalization; the positive group of the one-year mortality prediction model is data of patients who died within one year after hospitalization, and the negative group of the one-year mortality prediction model is data of patients who survived within one year after hospitalization.

6. The acute respiratory distress syndrome mortality prediction system of claim 1, wherein: the model building module comprises:

The first division module is used for dividing the features with obvious difference among groups into a first training set and a test set by adopting a K-fold cross verification method;

the second division module is used for dividing the first training set into a second training set and a verification set by adopting a K-fold cross verification method;

the model construction module is used for finding out the optimal model parameters by adopting a grid optimization method according to the second training set and the verification set, and further constructing a plurality of acute respiratory distress syndrome mortality prediction models by adopting a random forest algorithm according to the optimal model parameters;

the test module is used for testing the test set by adopting each acute respiratory distress syndrome mortality prediction model respectively to obtain the prediction result of each factor;

the prediction performance acquisition module is used for averaging the prediction results of all folds to obtain the prediction performance results corresponding to all the acute respiratory distress syndrome mortality prediction models;

and the model selection module is used for selecting a model with the best predicted performance result from all the acute respiratory distress syndrome mortality prediction models according to the obtained predicted performance result as a final acute respiratory distress syndrome mortality prediction model.

7. A system for predicting acute respiratory distress syndrome mortality, comprising: the method comprises the following steps:

At least one processor;

at least one memory for storing at least one program;

when executed by the at least one processor, cause the at least one processor to implement an acute respiratory distress syndrome mortality prediction system of any of claims 1-6.