CN115579136A

CN115579136A - Delayed discharge risk prediction method

Info

Publication number: CN115579136A
Application number: CN202211068045.3A
Authority: CN
Inventors: 郑兴东; 刘军; 邹海东; 俞华; 胡国勇; 傅春瑜; 赵英英; 范骏翔; 黄陈; 石耀罡
Original assignee: Shanghai First Peoples Hospital
Current assignee: Shanghai First Peoples Hospital
Priority date: 2022-09-01
Filing date: 2022-09-01
Publication date: 2023-01-06

Abstract

The invention relates to the field of medical expert systems for computer-aided diagnosis, in particular to a method for predicting the risk of delayed discharge. A delayed discharge risk prediction method is characterized in that: the method is implemented in sequence according to the following steps: 1. data acquisition and treatment, two, inclusion and exclusion standards, three, and prediction factor preliminary screening: extracting at least 50 clinically relevant characteristic variables, and dividing a data set: and randomly dividing the data set into a training set and an independent testing set, namely a two-part model, a five-part model and a preliminary model. The invention has accurate risk prompt and improves the prognosis of patients.

Description

Delayed discharge risk prediction method

Technical Field

The invention relates to the field of medical expert systems for computer-aided diagnosis, in particular to a method for predicting the risk of delayed discharge.

Background

With continuous deepening of new medical improvement, the national requirements on the fine management level of hospitals are continuously improved, the management mechanism of the hospitals faces new opportunities and challenges, the fine management becomes a necessary way for the development of modern hospitals, and the fine management is a key factor for improving the operation efficiency and the medical quality of the hospitals.

The average hospitalization day is the average hospitalization time of a patient discharged within a certain time, and is an important index for evaluating the operation efficiency of the hospital.

The hospital bed turnover number is accelerated, the medical service efficiency is improved, and the social problems of difficult seeing and hospitalization and the like are relieved to a certain extent. Under the condition that the existing medical resources are not changed, the average hospitalization day is reduced, the hospitalization cost of the patient is reduced, the burden is reduced, the utilization rate of beds can be improved, the economic benefit of a medical institution is improved, the huge social benefit is generated, and the three-win situation of the medical institution, the patient and the society is realized.

The average hospital stay day controlled by the medical institution usually adopts a historical data deduction algorithm, namely, in order to shorten the average hospital stay day, a hospital can make an assessment target according to the historical data of the average hospital stay day of nearly 3 years or 5 years in each department, but the method does not fully consider the difficult and complicated degree of the department receiving and treating disease, and the scientificity and fairness of the method are usually questioned by clinical departments.

Disclosure of Invention

In order to overcome the defects of the prior art and provide a medical auxiliary system with accurate risk prompt and improved patient prognosis, the invention discloses a delayed discharge risk prediction method.

The invention achieves the purpose by the following technical scheme:

a delayed discharge risk prediction method is characterized in that: the method is implemented in sequence according to the following steps:

1. data acquisition and treatment:

a sample data set: 4700 patients in gastrointestinal surgery are selected from patients discharged from hospital, and diagnosis, medical history, examination, chief complaint, physical sign and medical advice of the patients are collected;

2. inclusion and exclusion criteria:

inclusion criteria were: the following patients were selected: gastrointestinal surgery inpatients, obtains the pathological diagnosis report of gastric cancer/colorectal cancer,

exclusion criteria: the following patients were not included: diagnosis of gastric/colorectal cancer; patients were not treated surgically; patients with incomplete data, wherein the incomplete data comprises the days of hospitalization less than 3 days after operation, patients who are automatically discharged from hospital, no inspection result and incomplete operation records; palliative or rescue surgery; a new adjuvant radiotherapy and chemotherapy patient; the operation does not reach the excision of R0;

3. preliminary screening of a prediction factor: extracting at least 50 clinically relevant characteristic variables;

4. dividing a data set: randomly dividing a data set into a training set and an independent test set;

5. primary modeling:

fifthly, 1, selecting a model:

the machine learning system can carry out batch experiments on various machine learning algorithms, complete automatic optimization search of the hyper-parameters, select a model with accurate prediction and stable performance to use in a product preferentially, and the algorithms participating in the experiments comprise:

a. xgboost，

b. scorecard，

c. neural network，

d. svm，

e. logistic regression，

f. decision tree，

g. random forest；

the summary of each algorithm is as follows:

a. xgboost: the algorithm provides parallel tree lifting (also called GBDT, GBM), can quickly and accurately solve many data science problems, the same code runs on a main distributed environment (Hadoop, SGE, MPI), and can solve problems beyond billions of examples;

b. scorecard: the card model is evaluated, and the algorithm can obtain a model which is consistent with the form of a traditional medical scale. The principle is that a model variable WOE coding mode is discretized, and then a logistic regression model is applied to carry out a generalized linear model of two classification variables;

c. neural network: the Neural Network is generally also called an Artificial Neural Network (ANN), and the algorithm learning rule may be a Hebb learning rule, a Delta learning rule, a gradient descent learning rule, a Kohonen learning rule, a back propagation learning rule, a probabilistic learning rule, a competitive learning rule, and the like;

d. svm: the algorithm Support Vector Machine (SVM) is a generalized linear classifier (generalized linear classifier) which performs binary classification on data according to a supervised learning (super learning) mode, and a decision boundary of the SVM is a maximum-margin hyperplane (maximum-margin hyperplane) for solving a learning sample;

e. logistic regression: the algorithm is a generalized linear regression analysis model and is commonly used in the fields of data mining, automatic disease diagnosis, economic prediction and the like. For example, risk factors causing diseases are discussed, and the probability of occurrence of diseases is predicted according to the risk factors;

f. precision tree: the decision tree is a prediction model in machine learning, represents a mapping relation between object attributes and object values, and is characterized in that entopy = the disorder degree of the system, the algorithms ID3, C4.5 and C5.0 are used for generating the Entropy, and the metric is the concept based on the Entropy in the informatics theory;

g. random forest: random forest, in machine learning, the algorithm is a classifier that contains multiple decision trees and whose output classes are determined by the modes of the classes output by the individual trees, and developed and deduced by Leo Breiman and Adele Cutler, the algorithm was a random decision forest (random decision trees) proposed by Tin karm Ho of bell laboratories in 1995, which combines the "boosting aggregation" idea of Breimans and the "random subspace method" of Ho to build a set of decision trees.

Fifth, 2 selection of model predictors:

five.2.1 calculate the relevance of each variable to outcome:

information Value (IV) is used as the correlation index. The information value IV is a statistical index of the information theory and is used for measuring the difference of the distribution of a certain variable of two groups of samples (a negative group and a positive group) and further describing the forecasting capacity of the variable on the outcome. Setting the threshold value of the IV value to be 0.2, and deleting variables smaller than 0.2;

five, 2.2 collinearity calculation:

calculating absolute values of correlation coefficients of the screened variables related to the outcome, investigating collinearity, setting a threshold value to be 0.8, removing variables with smaller IV values from the collinear variables exceeding the threshold value, and finally entering 23 modulus variables;

fifthly, 3, modeling process:

fifthly, 3.1, cross Validation (CV) division, namely dividing a 5-fold CV training set and a CV validation set in the training set;

3.2 model hyper-parameter search is carried out based on cross validation, the aim is to maximize the average AUC on the CV validation set, and the optimal hyper-parameter of the model is obtained through a Bayesian optimization (Bayesian optimization) strategy;

fifthly, 3.3, training the algorithm model on the whole training set again according to the optimal hyper-parameter to obtain a final model;

fifthly, 3.4, evaluating the effect of the model on a training set and an independent test set;

fifthly, 4, evaluating the fitting effect of the model:

the prediction capability of the model is characterized by using the area under the receiver operating characteristic curve (ROC) (AUC), the sensitivity (TPR), the specificity (TNR), the Accuracy (ACC) and the accuracy (PPV), the AUC value range is 0-1, and the closer to 1, the better the prediction effect of the model is; measuring the stability of the model in a training set sample and a verification set sample by using a model stability index (PSI), wherein the PSI <0.1 indicates that the model stability is good; PSI in the range of [0.1,0.25] indicates that the model is slightly unstable; PSI >0.25 indicates that the model is unstable.

The method for predicting the risk of delayed discharge is characterized by comprising the following steps:

and in the third step, extracting 79 clinically relevant characteristic variables which are respectively as follows:

Step four, according to the ending layering, randomly dividing the data set into a training set (N = 1433) and an independent test set (N = 615) according to the proportion of 7:3;

and in the fifth step, 4 prediction algorithms are selected: logistic regression, SVM, randomForest, and XGboost.

Colorectal cancer (CRC) is a common malignant tumor of the digestive tract, and the treatment method mainly includes surgical treatment, radiotherapy, chemotherapy, gene therapy and the like, but surgical treatment is still the core link for colorectal malignant tumor, and various endoscopic surgical methods can be adopted according to different patients, clinical conditions and technologies, including laparoscopic colorectal cancer radical surgery, hand assisted laparoscopic surgery, robotic colorectal cancer radical surgery, transanal rectal cancer surgery and the like, and the patients are easy to have complications such as infection, bleeding and the like after treatment, and the corresponding prolonged hospitalization time can also bring serious economic burden to the patients. If the number of days of hospitalization of a postoperative patient is pre-judged, close attention is paid to the patient with a high delay probability of the number of days of hospitalization, prevention and corresponding treatment are given, and new help can be effectively brought to complication occurrence and perioperative period management.

Delayed discharge is associated with postoperative complications, circulatory problems, respiratory diseases and injuries, digestive and genitourinary diseases and other risk factors, and the risk factors are not identical for different disease stages and different regional populations, which may be related to medical level, economic status, and individual status differences. Therefore, when a deferred discharge risk prediction model is established, case contrast research needs to be carried out firstly to screen out independent risk factors related to deferred discharge; secondly, calculating the relative risk degree of the delayed discharge according to the corresponding independent risk factors, and then deducing a prediction model according to the calculated relative risk degree; finally, in order to verify the reproducibility (internal validity) and the universality (external validity) of the model, the prediction model should be subjected to internal verification, that is, the prediction performance of the model is checked by randomly extracting a part of samples from the existing specimen, and external verification, that is, the prediction performance of the model needs to be checked in other samples, but the external verification usually needs multi-center research and is difficult to implement. For the prediction result of the model, the clinical effectiveness can be evaluated by evaluating the sensitivity and specificity of the model, and common indexes comprise sensitivity, specificity, a working curve of a subject and the like.

The invention finally determines meaningful risk indexes and variables through various machine learning methods, and establishes a set of prediction system with high accuracy. Firstly, modeling is carried out on all data, the model comprises Ridge regression, XGB, lightGBM, random forest, adaboost and SVM, and prediction characteristics are selected by combining medical professional knowledge.

The invention aims to deduce high risk factors related to poor progress of delayed discharge based on clinical data of patients and predict the progress of delayed discharge of patients according to a constructed prediction model so as to provide more sufficient support for selection of treatment decisions of patients. In addition, the delayed discharge risk prediction model is constructed and trained in a machine learning mode, so that the model is more accurate and the applicability is better.

The invention has the following beneficial effects:

the method comprises the steps of establishing a deferred discharge risk prediction model by utilizing an electronic medical record information base of a hospital and combining a machine learning method, carrying out real-time operation on patients who come in hospital and are hospitalized after software is deployed in hospital data, and prompting the deferred discharge patients in gastric cancer crowds for further evaluation and early intervention of doctors, delaying disease progress and improving patient prognosis. Moreover, the invention also makes corresponding analysis on the model prediction influence factors, thereby better assisting the clinical doctor in diagnosis and treatment.

Drawings

FIG. 1 shows ROC-curve of four machine learning modules of the present invention,

figure 2 is a schematic representation of the factor weights of the present invention,

FIG. 3 is a schematic diagram of AI performance indicators and grading results in accordance with the present invention.

Detailed Description

The invention is further illustrated by the following specific examples.

Example 1

1. Data acquisition and treatment:

a sample data set: 4700 patients in accordance with the standard are selected from all patients discharged from the gastrointestinal surgery, and the diagnosis, medical history, examination, chief complaint, physical signs and medical advice of the patients are collected;

2. inclusion and exclusion criteria:

exclusion criteria: the following patients were not included: diagnosis of gastric/colorectal cancer; patients were not treated surgically; patients with incomplete data (including hospitalization days less than 3 days after operation, patients who are automatically discharged from hospital, no test results, and incomplete operation records); palliative or rescue surgery; a new adjuvant radiotherapy and chemotherapy patient; the operation does not reach the excision of R0;

3. preliminary screening of a prediction factor:

the case data of nearly 5000 people in 10 years in the city is analyzed, 2048 samples meeting the inclusion and exclusion standards are finally extracted, and 79 clinically relevant characteristic variables are extracted, wherein the characteristics are as follows:

4. Data set partitioning:

according to the result layering, randomly dividing a data set into a training set (N = 1433) and an independent test set (N = 615) according to a 7:3 proportion;

5. and (3) primary modeling:

fifthly, 1, selecting a model:

the machine learning system can carry out batch experiments on various machine learning algorithms, complete automatic optimization search of the hyper-parameters, and preferentially select a model which is accurate in prediction and stable in performance to use in a product. Algorithms involved in the experiment include, but are not limited to:

a. xgboost，

b. scorecard，

c. neural network，

d. svm，

e. logistic regression，

f. decision tree，

g. random forest，

the summary of each algorithm is as follows:

b. scorecard: the card model is evaluated, and the algorithm can obtain a model which is consistent with the form of a traditional medical scale. The principle is that a generalized linear model of two classification variables is carried out by using a logistic regression model after a model variable WOE coding mode is discretized;

d. svm: the Support Vector Machine (SVM) of the algorithm is a generalized linear classifier (generalized-linear classifier) which performs binary classification on data according to a supervised learning mode, and a decision boundary of the SVM is a maximum-margin hyperplane (maximum-margin hyperplane) for solving learning samples;

e. logistic regression: the algorithm is a generalized linear regression analysis model and is commonly used in the fields of data mining, automatic disease diagnosis, economic prediction and the like. For example, risk factors causing diseases are discussed, and the probability of occurrence of diseases is predicted based on the risk factors;

f. precision tree: the decision tree is a prediction model in machine learning, the decision tree represents a mapping relation between object attributes and object values, entopy = the disorder degree of the system, the algorithms ID3, C4.5 and C5.0 are used for generating the use Entropy of the tree algorithm, and the measurement is based on the concept of Entropy in the informatics theory;

4 prediction algorithms are selected: logistic regression, SVM, random Forest and XGboost;

fifthly, 2, selecting model prediction factors:

five.2.1 calculate the relevance of each variable to the outcome:

information Value (IV) is used as the correlation index. The information value IV is a statistical index of an information theory and is used for measuring the difference of the distribution of a certain variable of two groups of samples (a negative group and a positive group) and further describing the predicting capability of the variable on the outcome. Setting the threshold value of the IV value to be 0.2, and deleting variables smaller than 0.2;

five, 2.2 collinearity calculation:

calculating absolute value of correlation coefficient to examine colinearity, setting threshold value to be 0.8, removing smaller variable of IV value from colinear variable exceeding threshold value, and finally entering 23 variables into module;

fifthly, 3, modeling process:

fifthly, 3.3, training the algorithm model again on the whole training set according to the optimal hyper-parameter to obtain a final model;

fifthly, 4, evaluating the fitting effect of the model:

the area under the characteristic curve (ROC) of a subject (AUC), the sensitivity (TPR), the specificity (TNR), the Accuracy (ACC) and the accuracy (PPV) are used for representing the prediction capability of the model, the AUC value range is 0-1, and the closer to 1, the better the prediction effect of the model is; measuring the stability of the model in a training set sample and a verification set sample by using a model stability index (PSI), wherein the PSI <0.1 indicates that the model stability is good; PSI in the range of [0.1,0.25] indicates a slightly unstable model; PSI >0.25 indicates that the model is unstable.

6. And (3) modeling results:

sixth.1 patient characteristics:

Clinical characteristics of CRC patients：

CRC， colorectal cancer； IQR， interquartile range； COPD， chronic obstructive pulmonary disease。

sixth, 2 model efficacy:

Table 2. Performance of the machine learning algorithms：

ROC-current of four machine learning modules is shown in FIG. 1.

Six.3 factor weight:

the body temperature, gastric cancer antigen CA724, hormone drug, albumin, serum uric acid, sodium, cholesterol, leukocyte, carcinoembryonic antigen CEA, and neutrophil ratio are shown in FIG. 2.

7. And (3) model verification:

the higher the AI score, the greater the likelihood of delayed discharge of the patient after surgery. It is recommended to risk-assess and prevent high-risk patients.

The AI performance index and the grading results are shown in fig. 3.

The following data are from the independent test set, currently selected:

low risk of delayed discharge: the risk of delayed discharge in this fractional group was approximately 7.67% (test set low risk group [0, 34.5] had 300, of which 23 were cases).

Risk of intermediate delayed discharge: the risk of delayed discharge in this fraction group was approximately 24.26% (169 in the test cohort [34.5, 44.6], with 41 cases, sensitivity above the mid-high line of 80.1%, specificity of 53.1%, positive predictive value of 28.4%).

High risk of delayed discharge: the risk of delayed discharge in this fraction group was about 52.08% (48 in the test set high risk group [44.6, 100], with 25 cases, 32.7% sensitivity above the high risk line, 90.9% specificity, 45.7% positive predictive value).

Patient A, simulating laparoscopic colorectal cancer radical treatment, preoperative body temperature is 37.5 ℃, lower limb edema occurs, hormone medicines are taken for a long time, blood shows that white blood cells are 20 x 10^9/L, far exceed normal values, the percentage of neutral granulocytes is 82%, and the abnormal (normal is 40% -75%); the biochemical report shows that 20g/L of albumin is lower than the normal value, 500umol/L of serum uric acid is higher, 129mmol/L of blood sodium exists, and hyponatremia exists; CA724 was 61U/mL, 0-6U/mL above normal.

The AI model judges a high risk of delayed discharge, i.e. the length of stay of patient a is expected to exceed 14 days with a risk proportion of over 52%.

Patient B, simulating laparoscopic colorectal cancer radical surgery, carrying out initial treatment, not taking hormone drugs or antitumor drugs, carrying out preoperative body temperature of 36.5 ℃, and displaying white blood cells 11 x 10^9/L slightly exceeding a normal value in a blood routine, wherein the percentage of neutrophils is 76 percent and slightly exceeds the normal value (40 to 75 percent of normal); serum uric acid 500umol/L, higher, blood sodium 133mmol/L, slight hyponatremia exists. CA724 was 9U/mL, slightly above normal by 0-6U/mL.

The AI model determines a low risk of delayed discharge, i.e., patient B is expected to be within 14 days of hospitalization and can be discharged on time.

When a deferred discharge risk prediction model is established, case contrast research needs to be carried out firstly to screen out independent risk factors related to deferred discharge; secondly, calculating relative risk degree of delayed discharge according to corresponding independent risk factors, and deducing a prediction model according to the calculated relative risk degree; finally, in order to verify the reproducibility (internal validity) and the universality (external validity) of the model, the prediction model should be subjected to internal verification, that is, the prediction performance of the model is checked by randomly extracting a part of samples from the existing specimen, and external verification, that is, the prediction performance of the model needs to be checked in other samples, but the external verification usually needs multi-center research and is difficult to implement. For the prediction result of the model, the clinical effectiveness can be evaluated by evaluating the sensitivity and specificity of the model, and common indexes comprise sensitivity, specificity, a working curve of a subject and the like.

And (3) diagnosis: according to 43 ICD-10+ keywords + manual review,

the operation name is as follows: human review was based on 71 ICD-9-CM-3 +.

When the patient has colorectal cancer diagnosis and the patient visits the diagnosis for the colorectal cancer related operation, the model is started after the notification of the operation is given (simultaneously, the 'fitting operation' or 'skin preparation' is included in the medical advice, and the electrocardiogram examination and the inspection examination of the auxiliary examination item are met). And running and refreshing the model result once when the clinical data is changed. And when the postoperative time is more than 168h or a discharge order is issued, stopping the operation of the model.

When the patient accords with the AI model starting condition, the risk of discharging is delayed, predicted and evaluated according to the self state of an illness of the patient, medical care is prompted to do prevention, treatment and treatment work in advance, and the possibility of delaying discharging is reduced.

Claims

1. A delayed discharge risk prediction method is characterized in that: the method is implemented in sequence according to the following steps:

1. data acquisition and treatment:

2. inclusion and exclusion criteria:

3. preliminary screening of a prediction factor: extracting at least 50 clinically relevant feature variables;

4. data set partitioning: randomly dividing a data set into a training set and an independent test set;

5. and (3) primary modeling:

fifthly, 1, selecting a model:

the machine learning system can carry out batch experiments on various machine learning algorithms, complete automatic optimization search of hyper-parameters, and preferentially select a model with accurate prediction and stable performance to use in a product, wherein the algorithms participating in the experiments comprise:

a、xgboost，

b、scorecard，

c、neural network，

d、svm，

e、logistic regression，

f、decision tree，

g、random forest；

fifth, 2 selection of model predictors:

five.2.1 calculate the relevance of each variable to outcome:

five, 2.2 collinearity calculation:

fifthly, 3, modeling process:

fifthly, 4, evaluating the fitting effect of the model:

the area under the characteristic curve (ROC) of a subject (AUC), the sensitivity (TPR), the specificity (TNR), the Accuracy (ACC) and the accuracy (PPV) are used for representing the prediction capability of the model, the AUC value range is 0-1, and the closer to 1, the better the prediction effect of the model is; measuring the stability of the model in a training set sample and a verification set sample by using a model stability index (PSI), wherein the PSI <0.1 indicates that the model stability is good; PSI in the range of [0.1,0.25] indicates that the model is slightly unstable; PSI >0.25 indicates that the model is unstable.

2. The method for predicting the risk of discharge with delay as set forth in claim 1, wherein:

in the third step, 79 clinically relevant characteristic variables are extracted, which are respectively as follows:

and in the fifth step, 4 prediction algorithms are selected: logistic regression, SVM, random Forest, and XGboost.