KR20210152047A - Apparatus for predicting acute kidney injury using machine learning based on artificial intelligence and method thereof - Google Patents

Abstract

The present invention relates to an apparatus for predicting acute renal failure using machine learning based on artificial intelligence and a method thereof. The apparatus for predicting acute renal failure using machine learning based on artificial intelligence, according to the present invention, comprises: a data input unit for configuring a retrospective cohort by receiving patient data collected by a retrospective study in response to a pre-selected criterion; a predictive model generation unit for extracting items related to data causing renal failure from the patient data, and generating a predictive model for predicting whether or not acute renal failure occurs by machine-learning the extracted data; and a prediction unit for predicting whether the acute renal failure of the patient will occur by reflecting the data items of the patient to be predicted in the prediction model. The present invention can reduce the risk of death due to the acute renal failure by predicting the occurrence of the acute renal failure at an early stage according to a clinical condition of the intensive care unit patient using the artificial intelligence-based machine learning model.

Description

Acute renal failure occurrence prediction device and method using artificial intelligence-based machine learning

The present invention relates to an apparatus and method for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning, and more particularly, to an artificial intelligence-based machine learning for early prediction of the occurrence of acute renal failure according to the clinical condition of an intensive care unit patient. It relates to an apparatus for predicting the occurrence of acute renal failure and a method therefor.

Acute kidney injury (AKI) is known to occur in about 10% of all hospitalized patients and up to 70% of patients receiving intensive care unit (ICU) treatment. The incidence of acute renal failure worldwide is increasing because it is difficult to detect and predict early. In Korea, the number of patients diagnosed with acute renal failure is steadily increasing, resulting in an increase in medical expenses.

The occurrence of acute renal failure in intensive care unit patients is a fatal complication that sharply reduces the patient's survival rate by 15 to 60%. There is a very high probability of transitioning to end-stage renal disease.

Recently, many studies have been conducted on acute renal failure occurring in the intensive care unit, and efforts have been made to diagnose and subdivide the degree of it. Moreover, there is a problem in that it is very difficult to accurately predict the factors causing acute renal failure because very complex factors such as the patient's past history, drug taking history, the cause of admission to the intensive care unit, and vital signs act.

Recently, a machine learning system based on artificial intelligence has been in the spotlight as a new paradigm in the development of disease diagnosis and treatment algorithms.

This artificial intelligence-based machine learning system can contribute greatly to the implementation of a predictive model for the risk of various diseases through big data analysis and a treatment decision-making model that allows prompt treatment response immediately after examination. It is expected that there will be

The technology underlying the present invention is disclosed in Korean Patent Application Laid-Open No. 10-2014-0076571 (published on June 20, 2014).

The technical task of the present invention is to use an artificial intelligence-based machine learning model to predict the occurrence of acute renal failure according to the clinical condition of an intensive care unit patient at an early stage so that diagnosis and treatment can be performed quickly. To provide an apparatus and method for predicting the occurrence of acute renal failure using

The apparatus for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention for achieving this technical task receives patient data collected in a retrospective study in response to a pre-selected criterion and performs a retrospective cohort a data input unit constituting a cohort study); a predictive model generator for extracting items related to data causing renal failure from the patient data, and generating a predictive model for predicting whether or not acute renal failure occurs by machine learning the extracted data; and a predictor for predicting whether acute renal failure occurs in the patient by reflecting the data items of the patient to be predicted in the predictive model.

In addition, the prediction model generator divides the extracted items into a static variable and a dynamic variable according to a preset criterion, and the data corresponding to the static variable is learned using a dense layer. And, the data corresponding to the dynamic variable can be learned using a Long Short Term Memory Layer (LSTM).

In addition, the static variable is at least one of the patient's age, sex, height, weight, hospitalization date, discharge date, BMI, underlying medical history, hospitalization cause, serum creatinine at the time of admission and within the hospitalization period, whether dialysis, drug administration history, and surgery. including at least one, wherein the dynamic variable includes data obtained by measuring at least one of white blood cell count change data, test results, food intake, and excretion at daily intervals, and at least any one or more of blood pressure, pulse, food intake, and excretion. It may include data measured in units of time.

In addition, the predictive model generator learns the predictive model using an Artificial Neural Network (ANN) algorithm including a Recurrent Neural Network (RNN) algorithm or a Convolutional Neural Network (CNN) algorithm. can do.

In addition, the data input unit may classify the collected patient data into a patient group with and without acute renal failure, and anonymize and encode the divided data to configure the retrospective cohort.

In addition, when it is predicted that acute renal failure has occurred through the prediction unit, it may further include a determination unit for determining whether the patient corresponds to any of the predefined efficacy evaluation steps.

In addition, the predefined efficacy evaluation step is divided into the first step when the serum creatinine rises 1.5 times or more from the basal level or the urine volume for 6 hours decreases to 0.5 ml/kg or less, and the serum creatinine is 2 If the urine output increases more than twice or decreases to 0.5 ml/kg or less for 12 hours, it is classified as the second stage. Serum creatinine rises more than 3 times from the basal level or the urine output decreases to 0.3 ml/kg or less for 24 hours. Alternatively, if there is no urine for 12 hours, it may be divided into a third stage, and the risk may be set to increase toward the third stage.

In addition, in the method for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to another embodiment of the present invention, the apparatus for predicting the occurrence of acute renal failure receives patient data collected in a retrospective study in response to a pre-selected criterion and performs a retrospective cohort (Retrospective cohort study) configuring; extracting items related to data causing renal failure from the patient data, and generating a predictive model for predicting whether acute renal failure occurs by machine learning the extracted data; And by reflecting the data items of the patient to be predicted to the prediction model may include the step of predicting whether or not acute renal failure of the patient.

As described above, according to the present invention, it is possible to reduce the risk of death due to acute renal failure by early prediction of the occurrence of acute renal failure according to the clinical condition of an intensive care unit patient using an artificial intelligence-based machine learning model.

In addition, according to the present invention, the average hospitalization period can be shortened and the patient's medical costs can be reduced by promptly performing diagnosis and treatment in the case of acute renal failure.

In addition, according to the present invention, it is possible to build a reliable predictive model by continuously learning the patient's medical information, and according to the learned data, it can be applied to not only kidney disease but also other chronic diseases. It has the effect of establishing a position.

1 is a block diagram illustrating an apparatus for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention.
2 is a diagram illustrating a dense layer.
3 is a diagram illustrating an LSTM layer.
4 is a flowchart illustrating an operation flow of a method for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention.
5 to 7 are diagrams for explaining step S420 in an embodiment of the present invention.
8 is a diagram illustrating a pre-defined validity evaluation step in step S450 in an embodiment of the present invention.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings. In this process, the thickness of the lines or the size of the components shown in the drawings may be exaggerated for clarity and convenience of explanation.

In addition, the terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to intentions or customs of users and operators. Therefore, definitions of these terms should be made based on the content throughout this specification.

First, an apparatus for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention will be described with reference to FIG. 1 .

1 is a block diagram illustrating an apparatus for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention.

As shown in FIG. 1 , the apparatus 100 for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention includes a data input unit 110 , a predictive model generation unit 120 , a prediction unit 130 , and a determination. part 140 .

First, the data input unit 110 configures a retrospective cohort study by receiving patient data collected through a retrospective observational study in response to a preset criterion.

At this time, the data input unit 110 divides the received patient data into a patient group with and without acute renal failure, and anonymizes and codes the divided data to configure the retrospective cohort.

In detail, patients aged 18 years or older who were admitted to the intensive care unit for the set period (3 10,000), the patient data is divided into a patient group with and without acute renal failure, and a retrospective cohort is constructed by anonymizing and coding the classified data.

In this case, the patient data may include basic demographic information of each subject, medical history taking records, medical questionnaire records, drug administration history, blood tests, urine tests, and image data after each subject first entered the intensive care unit.

However, at the time of admission to the intensive care unit, a serum creatinine of 4.0 mg/Dl or higher or a glomerular filtration rate (eGFR) ^{less than 15 ml/min/1.73 m 2 ,} or receiving renal replacement therapy for end-stage renal disease, or within 24 hours of admission to the intensive care unit It is desirable to exclude patients who died or those who could not be analyzed due to omission of relevant data from the collection.

And the predictive model generating unit 120 extracts items related to data causing renal failure from the patient data input from the data input unit 110, and machine learning the extracted data to generate a predictive model for predicting the occurrence of acute renal failure. .

The predictive model generator 120 divides the extracted items into static variables and dynamic variables according to preset criteria, respectively.

First, the static variable is at least one of the patient's age, sex, height, weight, hospitalization date, discharge date, BMI, underlying medical history, hospitalization cause, serum creatinine at the time of admission and within the hospitalization period, dialysis status, drug administration history, and surgery status. includes more than

In this case, data corresponding to the static variable is learned using a dense layer.

Here, the dense layer is also called a fully connected layer (FC Layer), and is a neural network designed to learn by connecting all input and output neurons.

2 is a diagram illustrating a dense layer.

As shown in FIG. 2 , the density layer connects both the input and the output, and includes a weight that connects the input and the output, respectively. That is, if there are 4 inputs and 8 outputs, there are a total of 32 weights. The density layer is the most basic layer in machine learning, and unless it is an image or data that are continuously correlated with each other, there is a lot of data that can be learned through this layer. Such a dense layer has the characteristic of designing a model and adding a layer layer through addition.

In addition, the dynamic variables include data obtained by measuring at least one of white blood cell count change data, test results, food intake, and excretion at daily intervals, and data measuring at least any one or more of blood pressure, pulse, food intake, and excretion in units of time. includes

In this case, data corresponding to the dynamic variable is learned using a Long Short Term Memory Layer (LSTM).

Here, the LSTM layer refers to a time-series data analysis layer of various time units, and is a special type of recurrent neural network designed to remember and learn better even if the distance between sequential input data is long.

3 is a diagram illustrating an LSTM layer.

As shown in FIG. 3, the LSTM layer has a cell state with a CC value flowing like a conveyor belt on top, and the LSTM has three gates (Forget gate, Input gate, Output gate) for protecting and controlling this cell state. ) to prevent vanishing gradients and allow gradients to flow effectively.

For the Forget gate ftft of the three gates, the output range of the sigmoid function is 0 to 1, so if the value is 0, the information in the previous state is forgotten; will do

And the gate (Input gate itit) for memorizing the current information has a value of 0 ~ 1 because it is a sigmoid, but ~CtC~t, which does the Hadamard product, is a result of the hyperbolic tangent, so -1 ~ becomes 1 Therefore, the result may be negative.

Finally, as for the gate (Output gate otot) for the final result htht, the value obtained by multiplying the hyperbolic tangent of the cell state by Hadamard becomes the final result of the LSTM layer.

In addition, the predictive model generator 120 is a predictive model using an artificial neural network (ANN) algorithm including a recurrent neural network (RNN) algorithm or a convolutional neural network (CNN) algorithm. can also learn

In addition, the prediction unit 130 reflects the patient's data items to be predicted in the prediction model generated by the prediction model generation unit 120 to predict whether or not acute renal failure occurs in the patient.

That is, when items related to data causing renal failure in the patient's data to be predicted are input to the prediction model, the prediction unit 130 may predict whether acute renal failure occurs in the patient.

Finally, when it is predicted that acute renal failure has occurred in the patient through the prediction unit 130 , the determination unit 140 determines which stage of the predefined efficacy evaluation steps the patient corresponds to.

At this time, the efficacy evaluation stage is divided into the first stage (mild, risk) when the serum creatinine rises 1.5 times or more from the basal level or the urine output decreases to 0.5 ml/kg or less for 6 hours, and the serum creatinine is higher than the basal level. If the urine output rises more than 2 times or decreases to 0.5 ml/kg or less for 12 hours, it is classified as Stage 2 (moderate, Injury). If it decreases to less than ml/kg or there is no urine for 12 hours, it can be classified as a third stage (severe, Failure), and the risk is set to increase as the third stage goes.

Hereinafter, a method for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention will be described with reference to FIGS. 4 to 8 .

4 is a flowchart illustrating an operation flow of a method for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention. A detailed operation of the present invention will be described with reference to this.

According to an embodiment of the present invention, first, the data input unit 110 of the apparatus 100 for predicting the occurrence of acute renal failure configures a retrospective cohort by receiving patient data collected in a retrospective study in response to a pre-selected criterion (S410) ).

At this time, the data input unit 110 divides the received patient data into a patient group with and without acute renal failure, and anonymizes and codes the divided data to configure the retrospective cohort.

Next, the predictive model generator 120 extracts items related to data causing renal failure from the patient data input in step S410, and generates a predictive model for predicting whether or not acute renal failure occurs by machine learning the extracted data (S420) ).

In this case, the predictive model generator 120 divides the extracted items into a static variable and a dynamic variable according to a preset criterion, respectively.

First, the static variable is at least one of the patient's age, sex, height, weight, hospitalization date, discharge date, BMI, underlying medical history, hospitalization cause, serum creatinine at the time of admission and within the hospitalization period, dialysis status, drug administration history, and surgery status. includes more than

In this case, data corresponding to the static variable is learned using the density layer.

In addition, the dynamic variables include data obtained by measuring at least one of white blood cell count change data, test results, food intake, and excretion at daily intervals, and data measuring at least any one or more of blood pressure, pulse, food intake, and excretion in units of time. includes

In this case, the data corresponding to the dynamic variable is learned using the LSTM layer.

5 to 7 are diagrams for explaining step S420 in an embodiment of the present invention.

5 is a diagram showing the configuration of a data usage set (Data Usage Set), and FIG. 6 is a diagram showing that the collected time series data is converted into a supervised learning data structure. The specific data analysis method is as in FIG. 5, A predictive model for the occurrence of acute renal failure at the next time point is constructed using the updated data up to the time point of acute renal failure, and the occurrence of acute renal injury the next day by combining basic demographic information and vital signs and test results cut in hourly units Data is preprocessed in the form of individual cases with whether or not it is a response variable. In this case, a predictive model is constructed on the refined data.

And, as shown in FIG. 6 , in order to predict the future after 48 hours, it is possible to predict the future 48 hours (Past History Time (48 hour) + Forecast Time (48 hour)) by using the result learning after 48 hours.

7 is a diagram for describing a process in which a static variable and a dynamic variable are combined for multivariate sequence classification.

As shown in Figure 7, the density layer (Static Layer) receives a static variable as an input (Input: Static Variable) and transmits it to the next layer. At this time, a repeat vector is used to match the time series to a concatenate layer. , and the LSTM layer receives a dynamic variable as an input (Input: Dynamic Day Variable, Input: Dynamic Hour Variable), analyzes the time unit time series, and then delivers it to the concatenate layer. At this time, the concatenate layer is composed of a method of classifying a class in the final output layer (Output: Static Variable, Output: Dynamic Day Variable, Output: Dynamic Hour Variable) by integrating static and dynamic variables. .

Accordingly, the predictive model generator 120 may generate and learn a predictive model for predicting whether or not acute renal failure occurs through such a learning process.

In addition, the predictive model generator 120 is a predictive model using an artificial neural network (ANN) algorithm including a recurrent neural network (RNN) algorithm or a convolutional neural network (CNN) algorithm. can also learn

Next, the prediction unit 130 reflects the patient's data items to be predicted in the prediction model generated in step S420 to predict whether or not acute renal failure of the patient will occur (S430).

That is, when items related to data causing renal failure in the patient's data to be predicted are input to the prediction model, the prediction unit 130 may predict whether acute renal failure occurs in the patient.

Finally, as a result of the prediction of step S430, when it is predicted that renal failure has occurred in the patient (S440), the determination unit 140 determines which step the patient corresponds to among the predefined efficacy evaluation steps (S450).

In this case, the efficacy evaluation stage can be defined by referring to the guidelines (RIFLE criteria) presented by the World Renal Association (KDIGO).

8 is a diagram illustrating a pre-defined validity evaluation step in step S450 in an embodiment of the present invention.

As shown in FIG. 8, when serum creatinine rises 1.5 times or more from the basal level or when the urine output decreases to 0.5 ml/kg or less for 6 hours, it is classified as the first stage (mild, risk), and serum creatinine is higher than the basal level. If the urine output rises more than 2 times or decreases to 0.5 ml/kg or less for 12 hours, it is classified as Stage 2 (moderate, Injury). If it decreases to less than ml/kg or there is no urine for 12 hours, it can be classified as a third stage (severe, Failure), and the risk is set to increase as the third stage goes.

Accordingly, the apparatus 100 for predicting the occurrence of acute renal failure may provide a treatment decision so that the treatment of the corresponding patient is appropriately performed in response to the determination result of step S450 .

As described above, the apparatus and method for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning according to an embodiment of the present invention use an artificial intelligence-based machine learning model according to the clinical condition of an intensive care unit patient. By predicting the onset early, the risk of death from acute renal failure can be reduced.

In addition, according to an embodiment of the present invention, the average hospitalization period can be shortened and the patient's medical cost can be reduced by promptly performing diagnosis and treatment in the case of acute renal failure.

In addition, according to an embodiment of the present invention, a highly reliable predictive model can be built by continuously learning the patient's medical information, and it can be applied to not only kidney disease but also other chronic diseases according to the learning data, so future advanced medical care It has the effect of establishing a position in the health industry.

Although the present invention has been described with reference to the embodiment shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and equivalent other embodiments are possible therefrom. will be. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the following claims.

100: device for predicting the occurrence of acute renal failure 110: data input unit
120: predictive model generator 130: predictor
140: judgment unit

Claims (14)

In the apparatus for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning,
a data input unit configured to configure a retrospective cohort study by receiving patient data collected in a retrospective study in response to a pre-selected criterion;
a predictive model generator for extracting items related to data causing renal failure from the patient data, and generating a predictive model for predicting whether or not acute renal failure occurs by machine learning the extracted data; and
Acute renal failure occurrence prediction apparatus including a prediction unit for predicting whether or not acute renal failure in the patient by reflecting the data items of the patient to be predicted in the prediction model. According to claim 1,
The predictive model generation unit,
The extracted items are divided into a static variable and a dynamic variable according to a preset criterion, and the data corresponding to the static variable is learned using a density layer and corresponds to the dynamic variable. A device for predicting the occurrence of acute renal failure that learns using the LSTM layer (Long Short Term Memory Layer). 3. The method of claim 2,
The static variable is
Including at least one or more of the patient's age, sex, height, weight, hospitalization date, discharge date, BMI, underlying medical history, hospitalization cause, serum creatinine at the time of admission and within the hospitalization period, whether dialysis, drug administration, and surgery,
The dynamic variable is
Acute, including data obtained by measuring at least one of white blood cell count change data, test results, food intake, and excretion at daily intervals, and data measuring at least any one or more of blood pressure, pulse, food intake, and excretion in units of time A device for predicting the occurrence of renal failure. According to claim 1,
The predictive model generation unit,
Acute renal failure occurrence prediction apparatus for learning the predictive model using an artificial neural network (ANN) algorithm including a recurrent neural network (RNN) algorithm or a convolutional neural network (CNN) algorithm. According to claim 1,
The data input unit,
A device for predicting the occurrence of acute renal failure by dividing the collected patient data into a patient group with and without acute renal failure, and anonymizing and encoding the separated data to configure the retrospective cohort. According to claim 1,
When it is predicted that acute renal failure has occurred through the prediction unit,
Acute renal failure occurrence prediction device further comprising a determination unit for determining whether the patient corresponds to any of the predefined efficacy evaluation steps. 7. The method of claim 6,
The predefined effectiveness evaluation step is,
If the serum creatinine rises more than 1.5 times from the basal level or the urine output decreases to 0.5 ml/kg or less for 6 hours, it is classified as the first stage,
When serum creatinine rises more than twice the basal level or when urine output decreases to 0.5ml/kg or less for 12 hours, it is classified as the second stage,
When serum creatinine rises 3 times or more from the baseline level, or when urine output is reduced to 0.3 ml/kg or less for 24 hours, or when there is no urine for 12 hours, acute renal failure is classified as a third stage and the risk increases as the third stage progresses. occurrence prediction device. In a method for predicting the occurrence of acute renal failure using artificial intelligence-based machine learning,
The apparatus for predicting the occurrence of acute renal failure comprises: configuring a retrospective cohort study by receiving patient data collected in a retrospective study in response to a pre-selected criterion;
extracting items related to data causing renal failure from the patient data, and generating a predictive model for predicting whether acute renal failure occurs by machine learning the extracted data; and
A method of predicting the occurrence of acute renal failure by reflecting the data items of the patient to be predicted in the predictive model, comprising the step of predicting whether or not the patient will have acute renal failure. 9. The method of claim 8,
The step of generating the predictive model comprises:
The extracted items are divided into a static variable and a dynamic variable according to a preset criterion, and the data corresponding to the static variable is learned using a density layer and corresponds to the dynamic variable. A method for predicting the occurrence of acute renal failure by learning the data using the LSTM layer (Long Short Term Memory Layer). 10. The method of claim 9,
The static variable is
Including at least one or more of the patient's age, sex, height, weight, hospitalization date, discharge date, BMI, underlying medical history, hospitalization cause, serum creatinine at the time of admission and within the hospitalization period, whether dialysis, drug administration, and surgery,
The dynamic variable is
Acute, including data obtained by measuring at least one of white blood cell count change data, test results, food intake, and excretion at daily intervals, and data measuring at least any one or more of blood pressure, pulse, food intake, and excretion in units of time A method for predicting the occurrence of renal failure. 9. The method of claim 8,
The step of generating the predictive model comprises:
Acute renal failure occurrence prediction method for learning the predictive model using an artificial neural network (ANN) algorithm including a recurrent neural network (RNN) algorithm or a convolutional neural network (CNN) algorithm. 9. The method of claim 8,
The step of constructing the retrospective cohort comprises:
A method for predicting the occurrence of acute renal failure by dividing the collected patient data into a patient group with and without acute renal failure, and anonymizing and encoding the separated data to configure the retrospective cohort. 9. The method of claim 8,
When it is predicted that renal failure has occurred in the predicting step,
Acute renal failure occurrence prediction method further comprising the step of determining whether the patient corresponds to any of the predefined efficacy evaluation steps. 14. The method of claim 13,
The predefined effectiveness evaluation step is,
If the serum creatinine rises more than 1.5 times from the basal level or the urine output decreases to 0.5 ml/kg or less for 6 hours, it is classified as the first stage,
When serum creatinine rises more than twice the basal level or when urine output decreases to 0.5ml/kg or less for 12 hours, it is classified as the second stage,
When serum creatinine rises 3 times or more from the baseline level, or when urine output is reduced to 0.3 ml/kg or less for 24 hours, or when there is no urine for 12 hours, acute renal failure is classified as a third stage and the risk increases as the third stage progresses. How to predict the occurrence.

Images