WO2019035639A1

WO2019035639A1 - Deep learning-based septicemia early detection method and program

Info

Publication number: WO2019035639A1
Application number: PCT/KR2018/009338
Authority: WO
Inventors: 감혜진; 김하영
Original assignee: 재단법인 아산사회복지재단
Priority date: 2017-08-16
Filing date: 2018-08-14
Publication date: 2019-02-21
Also published as: KR101886374B1

Abstract

The present invention relates to a deep learning-based septicemia early detection method and program. The deep learning-based septicemia early detection method, according to one embodiment of the present invention, comprises the steps of: acquiring, by a computer, a feature data set within N unit times before a reference time point (S200); inputting, by the computer, the feature data set in a septicemia detection model (S400); and providing, by the computer, a result of predicting the occurrence of septicemia at a specific prediction time point (S600).

Description

Deep Learning-based Sepsis Early Detection Method and Program

The present invention relates to a method and program for early detection of septicemia based on deep running.

Sepsis is a condition in which a microorganism is infected and a serious inflammatory reaction occurs in the whole body. Sepsis is rarely detected immediately after infection, and is confirmed by precise examination with suspected suspicious symptoms or developmental patterns based on various vital signs and test values. For example, hyperthyroidism with body temperature rising to 38 degrees or less, hypothermia with less than 36 degrees, breathing more than 24 times per minute (brisk breathing), heart rate more than 90 beats per minute (tachycardia), increased blood leukocyte count (SIRS), which is called septicemia when the systemic inflammatory response syndrome is due to infection by microorganisms. This is called systemic inflammatory response syndrome (SIRS). It takes time to first diagnose sepsis after preliminary signs such as SIRS, and if not prepared for sepsis early, multiple organ dysfunction syndrome (MODS) may occur and the patient may die.

In the past, various reference feature data were generated and applied to regression analysis and machine learning method. To generate the sepsis and to predict the exacerbation of the sepsis, new feature data was generated and used to grasp the correlation between each medical data. In this case, since the human-biased characteristic data processed in a human-understandable form affects the analysis result according to which new feature data is selected and used, the pattern of occurrence of the disease or the alternative disease change pattern Because of the lack of consideration of multiple angles.

Patients admitted to the ICU need to have a precise method for detecting sepsis earlier because the patient's death due to sepsis can be prevented before the occurrence of sepsis as early as possible before the occurrence of sepsis. To this end, the present invention provides a deep-learning-based sepsis early detection method and program for calculating the possibility of sepsis after a specific time by applying a data set generated based on medical data described in medical records to a deep learning algorithm I want to.

The problems to be solved by the present invention are not limited to the above-mentioned problems, and other problems which are not mentioned can be clearly understood by those skilled in the art from the following description.

According to an embodiment of the present invention, there is provided a method for early detection of septicemia based on deep learning, comprising: acquiring a feature data set within a unit time of N units before a reference point; Inputting the feature data set to a sepsis detection model by a computer; And providing a prediction result of a sepsis occurrence at a specific predicted time point, wherein the sepsis detection model is generated by learning learning data based on deep learning, and the prediction time is calculated from k k is a specific natural number), the septic event occurrence prediction result is a result of whether an event of occurrence of septicemia occurs at the predicted time, and the feature data set includes medical data stored in the electronic medical record . &Lt; / RTI >

Further, in another embodiment, the computer further comprises calculating a correlation between each basic feature data for one or more medical data by only basic feature data through the septicemia detection model, And extracting at least one representative value from the at least one medical data recorded in the medical record.

In another embodiment, the learning data includes a target time point for a plurality of sepsis patients and a characteristic data set in N unit time periods before a specific time from the target time point, And the septicemia detection model is generated by applying the feature data set in N unit time periods in the learning data to the deep learning algorithm by matching the k pieces of septicemia occurrence results after k units of unit time elapses.

In another embodiment, the time point at which the sepsis development pattern is first confirmed is an initial time point when the systemic inflammatory response syndrome persists beyond the reference time.

In another embodiment, the step of acquiring the feature data set includes extracting at least one representative value for at least one of systolic blood pressure, pulse pressure, heart rate, body temperature, respiratory rate, leukocyte count, hydrogen ion index, .

In another embodiment, the unit time is any one of a time interval for recording specific first medical data on the electronic medical record or an integer multiple of the time interval.

In another embodiment, the characteristic data set acquisition step may include calculating at least one of successive second medical data values when specific second medical data is acquired at a time interval longer than the unit time .

In another embodiment, when there are a plurality of pieces of the medical data, when each of the pieces of medical data measured for each measurement period does not include the unit time of the reference time point by the first medical data, Characterized in that the feature data is constructed by interpolating or interpolating a value measured at a unit time at a point adjacent to a reference time point of the first medical data from the medical data, A predetermined number of unit times as the unit time of the adjacent time in the order of closest to the unit time of the reference time including the unit time of the nearest time point from the unit time of the base time of the medical data.

In another embodiment, the septicemia detection model uses a long short-term memory (LSTM) algorithm.

In another embodiment, the sepsis generation prediction result providing step performs a sepsis occurrence prediction with N feature data sets changed every unit time for a specific patient.

The deep learning-based septicemia early detection program according to another embodiment of the present invention is combined with a hardware computer to execute the above-mentioned deep learning-based septicemia early detection method and is stored in the medium.

According to the present invention as described above, the following various effects are obtained.

First, since the learned septicemia sensing model optimally generates feature data based on the basic feature data, the user does not have to perform a process of generating a feature (i.e., a Referece Feature) advantageous to prediction.

Second, the occurrence of sepsis can be predicted early and accurately compared with the existing regression model. For example, it is possible to predict the onset of systemic inflammatory response syndrome (SIRS), which is one of the pre-symptom symptoms prior to the occurrence of sepsis, and to prepare for the progression of the patient's sepsis.

Thirdly, it is possible to predict sepsis early by only the data acquired in the ICU and recorded in the electronic medical record (EMR), without having to obtain separate medical data for early detection of sepsis from the patient. Therefore, early detection of septicemia can be performed using a computer without using a separate device in a hospital.

Fourth, data recorded in electronic medical records can be directly applied to a septicemia detection model to predict the possibility of SIRS occurrence after a specific time in real time.

FIG. 1 is a flowchart of an early detection method of sepsis according to an embodiment of the present invention.

2 is an exemplary table of feature data sets obtained from a plurality of medical data sets in accordance with an embodiment of the present invention.

FIG. 3 is a diagram illustrating a process of constructing learning data based on SIRS initial occurrence time, which is one of symptoms of septic shock according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention and the manner of achieving them will become apparent with reference to the embodiments described in detail below with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. To fully disclose the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

Unless defined otherwise, all terms (including technical and scientific terms) used herein may be used in a sense commonly understood by one of ordinary skill in the art to which this invention belongs. Also, commonly used predefined terms are not ideally or excessively interpreted unless explicitly defined otherwise.

The terminology used herein is for the purpose of illustrating embodiments and is not intended to be limiting of the present invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification. The terms " comprises " and / or " comprising " used in the specification do not exclude the presence or addition of one or more other elements in addition to the stated element.

The term " computer " as used herein includes various devices capable of performing arithmetic processing to visually present results to a user. For example, the computer may be a smart phone, a tablet PC, a cellular phone, a personal communication service phone (PCS phone), a synchronous / asynchronous A mobile terminal of IMT-2000 (International Mobile Telecommunication-2000), a Palm Personal Computer (PC), a personal digital assistant (PDA), and the like. The computer may also be a medical device that acquires or observes medical images. In addition, the computer may be a server computer connected to various client computers. The computer may also be comprised of one or more devices.

Hereinafter, a method and program for predicting a deep-run-based sepsis according to embodiments of the present invention will be described with reference to the drawings.

Referring to FIG. 1, a method for early detection of a deep-learning based septicemia according to an embodiment of the present invention includes a step S200 of acquiring a feature data set within a unit time of N units before a reference time point; The computer inputting the feature data set into a sepsis detection model (S400); And a step (S600) in which the computer provides a prediction result of a sepsis occurrence at a specific predicted time point. Hereinafter, a detailed description of each step will be described.

First, the computer acquires the feature data set within N unit time before the reference point (S200; feature data set acquisition step). That is, the computer acquires a dataset for input to the sepsis detection model described below.

The feature data set is calculated based on medical data stored in electronic medical records. In other words, the computer does not acquire new medical data from the patient using a separate sensor or device from the patient to predict the occurrence of sepsis, and periodically measures the patient's condition in the Intensive Care Unit (ICU) . In one embodiment, in the obtaining of the feature data set (S200), the computer stores medical data used for forming the feature data set, such as systolic blood pressure, pulse pressure, heart rate, body temperature, respiratory rate, leukocyte count, hydrogen ion index, At least one of them.

Further, in another embodiment, the feature data set is formed by extracting one or more representative values (e.g., average value, maximum value, and minimum value) for one or more medical data recorded in the electronic medical record. For example, as shown in FIG. 2, the computer obtains three characteristic data (features) by calculating an average value, a maximum value, and a minimum value for specific medical data (for example, pulse pressure).

In this way, the computer uses only the basic feature data calculated from each medical data obtained from the electronic medical record, and sets the reference feature data (reference data) set to reflect the correlation, temporal change, Reference Feature Data) is not performed. As described later, since the computer calculates various correlations using the basic feature data through the sepsis detection model using the deep learning algorithm, the reference feature data may not be used.

In another embodiment, the unit time is any one of a time interval for recording specific first medical data on the electronic medical record or an integer multiple of the time interval. For example, as shown in FIG. 2, since the patient's specific first medical data (for example, systolic pressure) is measured every 1 hour, 30 minutes, or 15 minutes in the ICU, One hour, which is the same or an integer multiple of the time interval (that is, the measurement period) for recording the medical data, can be set as the unit time.

Further, since a plurality of medical data is measured for each measurement period, the computer can set a time, which is equal to or an integral multiple of the measurement period of various medical data, as a unit time. For example, if a sick blood pressure is recorded every 30 minutes in a particular hospital intensive care unit and the heart rate is recorded every 15 minutes, the computer can set an hour, which is an integral number of 30 minutes and 15 minutes, in unit time.

The computer calculates one or more representative values (e.g., average value, maximum value, and minimum value) based on one or more measured values obtained within a unit time for specific medical data, and uses each representative value as the characteristic data. For example, when the body temperature is recorded in the medical record every 10 minutes, the computer calculates the average value (Average) and the maximum value (Max) of the six body temperature data measured within one hour as the unit time do.

Further, in another embodiment, in the characteristic data set acquiring step (S200), when the specific second medical data is acquired at a time interval longer than the unit time, the computer can acquire at least one of successive second medical data values . That is, if there is no value to be measured within the unit time, the computer builds the feature data by filling the adjacent value. For example, as shown in the figure, the hydrogen ion exponent is irregularly measured at a longer time period than the unit time at 2 hours, 4 hours intervals, etc. Therefore, when the hydrogen ion exponent is not measured in a unit time, The value measured at the time is applied or corrected.

The computer acquires a plurality of feature data for N consecutive unit times to form a feature data set. For example, when 20 characteristic data are calculated per unit time using systolic blood pressure, pulse pressure, heart rate, body temperature, respiratory rate, leukocyte count, hydrogen ion index and blood oxygen concentration as medical data as in FIG. 2, The computer sets 5 * N (e.g., 100) pieces of feature data acquired for N consecutive unit times (e.g., 5 hours) into one feature data set.

Also, in one embodiment, the computer can set the current time as a reference time point. That is, in order to determine whether there is a possibility of occurrence of a sepsis occurrence pattern defined previously after a predetermined time from the present time, the computer uses a feature data set acquired within N unit hours before the present point in time, And the possibility of post-transplant sepsis. For example, since systemic inflammatory response syndrome (SIRS) occurs in sepsis patients in many cases, the computer can not detect N (N) hours before the current time (for example, 3 hours) The computer calculates the probability of occurrence of SIRS after a lapse of a specific time in real time by using the feature data set obtained within the unit time of each unit. , It is possible to continuously calculate the possibility that the sepsis will start at a certain time (for example, 3 hours). Through this, the computer can continuously perform the possibility of sepsis in a patient admitted to the ICU at a unit time interval.

Thereafter, the computer inputs the feature data set to the sepsis detection model (S400). That is, the computer inputs a feature data set obtained from a patient in unit time intervals or in real time to a sepsis detection model. The septicemia detection model is generated by learning learning data based on deep learning.

In one embodiment, the training data for training the septicemia sensing model includes a target time point for a plurality of sepsis patients and a set of N characteristic data within a time unit of N times before a specific time from the target time point. That is, the septicemia detection model includes a feature data set in N units of time in the learning data and a target time point after k units of unit time elapse (that is, a time point when a condition that can be judged as the occurrence of sepsis (hereinafter referred to as a sepsis judgment condition) ) And applying it to the deep learning algorithm.

For example, a time point for determining the criteria for determining sepsis may be the time of the first occurrence of SIRS. Since SIRS occurs in a large number of patients prior to sepsis, the occurrence of SIRS can be used as an element for early detection of sepsis. That is, as shown in FIG. 3, the target time point is the time when the SIRS first occurs in a patient admitted to the intensive care unit (particularly, the Medical Intensive Care Unit (MICU)) before diagnosis of sepsis. (3) respiratory rate> 20 / minute, or PaCO2 <32 mmHg, (3) respiratory rate> 20 / minute, or (4) at least two of the leukocyte count> 12,000 / μl, <4,000 / μl, or band neutrophil> 10% (ie, satisfying the SIRS condition) To the target point in time.

In addition, the method of constructing the feature data sets in N unit time for each patient is performed in the same manner as the method of acquiring the feature data set inputted to the septicemia detection model in order to calculate the possibility of occurrence of SIRS after a specific time.

In addition, the computer extracts a patient who has experienced SIRS among patients diagnosed with sepsis in the past, and then generates a feature data set based on the medical data of the patients. Since the learning data is generated only from the medical data of the patient who has been diagnosed as sepsis after the SIRS condition, the computer can construct the learning data for predicting the sepsis symptom accompanied by the SIRS condition.

Specifically, an embodiment of a process for constructing learning data used for training (learning) of a sepsis detection model for early diagnosis of sepsis accompanied by SIRS state is as follows. First, the computer extracts patients with a history of sepsis diagnosis. Thereafter, the computer extracts the patient whose SIRS state lasted for a certain time or longer before the time of the sepsis diagnosis. For example, the computer continuously extracts patients who continue to have SIRS status for 5 hours. Then, as shown in FIG. 3, the computer extracts the starting point at which the SIRS state starts, as the target point in time. Then, the computer extracts the feature data set within N unit time before the predetermined predicted interval time (k unit time, for example, 3 hours) from the target time. At this time, the feature data set is formed by calculating representative values (for example, an average value, a maximum value, a minimum value, and the like) for general medical data included in the electronic medical record of a plurality of extracted patients by each unit time.

A variety of deep-running algorithms can be applied to the sepsis detection model that trains the training data. In one embodiment, the septicemia sensing model may utilize Deep Feedforward Network (DFN) or Long Short-Term Memory (LSTM) algorithms.

A typical DFN has a network structure consisting of an input layer, one or more hidden layers, and an output layer. The data is input to the neuron of the input layer and calculated final values by calculating the weights of the connected edges and the sum of the values delivered from the previous node and functions until reaching the output layer, Through the comparison, learning proceeds in the direction of minimizing the difference between them.

LSTM is one of the recurrent neural network (RNN) methodologies that reflect temporal variability in time series data among several techniques of deep learning. It solves the problem of learning disability on deep network (eg Vanishing Gradient) in general RNN And is used in various research and development fields. The RNN is advantageous for learning patterns within a temporal change, and thus is useful for analyzing biomedical data in which data is acquired continuously over time. Thus, the sepsis detection model for analyzing learning data can provide higher accuracy when applying LSTM.

The computer provides a prediction result of occurrence of sepsis at a specific predicted time point (S600). The prediction time is a time point that has elapsed by k unit time intervals from the reference time point. That is, the prediction time is a time point that has elapsed by a time interval between the N unit time and the target time for acquiring the feature data set in the learning data from the reference time point (for example, the current time point). The computer uses the septicemia detection model to calculate the likelihood that one of the defined sepsis incidence patterns will occur at the time of the prediction. That is, the prediction result of the sepsis occurrence indicates whether any one of the precondition of the occurrence of the sepsis occurs at the predicted time.

In one embodiment, when the specified sepsis occurrence pattern is the occurrence of the SIRS state, the sepsis detection model detects the target time at which the SIRS state in the learning data starts and the characteristic in the N unit time obtained before k unit time from the target time The data set is learned, and a prediction result of a sepsis occurrence at a prediction time when k unit times has elapsed is calculated by using the feature data set within N unit time before the present point in time. At this time, the predicted sepsis occurrence result indicates whether the SIRS state, which is a precursor state of the occurrence of sepsis, starts at the predicted time.

According to another embodiment, the step of providing the sepsis occurrence prediction result (S600) may calculate the possibility of sepsis by using N feature data sets changed every unit time for a specific patient.

As described above, the deep-learning-based septicemia early detection method according to an embodiment of the present invention may be implemented as a program (or an application) to be executed in combination with a hardware computer and stored in a medium.

The above-described program may be stored in a computer-readable medium such as C, C ++, JAVA, machine language, or the like that can be read by the processor (CPU) of the computer through the device interface of the computer, And may include a code encoded in a computer language of the computer. Such code may include a functional code related to a function or the like that defines necessary functions for executing the above methods, and includes a control code related to an execution procedure necessary for the processor of the computer to execute the functions in a predetermined procedure can do. Further, such code may further include memory reference related code as to whether the additional information or media needed to cause the processor of the computer to execute the functions should be referred to at any location (address) of the internal or external memory of the computer have. Also, when the processor of the computer needs to communicate with any other computer or server that is remote to execute the functions, the code may be communicated to any other computer or server remotely using the communication module of the computer A communication-related code for determining whether to communicate, what information or media should be transmitted or received during communication, and the like.

The medium to be stored is not a medium for storing data for a short time such as a register, a cache, a memory, etc., but means a medium that semi-permanently stores data and is capable of being read by a device. Specifically, examples of the medium to be stored include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage, and the like, but are not limited thereto. That is, the program may be stored in various recording media on various servers to which the computer can access, or on various recording media on the user's computer. In addition, the medium may be distributed to a network-connected computer system so that computer-readable codes may be stored in a distributed manner.

While the present invention has been described in connection with what is presently considered to be practical exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, You will understand. Therefore, it should be understood that the above-described embodiments are illustrative in all aspects and not restrictive.

Claims

The computer acquiring the feature data set within N unit time prior to the reference time point;

Inputting the feature data set to a sepsis detection model by a computer; And

The computer providing a prediction result of a sepsis occurrence at a specific predicted time point,

The septicemia detection model is generated by learning learning data based on deep learning,

The prediction time point is a time point that has elapsed from the reference point by k unit time intervals (k is a specific natural number)

The result of predicting the occurrence of sepsis is a result of whether or not the occurrence of the sepsis occurs at the predicted time,

Wherein the feature data set is computed based on medical data stored in an electronic medical record.
The method according to claim 1,

Further comprising the step of the computer calculating a correlation between each basic characteristic data for one or more medical data by only the basic characteristic data through the sepsis detection model,

Wherein the basic feature data comprises:

Wherein at least one representative value is extracted from the at least one medical data recorded in the electronic medical record.
The method according to claim 1,

The learning data includes:

And a feature data set in N time units before a specific time from the target time point for a plurality of sepsis patients,

The target time point is a point in time when a sepsis occurrence pattern is first confirmed in a sepsis patient,

In the sepsis detection model,

Wherein the feature data sets in N time units in the learning data are generated by applying the feature data sets to the deep learning algorithm by matching k pieces of septicemia occurrence results after k unit time elapses.
The method of claim 3,

When the appearance of the sepsis was first confirmed,

An early detection method for sepsis based on deep running, which is the earliest time when the systemic inflammatory response syndrome persists beyond the reference time.
The method according to claim 1,

The feature data set acquiring step includes:

Wherein at least one representative value is extracted for at least one of systolic blood pressure, pulse pressure, heart rate, body temperature, respiratory rate, white blood cell count, hydrogen ion index, and blood oxygen concentration.
The method according to claim 1,

The unit time may be,

Wherein the electronic medical record is one of a time interval for recording specific first medical data on the electronic medical record or an integral multiple of the time interval.
The method according to claim 6,

The feature data set acquiring step includes:

Wherein when the specific second medical data is acquired at a time interval longer than the unit time, the calculation is performed using at least one of successive second medical data values.
The method according to claim 6,

When there are a plurality of medical data,

When each medical data measured for each measurement period does not include a unit time at a reference time point by the first medical data,

Wherein the computer constructs the characteristic data by interpolating or interpolating a value measured at a unit time at a point adjacent to the reference time point of the first medical data among the medical data,

The unit time of the adjacent time point

Wherein the unit time of the first time point is a unit time of a time point adjacent to the time point of the reference time point including the unit time of the nearest time point from the unit time of the reference time point of the first medical data,

Deep learning based early detection of sepsis.
The method according to claim 1,

In the sepsis detection model,

Using a long-short-term memory (LSTM) algorithm.
The method according to claim 1,

Wherein the step of providing the sepsis occurrence prediction result comprises:

A method for early detection of sepsis based on deep running, the method comprising: predicting occurrence of sepsis with N feature data sets changed every unit time for a particular patient.
10. A deep-run-based septicemia early detection program stored in a medium for performing the method of any one of claims 1 to 10 in combination with a computer which is hardware.