KR102446376B1 - Method for predicting of mortality risk or sepsis risk and device for predicting of mortality risk or sepsis risk using the same - Google Patents

Abstract

The present invention provides a method for predicting an emergency situation implemented by a processor, comprising: receiving biosignal data for an individual from a biosignal prediction device; and a risk sequence generation model configured to generate a risk sequence based on the biosignal data It provides a method for predicting the risk of death or sepsis and a device using the same, comprising the steps of generating a risk sequence for an individual based on biosignal data using .

Description

A method for predicting the risk of death or sepsis and a device for predicting the risk of death or sepsis using the same

The present invention relates to a method for predicting the risk of death or sepsis and a device for predicting the risk of death or sepsis using the same, and more particularly, to a method for predicting the risk of death or sepsis that can predict the risk based on a biosignal for an individual, and death using the same or to a device for predicting sepsis risk.

Many patients using medical services are easily exposed to fatal diseases, and in some cases, they need continuous health status check and measures accordingly. In particular, continuous observation of the patient's condition may be more important for critical patients, such as intensive care patients accommodated in an intensive care unit of a hospital.

In order to check the patient's condition with respect to matters related to disease progression or life maintenance, such as pulse rate, blood pressure, or respiration, various predictive devices may be provided for critically ill patients. These predictive devices measure the patient's condition and display the results so that medical staff can check them. However, the conventional predictive equipment has a limitation in that the medical staff must continuously monitor the predictive equipment as it displays predictive information about the patient's condition. That is, the risk prediction system based on the conventional prediction equipment has a problem in that, when continuous monitoring is impossible, an appropriate preliminary action according to the prediction information cannot be provided.

In particular, in the case of patients who are admitted to the intensive care unit due to non-specific symptoms and various disease names, it may be more difficult to predict the prognosis using conventional predictive equipment because there is a tendency to show rapid changes in status even after admission.

Therefore, the development of a new risk prediction system that can predict an emergency situation in advance and quickly provide information related to the emergency situation is continuously required for patients who have received medical services, especially critical patients with high importance of prognosis. there is a situation.

The description underlying the invention has been prepared to facilitate understanding of the invention. It should not be construed as an admission that the matters described in the background technology of the invention exist as prior art.
Prior literature: Registered Patent Publication No. 10-1886374 (2018.08.07.)

On the other hand, the inventors of the present invention found that changes in biosignals as part of the physiological response of the human body are preceded before a clinically suspected morbidity or emergency such as death occurs. It was noted that

In particular, the inventors of the present invention related to the risk of a patient with a serious condition whose condition changes over time, the change in biosignal data obtainable in the intensive care unit is related to the onset of a specific disease, furthermore, to a potentially serious condition such as death. Note that you can have it.

As a result, the inventors of the present invention predict the risk of a patient, particularly a critically ill patient, using time-series biosignal data such as temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, mean blood pressure, and respiration rate acquired for an arbitrary time in advance. was aware that it could be used for

Furthermore, the inventors of the present invention have noted that clinical data of a biological sample obtained from a patient can be closely related to the risk of developing a disease and further, the risk of death.

More specifically, the inventors of the present invention measured the Glasgow coma scale (GCS), arterial blood oxygen saturation, inhaled oxygen concentration, bicarbonate ion concentration, bilirubin level, creatinine level, platelet count, total urine output It was recognized that biological test data such as , potassium concentration, sodium concentration, white blood cell count, lactate concentration, APH level, and hematocrit level can contribute to improving the accuracy of risk prediction.

In particular, the inventors of the present invention were able to find that the biological test data are closely related to the occurrence of sepsis, which is a high cause of death for patients with severe conditions.

On the other hand, the inventors of the present invention have paid attention to providing not only early prediction of the risk state but also early notification according to the prediction result as a way to overcome the limitations of the conventional risk prediction system.

Accordingly, the inventors of the present invention have attempted to develop a new risk prediction system that provides a risk prediction using biosignal data and biological test data obtained from a patient and a notification according to the prediction result.

In particular, the inventors of the present invention have attempted to develop a risk prediction algorithm for potentially serious conditions such as disease onset and death using biosignal data and further biological test data.

The inventors of the present invention could expect that through the development of such a risk prediction algorithm, the patient's disease state could be detected quickly, and the patient's emergency situation could be predicted in advance, thereby improving the treatment performance by advancing the treatment time for the patient. . Furthermore, the inventors of the present invention were able to recognize that the development of a risk prediction algorithm can provide the effects of increasing the survival rate of patients, preventing complications, and reducing treatment costs.

As a result, the inventors of the present invention have developed a new risk prediction system that predicts the risk of death or sepsis based on data obtained from a patient, and provides a notification according to the prediction result.

More specifically, the risk prediction system of the present invention uses a risk sequence generation model configured to calculate a risk score based on biosignal data obtained from an individual such as a patient, furthermore, biological test data, and generate a risk sequence based on this. It can be configured to predict and provide a dangerous situation in advance of any time.

Furthermore, the risk prediction system of the present invention may be configured to provide an early notification to a medical staff or guardian based on the prediction result.

On the other hand, the inventors of the present invention, when providing a notification simply based on the degree of risk, for example, when providing a notification at a level of risk greater than or equal to a threshold, a false notification may be frequently provided for a patient in a serious condition. Attention was paid to the problem.

As a result, the inventors of the present invention were able to apply a notification sending model configured to determine whether to send a notification by further considering biosignal data and biological test data for an individual to the risk prediction system.

Accordingly, the problem to be solved by the present invention is to generate a risk sequence based on biosignal data for an individual, furthermore, biological test data, and predict the risk of death or sepsis based on the risk sequence, risk of death or sepsis It is to provide a method of predicting risk.

Another object to be solved by the present invention is to provide a method for predicting the risk of death or sepsis, configured to provide a notification based on the risk to an individual using a risk notification model.

Another problem to be solved by the present invention is a receiver configured to receive biosignal data, furthermore, biological test data for an object, and a risk of death or sepsis for an object connected to communicate with the receiver and based on various predictive models To provide a device for predicting risk of death or risk of sepsis, comprising a processor configured to predict.

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

In order to solve the problems as described above, a method for predicting the risk of death or sepsis according to an embodiment of the present invention is provided. A method for predicting risk of death or sepsis implemented by a processor, the method comprising: receiving biosignal data for the subject, using a risk sequence generation model configured to generate a risk sequence based on the biosignal data; generating a risk sequence of , and predicting a risk of death or a risk of sepsis for the subject based on the risk sequence.

According to a feature of the present invention, the step of receiving the biosignal data includes receiving at least one biosignal data from the group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, and mean blood pressure for the subject. can do. Furthermore, predicting the risk of death or the risk of sepsis may include predicting the risk of death based on the risk sequence.

According to another aspect of the present invention, the step of receiving the biosignal data includes receiving the biosignal data a plurality of times in a predetermined time unit, and the generating the risk sequence includes: using a risk sequence generation model , generating a risk sequence of a predetermined time unit based on the biosignal data received a plurality of times in a predetermined time unit. Furthermore, predicting the risk of death or the risk of sepsis may include predicting the risk of death based on a risk sequence of a predetermined time unit.

According to another feature of the present invention, Glasgow coma scale (GCS), arterial blood oxygen saturation, inhaled oxygen concentration, bicarbonate ion concentration, bilirubin level, creatinine level, platelet count, total urine output The method may further include receiving biological test data for at least one subject selected from the group consisting of , potassium concentration, sodium concentration, white blood cell count, lactate concentration, APH level, and hematocrit. Furthermore, the generating the risk sequence further includes generating a sepsis risk sequence based on the biosignal data and the biological test data using the risk sequence generation model, wherein the predicting the risk of death or the risk of sepsis comprises: , based on the sepsis risk sequence, may further include predicting the sepsis risk.

According to another feature of the present invention, the receiving of the biological test data may include receiving the maximum and minimum values of the biological test data measured a plurality of times in a predetermined time unit.

According to another aspect of the present invention, the receiving of the biosignal data includes: receiving at least one biosignal data from the group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, and mean blood pressure for an individual may include

According to another feature of the present invention, the method may further include the step of providing a risk notification to the subject, which is performed after the step of predicting the risk of death or the risk of sepsis.

According to another feature of the present invention, the risk sequence generation model is further configured to calculate a death risk score based on the biosignal data, and performs biological test data for the subject before the step of providing the risk notification. It may further include the step of receiving. Meanwhile, the risk notification model includes a first notification model configured to output a vector value based on the death risk score calculated by the risk sequence generation model, a second notification model configured to output a vector value based on biosignal data, and at least one of the third models configured to output a vector value based on the biological test data, and the risk of death based on the vector value output by at least one of the first alert model, the second alert model, and the third alert model and a fourth notification model configured to determine whether to provide a notification.

According to another feature of the present invention, the method may further include receiving a drug administration record or a notification sending record for the subject, which is performed before the step of providing a risk level notification. Furthermore, the first notification model may be further configured to output a vector value based on the death risk score and the drug administration record and/or the notification sending record.

According to another feature of the present invention, the risk of death may be defined as a risk of occurrence of death before a predetermined time, and the risk of sepsis may be defined as a risk of developing sepsis before a predetermined time.

According to another feature of the present invention, the risk sequence generation model comprises the steps of: receiving bio-signal data for learning obtained before a predetermined time before the risk occurs with respect to the sample entity; It may be a model trained by generating and estimating the risk for the sampled subject at any time before the risk occurs based on the training risk sequence.

According to another feature of the present invention, the sample subject is a dead subject, and the step of predicting the risk for the sample subject may include predicting the risk of death for the sample subject based on the learning risk sequence. .

According to another feature of the present invention, the sample subject is a sepsis onset subject, and further comprising the step of receiving biological test data for learning obtained before a predetermined time from the onset of sepsis of the sample subject, and generating a learning risk sequence The performing may include generating a sepsis risk sequence for learning based on the biological test data for learning and the biosignal data for learning. Furthermore, predicting the risk for the sample object may include predicting the risk of developing sepsis for the sample object based on the sepsis risk sequence for learning.

According to another aspect of the present invention, the receiving of the biosignal data may include receiving the maximum value, the minimum value, and the average value of the biosignal data measured a plurality of times in a predetermined time unit. Also, the generating of the individual risk sequence may include generating the individual risk sequence based on the maximum value, the minimum value, and the average value of the biosignal data using the risk sequence generation model.

According to another feature of the present invention, the method of the present invention may further include receiving age data for the subject. In this case, the generating of the individual risk sequence may further include generating the individual risk sequence based on the biosignal data and the age data using the risk sequence generation model. In order to solve the problems as described above, there is provided a device for predicting the risk of death or sepsis according to an embodiment of the present invention. The device includes a receiver configured to receive biosignal data for the entity, and a processor coupled to communicate with the receiver. In this case, the processor is configured to generate a risk sequence of the subject by using a risk sequence generation model configured to generate a risk sequence based on the biosignal data, and predict a risk of death or a risk of sepsis for the subject based on the risk sequence do.

According to a feature of the present invention, the receiver is configured to receive at least one biosignal data from the group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, and mean blood pressure for the subject, and the processor is configured to: As such, it may be further configured to predict the risk of death.

According to another feature of the present invention, the receiving unit is configured to receive the biosignal data a plurality of times in a predetermined time unit, and the processor uses the risk sequence generation model, based on the biosignal data received a plurality of times in a predetermined time unit It may be further configured to generate a risk sequence of a predetermined time unit and predict the risk of death based on the risk sequence of a predetermined time unit.

According to another feature of the present invention, the receiver unit includes Glasgow Coma Scale, arterial blood oxygen saturation, inhaled oxygen concentration, bicarbonate ion concentration, bilirubin level, creatinine level, platelet count, total urine output, potassium concentration, sodium concentration, white blood cell count, lactic acid and receive biological test data for at least one subject selected from the group consisting of concentration, APH level, and hematocrit. In addition, the processor may be further configured to generate the sepsis risk sequence based on the biosignal data and the biological test data using the risk sequence generation model, and predict the sepsis risk based on the sepsis risk sequence.

According to another feature of the present invention, the processor may be further configured to provide a risk notification to the entity.

According to another feature of the present invention, the risk sequence generation model may be further configured to calculate a death risk score based on the biosignal data, and the receiving unit may be further configured to receive biological test data for the subject. Furthermore, the risk notification model includes a first notification model configured to output a vector value based on the death risk score calculated by the risk sequence generation model, a second notification model configured to output a vector value based on biosignal data, and at least one of the third models configured to output a vector value based on the biological test data, and the risk of death based on the vector value output by at least one of the first alert model, the second alert model, and the third alert model and a fourth notification model configured to determine whether to provide a notification.

According to another feature of the present invention, the receiver may be further configured to receive the maximum value, minimum value, and average value of the biosignal data measured a plurality of times in a predetermined time unit. In this case, the processor may be further configured to generate the risk sequence of the individual based on the maximum value, the minimum value, and the average value of the biosignal data using the risk sequence generation model.

According to another feature of the present invention, the receiving unit may be further configured to receive the maximum, minimum and average values of the biological test data, measured a plurality of times in a predetermined time unit. In this case, the processor may be further configured to generate the sepsis risk sequence based on the maximum value, the minimum value, and the average value of the biological test data using the risk sequence generation model.

According to another feature of the present invention, the receiving unit may be further configured to receive age data for the subject. In this case, the processor may be further configured to generate the risk sequence of the individual based on the biosignal data and the age data using the risk sequence generation model.

The present invention receives biosignal data related to the state of an individual and furthermore, death of an individual, and provides a risk prediction system based on the received data, thereby rapidly detecting whether an individual has a disease or not, and collecting information related to the individual's life. By providing it, it is possible to predict an emergency situation.

Accordingly, the present invention can provide a good prognosis for treatment by advancing the time of treatment for an individual.

Furthermore, the present invention is capable of predicting with high accuracy the risk of developing sepsis, which provides a high cause of death, as well as the risk of death for an individual by further considering biological test data, which may be clinical data for a biological sample obtained from an individual. It works.

According to the present invention, by providing an early notification according to the predicted risk of death or sepsis, it may be possible to quickly treat an individual in a serious condition such as a critically ill patient according to a predicted risk situation.

The present invention based on the provision of such early notification has effects that can be attributed to improvement of medical service as well as increase in survival rate, prevention of complications, and reduction in treatment cost.

Accordingly, the present invention has the effect of overcoming the limitations of the risk prediction system that cannot provide an appropriate preliminary action according to the prediction information when continuous monitoring is not possible as the predicted information is provided in a display manner.

In particular, the present invention uses a notification sending model configured to determine whether to send a notification by further considering biosignal data and biological test data for an individual, so that frequent notifications that occur when a notification is simply provided based on the level of risk It has the effect of solving the problems associated with it.

The effect according to the present invention is not limited by the contents exemplified above, and more various effects are included in the present specification.

1A illustrates an exemplary risk prediction system using a device for predicting the risk of death or sepsis according to an embodiment of the present invention.
Figure 1b shows the configuration of a device for predicting the risk of death or sepsis according to an embodiment of the present invention.
2 exemplarily shows a procedure of a method for predicting the risk of death or sepsis according to an embodiment of the present invention.
3A to 3D show data for training of a risk sequence generation model used in various embodiments of the present invention.
3E is an exemplary diagram illustrating a pre-processing procedure of data input to a risk sequence generation model used in various embodiments of the present invention.
3F exemplarily shows the configuration of a risk sequence generation model used in various embodiments of the present invention.
Figure 3g shows the training data of the risk sequence generation model used in the device for predicting the risk of death or sepsis according to another embodiment of the present invention.
FIG. 3H exemplarily illustrates the configuration of a risk sequence generation model used in a device for predicting the risk of death or sepsis according to another embodiment of the present invention.
4 exemplarily shows the configuration of a risk risk notification model used in various embodiments of the present invention.
5A and 5B are diagrams illustrating a result of generating a risk sequence according to whether or not death according to the device for predicting the risk of death or sepsis according to an embodiment of the present invention.
5C is a diagram illustrating a result of predicting whether or not death according to the device for predicting the risk of death or sepsis according to an embodiment of the present invention.
5D is a diagram illustrating a result of predicting whether sepsis occurs according to the device for predicting the risk of death or sepsis according to an embodiment of the present invention.
6A illustrates a result of predicting whether or not death according to the device for predicting the risk of death or sepsis according to another embodiment of the present invention.
6B is a diagram illustrating a result of predicting whether sepsis occurs according to a device for predicting the risk of death or sepsis according to another embodiment of the present invention.

Advantages of the invention, and methods of achieving them, will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in a variety of different forms, only these embodiments allow the disclosure of the present invention to be complete, and common knowledge in the technical field to which the present invention belongs It is provided to fully inform the possessor of the scope of the invention, and the present invention is only defined by the scope of the claims.

The shapes, sizes, proportions, angles, numbers, etc. disclosed in the drawings for explaining the embodiments of the present invention are illustrative and the present invention is not limited to the illustrated matters. In addition, in describing the present invention, if it is determined that a detailed description of a related known technology may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. When 'including', 'having', 'consisting', etc. mentioned in this specification are used, other parts may be added unless 'only' is used. When a component is expressed in the singular, the case in which the plural is included is included unless otherwise explicitly stated.

In interpreting the components, it is interpreted as including an error range even if there is no separate explicit description.

Each feature of the various embodiments of the present invention may be partially or wholly combined or combined with each other, and as those skilled in the art will fully understand, technically various interlocking and driving are possible, and each embodiment may be independently implemented with respect to each other. It may be possible to implement together in a related relationship.

For clarity of interpretation of the present specification, terms used herein will be defined below.

As used herein, the term “subject” may refer to a subject whose risk of developing an emergency, for example, death or disease, such as sepsis, is predicted. On the other hand, the subject within the present specification may be a 'patient' or a 'critical patient', but is not limited thereto, and may include all subjects for which the risk of death or sepsis is to be predicted.

As used herein, the term “biological signal data” refers to vital signs and vital signs, and may refer to data related to the state of an individual. Such biosignal data may be associated with the risk of death of an individual, and further, whether or not a disease such as sepsis develops. In this case, the biosignal data may be temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, average blood pressure, and respiration rate of the individual measured by the biosignal measuring device. However, the present invention is not limited thereto, and the biosignal data may include various measurement data related to a health state of an individual. On the other hand, such biosignal data may be time-series data obtained from an individual at any time before the prediction of death or sepsis.

As used herein, the term “biological test data” may refer to clinical data on a biological sample obtained from an individual. Such biological test data may be associated with the risk of death of an individual and further, whether or not a disease such as sepsis develops. At this time, biological test data are Glasgow coma scale (GCS), arterial blood oxygen saturation, inspiratory oxygen concentration, bicarbonate ion concentration, bilirubin level, creatinine level, platelet count, total urine output, potassium concentration , sodium concentration, white blood cell count, lactate concentration, APH level and hematocrit level, but is not limited thereto. On the other hand, the biological test data may be the maximum value or the minimum value among data obtained from the subject at any time before the prediction of death or sepsis.

As used herein, the term “risk sequence generation model” may refer to a model configured to generate a risk sequence based on biosignal data, furthermore, biological test data. Furthermore, the risk sequence generation model may be a model further configured to predict death or sepsis based on the risk sequence.

For example, the risk sequence generation model calculates a risk score based on biosignal data obtained at any time before death or sepsis, and further biological test data, from an individual who has died or developed sepsis, and calculates a risk score, may be a model trained to generate a risk sequence. Furthermore, the risk sequence generation model may be a model further trained to predict the risk of death or sepsis based on the risk sequence.

Meanwhile, the risk sequence generation model of the present invention may be a prediction model based on a deep learning algorithm based on a convolutional neural network (CNN). For example, the risk sequence generation model is based on an analysis module configured to analyze risk through Conv step, BatchNorm and ReLu standardization step, MaxPool and Dropout step based on input biosignal data and Fully-based on biological test data. It may consist of an analysis module configured to analyze risk through connected and BacthNorm steps. In this case, the value output by each analysis module in the risk sequence generation model is combined through a logit step that is converted into a probability logo, and finally, the risk of death or sepsis for the individual can be predicted.

However, the structure of the risk sequence generation model of the present invention is not limited thereto, and may be based on various algorithms capable of predicting the risk of death or sepsis based on data for an individual. For example, the risk sequence generation model of the present invention may be a model based on a clustering algorithm algorithm such as k-Means or SOM algorithm to form a pattern based on biosignal data and/or biological test data obtained from an individual. may be

As used herein, the term "risk notification model" refers to an individual that is expected to be a high-risk group based on time-series data such as a risk sequence, biosignal data, biological test data, and drug administration record generated by the risk sequence generation model. It may be a model configured to provide notifications for

More specifically, the risk notification model learns and notifies the phenotype of the risk score calculated in the risk sequence generation step, the biosignal data for the individual, further biological test data, the time series data of the antibiotic/hypertensive administration record, and the notification sending record. It may be a model configured to determine whether to send. Accordingly, the risk notification model of the present invention can reduce false notifications compared to when a notification is simply provided based on the critical level of the risk sequence. That is, as the risk notification model further considers various data, it is possible to provide an accurate notification when an action is required for an entity.

Meanwhile, the risk notification model of the present invention may be a model based on a deep learning algorithm of a Recurrent Neural Network (RNN) or CNN.

For example, the risk notification model is a plurality of models in which a plurality of notification models composed of independent RNN units and a notification model composed of a layer that determines whether to send a notification based on a value output to each unit are multi-layered. can be configured.

More specifically, the risk notification model is a first of the RNN unit configured to receive the time series data of the death risk score, the antibiotic/hypertensive administration record, and the notification sending record calculated by the risk sequence generation model as input and encode it into a multidimensional vector. a notification model, a second notification model of the RNN unit configured to receive time-series biosignal data and encode it into a multidimensional vector, and a third notification model of the RNN unit configured to receive and encode biological test data into a multidimensional vector. can Furthermore, the risk notification model may further include a fourth notification model of the Dense layer configured to determine whether to send a notification by outputting 0 or 1 based on the output value of each RNN unit.

In this case, each of the first notification model, the second notification model, and the third notification model may be configured as an RNN unit configured with a two-layer long short-term memory (LSTM). Furthermore, data input to each of the first notification model, the second notification model, and the third notification model may be converted into a linear matrix through an embedding layer before being input to each model. In addition, the output value is finally input to one Dense layer by the RNN unit of each model to be concatenated, and output as 0 or 1 through the sigmoid function to finally determine whether to send a notification.

For example, if the value finally output by the fourth notification model is 0, the risk notification model does not provide an alarm, and if the value output by the fourth notification model is 1, it may be configured to provide a notification. have.

On the other hand, according to various embodiments of the present invention, the risk notification model of the present invention includes at least one of a first notification model, a second notification model, and a third notification model that encodes a vector value based on various time series data; and a fourth notification model configured to determine whether to provide a death risk notification based on a vector value output by the at least one model.

That is, the risk notification model may be configured to determine whether to send a risk notification based on more diverse data obtained from an entity as well as a risk sequence. Therefore, the device for predicting the risk of death or sepsis of the present invention can send an alarm more precisely by using the risk notification model.

Preferably, the risk notification model may include, but is not limited to, a first notification model, a second notification model, a third notification model, and a fourth notification model.

Furthermore, the structure of the risk alert model may be based on various machine learning algorithms, as long as it provides alerts to high risk groups based on risk sequences and/or data about entities. For example, the risk notification model of the present invention is based on an algorithm by machine learning of a Randomized Decision forest algorithm, a Penalized Logistic Regression algorithm, and more diverse deep learning algorithms, an input risk sequence score, further biosignal data and/or It may be configured to determine (classify) whether to send an alert based on biological test data.

Hereinafter, a device for predicting the risk of death or sepsis and a risk prediction system using the same according to an embodiment of the present invention will be described in detail with reference to FIGS. 1A and 1B . 1A illustrates an exemplary risk prediction system using a device for predicting the risk of death or sepsis according to an embodiment of the present invention. Figure 1b shows the configuration of a device for predicting the risk of death or sepsis according to an embodiment of the present invention.

Referring to FIG. 1A , the risk prediction system 1000 according to an embodiment of the present invention provides the device 100 for predicting the risk of death or sepsis, the biosignal data 310 and the biological test obtained for the subject 200 It is composed of data 300 including data 320 , biosignal measuring device 400 providing biosignal data 310 , and medical staff device 500 .

More specifically, the device 100 for predicting the risk of death or sepsis in the risk prediction system 1000 receives the biosignal data 310 measured with respect to the subject 200 and generates a risk sequence for the subject 200 . and may be configured to predict the risk of death based on this. Furthermore, the device 100 for predicting the risk of death or sepsis may further consider the biological test data 320 for the subject 200 to generate a risk sequence and further predict the risk of sepsis for the subject 200 .

At this time, the biosignal measuring device 200 transmits biosignal data of at least one of the group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, mean blood pressure, and respiration rate to the subject 200 to predict the risk of death or sepsis may be configured to transmit to device 100 .

On the other hand, the device 100 for predicting the risk of death or sepsis in the risk prediction system 1000 includes a risk sequence for an individual (or a risk score constituting a risk sequence), furthermore, biosignal data 310 and/or biological test data It may be further configured to determine whether to send a notification based on 320 .

Accordingly, the device 100 for predicting the risk of death or sepsis may provide a notification to the medical staff device 500 about the individual 200 requiring an action. Accordingly, the medical staff may recognize the notification and provide feedback such as taking an action according to the symptoms of the individual 200 .

More specifically, referring to FIG. 1B , the device 100 for predicting the risk of death or sepsis includes a receiving unit 110 , an input unit 120 , an output unit 130 , a storage unit 140 , and a processor 150 . .

Specifically, the receiver 110 may be configured to receive various data 300 of the biosignal data 310 and the biological test data 320 for the subject 200 . For example, the receiving unit 110 receives the biosignal data 310 of the temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, mean blood pressure, and respiration rate for the object 200 from the biosignal measuring device 200 . can do. Further, the receiving unit 110, the Glasgow coma scale, arterial blood oxygen saturation, inhaled oxygen concentration, bicarbonate ion concentration, bilirubin level, creatinine level, platelet level, total urine excretion amount, potassium concentration, sodium concentration, leukocytes for the subject 200, and biological test data 320 of levels, lactate concentrations, APH levels, and hematocrit levels.

According to a feature of the present invention, the receiving unit 110 may be further configured to transmit, to the medical staff device 500 , a result predicted for the entity 200 determined by the processor 150 , which will be described later.

According to another feature of the present invention, the receiver 110 may be further configured to receive the maximum value, the minimum value, and the average value of the biosignal data measured a plurality of times in a predetermined time unit. In addition, the receiving unit 110 may be further configured to receive the maximum value, the minimum value, and the average value of the biological test data measured a plurality of times in a predetermined time unit.

According to another feature of the present invention, the receiving unit 110 may be further configured to receive age data for the individual.

The input unit 120 is not limited to a keyboard, a mouse, a touch screen panel, and the like. The input unit 120 may set the device 100 for predicting the risk of death or sepsis and may instruct the operation of the device 100 for predicting the risk of death or sepsis.

Meanwhile, the output unit 130 may display the biosignal data 310 or the biological test data received by the receiving unit 110 . Furthermore, the output unit 130 may display the risk sequence for the entity generated by the processor 150 and display the emergency situation predicted by the processor 150 . Furthermore, the output unit 130 may be further configured to output a notification sound when a notification transmission is determined by the processor 150 .

The storage unit 140 stores the data 300 of the biosignal data 310 or the biological test data 320 of the object 200 received through the receiving unit 110, and the death set through the input unit 120. Or it may be configured to store an indication of the device 100 for predicting the risk of sepsis. Furthermore, the storage unit 140 is configured to store the risk sequence for the entity 200 generated by the processor 150, which will be described later, and to store the predicted risk level for the entity 200, furthermore, whether to send a notification. However, it is not limited to the above, and the storage unit 140 may store various pieces of information determined by the processor 150 for risk prediction.

The processor 150 may be a component for providing an accurate prediction result of the device 110 for predicting the risk of death or sepsis. In this case, for accurate risk prediction, the processor 150 may be based on a biosignal risk sequence generation model trained to generate a risk sequence based on the biosignal data 310 and/or the biologic test data 320 . For example, the processor 150 generates a risk sequence based on the biosignal data 310 and/or the biological test data 320 using the risk sequence generation model, and calculates the risk of death or the risk of sepsis based on the generated risk sequence. predictable.

In another embodiment, the processor 150 may be further configured to send a notification according to the risk level of the entity 200 by using the risk notification model. For example, the processor 150 may be configured to determine whether to send a notification based on the risk score calculated by the risk sequence generation model and further the biosignal data 310 and/or the biological test data 320 based on the model. can do. Accordingly, the processor 150 may provide a notification when the entity 200 is predicted to be a high risk group by using the risk notification model. Accordingly, the medical staff can take quick action on the subject 200 .

Hereinafter, a method for predicting the risk of death or sepsis according to an embodiment of the present invention will be described in detail with reference to FIG. 2 . 2 illustrates a procedure of a method for predicting the risk of death or sepsis according to an embodiment of the present invention.

Referring to Figure 2, the procedure of predicting the risk of death or sepsis according to an embodiment of the present invention is as follows. First, biosignal data for an object is received (S210). Then, using the risk sequence generation model, a risk sequence for the individual is generated based on the received biosignal data (S220). Next, the degree of risk to the individual is predicted based on the risk pattern (S230), and finally, a notification is provided based on the degree of risk (S240).

Specifically, in the step of receiving the biosignal data ( S210 ), biosignal data for an object may be received. For example, in the step of receiving the biosignal data ( S210 ), at least one biosignal data from the group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, mean blood pressure, and respiration rate for the subject may be received. have.

Meanwhile, according to a feature of the present invention, in the step of receiving the biosignal data ( S210 ), the biosignal data may be received a plurality of times in a predetermined time unit. For example, in the step of receiving the biosignal data ( S210 ), biosignal data for 24 hours may be received from a patient entering the room at 1 hour intervals.

According to another feature of the present invention, in the step of receiving the biosignal data ( S210 ), the maximum value, the minimum value, and the average value of the biosignal data measured a plurality of times in a predetermined time unit may be received.

According to another feature of the present invention, Glasgow coma scale (GCS), arterial blood oxygen saturation, inhaled oxygen concentration, bicarbonate ion concentration, bilirubin (bilirubin) level, creatinine (creatinine) level, platelet count, total urine output, receiving biological test data for the subject at least one selected from the group consisting of potassium concentration, sodium concentration, white blood cell count, lactate concentration, anterior pituitary hormone (APH) level and hematocrit level can be

Meanwhile, according to another feature of the present invention, the step of receiving age data for the individual may be further performed.

Next, in the step of generating the risk sequence ( S220 ), the risk sequence for the individual may be generated using the risk sequence generation model learned to generate the risk sequence based on the biosignal data and/or biological test data. . In this case, the risk sequence generation model may be a model based on a deep learning algorithm forming a plurality of layers, configured to receive biosignal data and/or biological test data, calculate a risk score, and generate a risk sequence based on this.

According to a feature of the present invention, in the step of generating the risk sequence ( S220 ), the death risk sequence may be generated based on the biosignal data by the risk sequence generation model.

According to another feature of the present invention, in the step of generating the risk sequence ( S220 ), the sepsis risk sequence may be generated based on the biosignal data and/or the biological test data by the risk sequence generation model.

According to another feature of the present invention, in the step of generating the risk sequence ( S220 ), the risk sequence of the individual may be generated based on the maximum value, the minimum value, and the average value of the biosignal data using the risk sequence generation model. .

According to another feature of the present invention, in the step of generating the risk sequence ( S220 ), the risk sequence of the individual may be generated based on the biosignal data and the age data using the risk sequence generation model.

Next, in the step of predicting the risk ( S230 ), the risk of death may be probabilistically predicted based on the risk sequence. In this case, in the step of predicting the risk ( S230 ), the risk prediction may be performed by the aforementioned risk sequence generation model.

According to a feature of the present invention, in the step of predicting the risk ( S230 ), the risk of death may be predicted based on the risk sequence. Furthermore, the sepsis risk may be predicted based on the risk sequence. For example, in the step of predicting the risk ( S230 ), the risk of death or the risk of sepsis before any predetermined time may be predicted.

Finally, in the step of providing the risk notification ( S240 ), based on the risk score obtained from the above-described risk sequence generation model, furthermore, biosignal data and/or biological test data, whether to send the notification may be determined.

Meanwhile, in the step of providing the risk notification ( S240 ), the notification may be sent by the learned risk notification model to determine whether to send the notification based on the risk score, further biosignal data and/or biological test data. For example, in the step of providing the notification ( S240 ), a risk score and further biosignal data and/or biological test data may be input to the risk ali model to determine whether to send the notification.

In this case, the risk notification model may be composed of a plurality of RNN units and a plurality of models in which a layer for determining whether to send a notification based on a value output to each unit is multi-layered.

For example, the risk notification model may include a first of the RNN unit configured to receive the time series data of the death risk score, the antibiotic/hypertensive administration record, and the notification sending record calculated by the risk sequence generation model as input and encode it into a multidimensional vector. a notification model, a second notification model of the RNN unit configured to receive time-series biosignal data and encode it into a multidimensional vector, and a third notification model of the RNN unit configured to receive and encode biological test data into a multidimensional vector. can Furthermore, the risk notification model may further include a fourth notification model of the Dense layer configured to determine whether to send a notification by outputting 0 or 1 based on the output value of each RNN unit.

In this case, each of the first notification model, the second notification model, and the third notification model may be configured as an RNN unit configured with a two-layer LSTM. Furthermore, data input to each of the first notification model, the second notification model, and the third notification model may be converted into a linear matrix through an embedding layer before being input to each model. In addition, the final output value by the RNN unit of each model is inputted to and joined to the fourth notification model of one Dense layer, and output as 0 or 1 through the sigmoid function to finally determine whether to send the notification.

As a result, according to the method for predicting the risk of death or sepsis according to an embodiment of the present invention, an emergency situation such as death or sepsis onset is predicted for an individual, for example, a critically ill patient, and any time before the emergency situation occurs A notification may be provided. Accordingly, the medical staff can take effective measures for the subject according to the change of the emergency situation.

Hereinafter, a risk sequence generation model used in various embodiments of the present invention will be described in detail with reference to FIGS. 3A to 3F .

At this time, the risk sequence generation model used in the device for predicting the risk of death or sepsis according to various embodiments of the present invention is not only generated a risk sequence associated with the risk of death or sepsis, but is also learned to predict the risk based on this. can be a model. However, the present invention is not limited thereto, and the generation of the risk sequence and the prediction of the risk may be performed in more various ways.

Furthermore, as the risk sequence generation model of the present invention can receive more diverse data for an individual, it can predict not only sepsis but also more various diseases.

3A to 3D show data for training of a risk sequence generation model used in various embodiments of the present invention.

Referring to FIG. 3A , as data for verification of the risk sequence generation model of the present invention, MICU (medical intensive care unit), SICU (surgical intensive care unit), TICU (trauma intensive care unit), CSRU included in MIMIC-III ICU admission record data of 38597 adult patients obtained from the clinical safety research unit (CCU) and the coronal care unit (CCU) were used. The learning data of the present invention utilized Severance ICU data.

More specifically, referring to FIG. 3B , the intensive care unit admission record data may include patient ID, admission time, discharge time, death time, and the like. In this case, the risk sequence generation model of the present invention may be trained to predict the overall risk associated with death by labeling the death in the room.

Referring to (a) of FIG. 3C , the minimum time point among the culture time point and the time point of the antibiotic administration in the period of 48 hours before and after the culture time, or during the period of 4 days from the time of antibiotic administration, is suspected of infection of a pathogen causing infection such as sepsis. It can be defined as a point in time. Furthermore, referring to FIG. 3c (b), the time point when the SOFA score (Sequential Organ Failure Assessment Score) configured to evaluate organ failure and prognosis for 48 hours before and 24 hours after the suspected infection time is 2 or more, It can be defined as the time of onset of sepsis. At this time, the risk sequence generation model of the present invention can learn to predict the sepsis risk by using the sepsis onset time as a label.

Referring to FIG. 3D , the training data of the risk sequence generation model of the present invention is specifically illustrated. At this time, the learning data is composed of time series data (dynamic data) of vital signs acquired for 24 hours, and biological test data (static data) in which the minimum and maximum values for 24 hours are extracted. can

More specifically, the biosignal data includes temperature, pulse, oxygen saturation (SpO ₂ ), systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean blood pressure (MBP) for an individual acquired for 24 hours. ) and respiration rate.

Furthermore, the biological test data were measured for 24 hours on the Glasgow coma scale (GCS), arterial blood oxygen saturation (SaO ₂ ), inhaled oxygen (FiO ₂ ) concentration, bicarbonate ion (HCO ₃ ) concentration, bilirubin (bilirubin). ) level, creatinine level, platelet level, total urine output (Urine output), potassium (potassium,) concentration, sodium concentration, white blood cell (WBC) level, lactate concentration and hippocampal Minimum and maximum values for each of the hematocrit values may be included.

The risk sequence generation model of the present invention may be trained to calculate a risk score based on the biosignal data for learning and/or biological test data as described above and generate a risk sequence based thereon. Furthermore, the risk sequence generation model may be further trained to predict the risk of death or the risk of sepsis based on the risk sequence.

On the other hand, biosignal data and biological test data for learning the risk sequence generation model are not limited thereto, and may be learned to generate a risk sequence by more diverse data and predict the risk of death or sepsis based on this.

3E is an exemplary diagram illustrating a pre-processing procedure of data input to a risk sequence generation model used in various embodiments of the present invention.

Referring to FIG. 3E (a) , the training data may include positive data and negative data. More specifically, positive data may be data for any time, prior to a predetermined time (K hours) from the onset of death or sepsis. Meanwhile, voice data may be obtained from data sampled for 24 hours at any point in time.

Referring to (b) of FIG. 3E , positive data may have a constant length according to a set arbitrary time. In this case, the negative data may be adjusted to have the same length as the positive data with respect to data sampled for 24 hours at any time point through the preprocessing procedure. For example, when positive data is data acquired for 10 hours, in the preprocessing step, negative data may be corrected to have a length of 10 hours with respect to data sampled for 24 hours.

On the other hand, this pre-processing procedure is not limited to the learning stage of the risk sequence generation model and is not performed. For example, time-series bio-signal data obtained from an individual for an arbitrary time may be corrected so that each bio-signal data has the same length through a pre-processing process.

3F exemplarily shows the configuration of a risk sequence generation model used in various embodiments of the present invention.

Referring to FIG. 3F , the risk sequence generation model 600 of the present invention may be a predictive model based on a deep learning algorithm based on a Convolutional Neural Network (CNN). More specifically, the risk sequence generation model 600 is configured to analyze the risk through the Conv step, the standardization step of BatchNorm and ReLu, the MaxPool and Dropout steps, based on the input biosignal data 310 . and an analysis module 620 configured to analyze risk through Fully-connected and BacthNorm stages based on the biological test data 320 . At this time, the values output by the respective analysis modules 610 and 620 in the risk sequence generation model 600 are finally analyzed by the module 640 that is interpretable through the Logit module 630 that is converted into a probability logo. Risk of death or sepsis can be predicted.

However, the structure of the risk sequence generation model 600 of the present invention is not limited thereto, and may be based on various algorithms capable of predicting the risk of death or sepsis based on data for an individual. For example, the risk sequence generation model 600 of the present invention is applied to a clustering algorithm algorithm such as k-Means or SOM algorithm to form a pattern based on biosignal data and/or biological test data obtained from an individual. It may be a model based.

Meanwhile, according to another embodiment of the present invention, the risk sequence generation model may be learned by more diverse learning data and may have more diverse structures.

Hereinafter, a risk sequence generation model used in a device according to another embodiment of the present invention will be described in detail with reference to FIGS. 3G and 3H .

Referring to FIG. 3G , training data of a risk sequence generation model used in a device according to another embodiment of the present invention is specifically illustrated. At this time, the learning data includes time series data of vital signs obtained at 1 hour intervals for 24 hours, and time series data of biological tests obtained at 1 hour intervals for 24 hours (dynamic data) data), and reference feature data, consisting of a minimum value, a maximum value, and an average value for each of the biosignal data and the biological test data acquired at 1-hour intervals for 24 hours.

More specifically, the biosignal data includes temperature, heart rate, oxygen saturation (SpO ₂ ), systolic blood pressure (SBP), and diastolic blood pressure (DBP) for an individual acquired at one-hour intervals for 24 hours. , mean blood pressure (MBP) and respiration rate.

Furthermore, the biological test data are, Glasgow coma scale (GCS), anterior pituitary hormone (APH) level, bicarbonate ion (HCO ₃ ) concentration, bilirubin level, creatinine ( creatinine) level, platelet level, potassium (potassium) level, sodium (sodium) level, white blood cell (WBC) level, lactate level, and hematocrit level.

As described above, the reference characteristic data may include a minimum value, a maximum value, and an average value for each of the biosignal data and the biological test data acquired at 1-hour intervals for 24 hours.

On the other hand, according to another embodiment of the present invention, the risk sequence generation model of the present invention may further use the age of the subject as learning data in order to predict the risk of sepsis or death.

The risk sequence generation model of the present invention may be trained to calculate a risk score based on the biosignal data for learning and/or biological test data and/or reference feature data as described above and generate a risk sequence based thereon. Furthermore, the risk sequence generation model may be further trained to predict the risk of death or the risk of sepsis based on the risk sequence.

FIG. 3H exemplarily illustrates the configuration of a risk sequence generation model used in a device for predicting the risk of death or sepsis according to another embodiment of the present invention.

At this time, according to another embodiment of the present invention, the risk sequence generation model 600 ′ includes a plurality of layers and a reference to which the age data 330 together with the biosignal data 310 and the biological test data 320 are input. It may be composed of a plurality of layers to which the feature data 340 is input.

More specifically, when biosignal data 310 and biological test data 320 are input to a 1x3 convolution layer, and age data 330 is input to a 1x1 convolution layer, 6 Res blocks (residual block) ) layer, a batch norm (Batch normalization) layer, a dropout layer, and an average pooling layer. At the same time, the reference feature data 330 is input to a density layer, and then goes through a batch norm (Batch normalization) layer and a dropout layer. Thereafter, all data 310 , 320 , 330 are integrated into the density layer and inputted, then the batch norm layer, the dropout layer, and the density layer are repeated twice, and finally associated with the risk of sepsis or death. Results can be output.

The risk sequence generation model 600' of the above structure may provide interpretation possibilities.

Hereinafter, the risk notification model of the present invention will be described in detail with reference to FIG. 4 . 4 exemplarily shows the configuration of a risk risk notification model used in various embodiments of the present invention. The risk notification model of the present invention, used in various embodiments of the present invention, is a Recurrent Neural Network (RNN) , or a model based on CNN's deep learning algorithm.

For example, the risk notification model may have a plurality of RNN units and may be configured as a multi-layered structure of layers that determine whether to send a notification based on a value output to each unit. More specifically, the risk notification model may include a long short-term memory (LSTM) unit configured to receive the death risk score calculated by the risk sequence generation model as input and provide an output value related to whether or not to provide a notification, and biosignal data. The RNN unit may include an RNN unit configured to provide an output value related to whether a notification is provided, and an RNN unit configured to receive biosignal data and provide an output value related to whether or not a notification is provided. Furthermore, the risk notification model may further include a Dense layer (Fully connected layer) configured to output 0 or 1 based on the output value of each LSTM to determine whether to send a notification.

On the other hand, the device for predicting the risk of death or sepsis according to various embodiments of the present invention is configured to provide a notification to an entity expected to be a high risk group based on the risk sequence generated by the risk sequence generation model, a risk notification model can be used to send notifications.

Referring to FIG. 4 , the risk notification model 700 of the present invention may be a model based on a deep learning algorithm of RNN or CNN.

More specifically, the risk notification model 700 includes a first notification model 710 (a), a second notification model 710 (b)) and a third notification model 710 (c), and a plurality of notification models ( 710) The fourth notification model 720 of the Dense layer that determines whether to send a notification based on a value output to each of them may be a model forming two layers.

The risk notification model 700 inputs time series data of risk scores calculated by the above-described risk sequence generation model, drug administration records such as antibiotics/hypertensives, and previously sent notification sending records (binary vector). A first notification model of the RNN unit configured to receive and encode into a multidimensional vector (710 (a)), a second notification model (710 (b)) of the RNN unit configured to receive and encode time-series biosignal data into a multidimensional vector, and and a third notification model 710 (c) of the RNN unit configured to receive time series biological test data and encode it into a multidimensional vector. Furthermore, the risk notification model 700 combines the last output values of each of the plurality of notification models 710 and outputs 0 or 1 through the sigmoid function to finally determine whether to send a notification or not, the fourth notification model of the Dense layer. 720 may be further included.

At this time, each of the first notification model 710 (a), the second notification model 710 (b)) and the third notification model 710 (c) is a two-layer LSTM (Long short- term memory) and may be configured as an RNN unit. Furthermore, the data input to each of the first notification model 710 (a), the second notification model 710 (b)) and the third notification model 710 (c)) are embedded in an embedding layer before being input to each model. can be transformed into a linear matrix through

) 은, Embedding 레이어를 통해 선형 행렬로 변환된 후 (

) 제1 알림 모델 (710 (a)) 에 입력될 수 있다. 나아가, Input 레이어에 입력된, 특정한 시점까지 획득된 7 가지의 생체 신호 데이터들 (

) 은, Embedding 레이어를 통해 선형 행렬로 변환된 후 (

) 제2 알림 모델 (710 (b)) 에 입력될 수 있다. 또한, Input 레이어에 입력된, 특정한 시점까지 획득된 3 가지의 생물학적 시험 데이터들 (

) 은, Embedding 레이어를 통해 선형 행렬로 변환된 후 (

) 제3 알림 모델 (710 (c)) 에 입력될 수 있다.More specifically, four four-dimensional time series data of risk score, antibiotic/hypertensive administration record, and notification sending record input to the input layer (

), after being converted to a linear matrix through the embedding layer (

) may be input to the first notification model 710 (a). Furthermore, 7 kinds of biosignal data input to the input layer and acquired up to a specific time point (

), after being converted to a linear matrix through the embedding layer (

) may be input to the second notification model 710 (b). In addition, three biological test data input to the input layer and acquired up to a specific time point (

), after being converted to a linear matrix through the embedding layer (

) may be input to the third notification model 710 (c).

Here, x_1, x_2, and x_3 denote original data (4-dimensional time series data, biosignals, and biological test data), and x_{e,1}, x_{e,2}, and x_{e,3} are linear It means data converted into a matrix. R denotes the Euclidean space, t denotes the sequence length of each original data, and W denotes the linear matrix defining the embedding layer.

At this time, seven kinds of biosignal data input to the second notification model 710 (b) are SBP, DBP, and MBP blood pressure, pulse (PR), respiration (BT), heart rate (HR), and oxygen saturation (SpO). ₂ ), and the three biological test data input to the third notification model 710 (c) may be a bilirubin level, a lactate concentration, and a creatinine level. However, it is not limited thereto.

Next, the 4D time series data input to the first notification model 710 (a) may be encoded into a multidimensional vector by the following [Equation 1] through the context layer.

[Equation 1]

,

,

Furthermore, the time series data of the biosignal input to the second notification model 710 (b) may be encoded into a multidimensional vector by the following [Equation 2] through the context layer.

[Equation 2]

,

,

In addition, time series data of biological test data input to the third notification model 710 (c) may be encoded into a multidimensional vector by the following [Equation 3] through the context layer.

[Equation 3]

,

,

Here, h denotes a hidden variable (latent variable) at the last time point created through RNN, and d denotes the dimension of the latent variable.

Finally, when vector values finally output from each of the plurality of notification models 710 are input to the fourth notification model 720 of the Dense layer, they are output as 0 or 1 through the sigmoid function of Equation 4 below. .

[Equation 4]

Here, W o means a linear vector for determining whether to send an alarm. Finally, when the value output by the fourth notification model is 0, the risk notification model may not provide an alarm, and may be configured to provide a notification when the value output by the fourth notification model is 1.

On the other hand, according to another feature of the present invention, the risk notification model 700, based on the objective function of the following [Equation 5], cross entropy loss (cross entropy loss) and L2 regularization (L2 regularization) weight prevention ( It may be a model configured to obtain a loss function based on weight decay) and learn the phenotype of time-series data based on this.

[Equation 5]

,

,

는 모델이 예측한 알람 발송 여부 (이진 변수, 0또는 1) 를 의미한다.here,

indicates whether the model predicts whether to send an alarm (binary variable, 0 or 1).

The risk notification model 700 of this configuration can reduce false alarms compared to when notifications are simply provided based on the critical level of the risk sequence. That is, the risk notification model 700 may provide a notification when an action is required for an entity by further considering various data.

Meanwhile, the structure of the risk notification model 700 of the present invention is not limited thereto, and may be based on various machine learning algorithms as long as a notification is provided to a high risk group based on a risk sequence and/or data about an entity. For example, the risk notification model of the present invention is based on an algorithm by machine learning of a Randomized Decision forest algorithm, a Penalized Logistic Regression algorithm, and more diverse deep learning algorithms, an input risk sequence score, further biosignal data and/or It may be configured to determine (classify) whether to send an alert based on biological test data.

Hereinafter, with reference to Example 1, a method for predicting the risk of death or sepsis according to an embodiment of the present invention and a result of evaluation of a device using the same will be described.

Example 1: First evaluation of the risk sequence generation model applied to various embodiments of the present invention

In the assessment, an assessment was performed on a risk sequence generation model, configured to assess the risk of sepsis or risk of death by means of biosignal data and biological test data.

5A and 5B are diagrams illustrating a result of generating a risk sequence according to whether or not death according to the device for predicting the risk of death or sepsis according to an embodiment of the present invention.

Referring to (a), (b), (c), (d), (e), (f), (g) and (h) of Figure 5a, the biosignal data and the biological test data for a living organism The generated sequence is shown. Furthermore, referring to FIG. 5A (i), a risk sequence generated by a risk sequence generation model used in various embodiments of the present invention is illustrated. More specifically, referring to (i) of FIG. 5A , in the case of a living individual, the risk score calculated based on the biosignal data and the biological test data appears to be close to zero. That is, in the case of an individual having a low risk of death, a low risk may be predicted by the risk sequence generation model.

In contrast, referring to (a), (b), (c), (d), (e), (f), (g) and (h) of FIG. 5B , biosignal data for a deceased individual and Referring to the sequences generated for the biological test data, it appears that the data values change significantly compared to the living individuals. Furthermore, referring to (i) of FIG. 5B , the risk score appears to be close to 1 in the risk sequence for the dead individual generated by the sequence generation model. That is, in the case of an individual having a low risk of death, a high risk may be predicted by the risk sequence generation model.

On the other hand, in various embodiments of the present invention, the device for predicting the risk of death or sepsis of the present invention, the risk sequence generated by the risk sequence generation model, more specifically, biosignal data and/or biological signal data for an individual together with a risk score A risk notification model configured to determine whether to send a notification based on the test data is further available.

Accordingly, the device for predicting the risk of death or sepsis of the present invention can reduce false notifications compared to when a notification is simply provided based on the critical level of the risk sequence. That is, the device for predicting the risk of death or sepsis of the present invention using the risk notification model may provide a notification at a time when an action is required for an individual by further considering various data.

5C is a diagram illustrating a result of predicting whether or not death according to the device for predicting the risk of death or sepsis according to an embodiment of the present invention. 5D is a diagram illustrating a result of predicting whether sepsis occurs according to the device for predicting the risk of death or sepsis according to an embodiment of the present invention.

Referring to (a) and (b) of Figure 5c, the AUC (Area Under the Curve) level and AP (Average Precision) value of death predicted by the risk sequence generation model of the present invention are shown. More specifically, referring to (a) of Figure 5c, as a result of evaluation based on data obtained from Severance ICU, the AUC value according to the death prediction 12 hours ago by the risk sequence generation model of the present invention is 0.992, The AP value is 0.909, which is high. In this case, the AUC value may mean a hit rate associated with excellent diagnostic ability. Referring to FIG. 5c (b), as a result of evaluation based on data obtained from MIMIC-III, AUC value according to death prediction 12 hours ago by the risk sequence generation model used in the device for risk prediction of the present invention is 0.849, and the AP value is 0.515.

That is, the risk sequence generation model used in various embodiments of the present invention may be used not only to generate the risk sequence but also to predict the risk of death.

Referring to (a) and (b) of FIG. 5D , the AUC level and AP value of sepsis occurrence predicted by the risk sequence generation model of the present invention are shown. More specifically, referring to (a) of Figure 5d, as a result of evaluation based on data obtained from Severance ICU, the AUC value according to the prediction of the onset of sepsis 6 hours ago by the risk sequence generation model of the present invention was 0.865. , the AP value is 0.276. Referring to (b) of Figure 5b, the result of evaluation based on the data obtained from MIMIC-III (more specifically, the model learned from Severance ICU is evaluated with MIMIC-III data), the device for predicting the risk of the present invention The AUC value according to the death prediction 6 hours before by the risk sequence generation model used in the study is 0.679, and the AP value is 0.166.

That is, the risk sequence generation model used in various embodiments of the present invention may be used not only to predict the risk of death, but also to predict the risk of developing sepsis.

*Example 2: Second evaluation of the risk sequence generation model applied to various embodiments of the present invention

In the assessment, an assessment was performed on a risk sequence generation model, configured to assess sepsis risk or mortality risk by biosignal data, biological test data, and reference characteristic data.

6A illustrates a result of predicting whether or not death according to the device for predicting the risk of death or sepsis according to another embodiment of the present invention. 6B is a diagram illustrating a result of predicting whether sepsis occurs according to a device for predicting the risk of death or sepsis according to another embodiment of the present invention.

6A (a) and (b), the AUC (Area Under the Curve) level and AP (Average Precision) level of death predicted by the risk sequence generation model of the present invention are shown. More specifically, referring to (a) of Figure 6a, as a result of evaluation based on data obtained from the Severance ICU, the AUC value according to the death prediction 12 hours ago by the risk sequence generation model of the present invention is 0.981, The AP value is 0.848, which is high. In this case, the AUC value may mean a hit rate associated with excellent diagnostic ability. Referring to FIG. 6a (b), as a result of evaluation based on data obtained from MIMIC-III, AUC value according to the prediction of death 12 hours ago by the risk sequence generation model used in the device for risk prediction of the present invention is 0.940, and the AP value is 0.723.

That is, the risk sequence generation model used in various embodiments of the present invention may be used not only to generate the risk sequence but also to predict the risk of death.

Referring to (a) and (b) of FIG. 6B , the AUC level and the AP level of the occurrence of sepsis predicted by the risk sequence generation model of the present invention are shown. More specifically, referring to (a) of Figure 6b, as a result of evaluation based on data obtained from Severance ICU, the AUC value according to the prediction of the onset of sepsis 6 hours ago by the risk sequence generation model of the present invention was 0.819. , the AP value is 0.185. Referring to (b) of Figure 5b, the result of evaluation based on the data obtained from MIMIC-III (more specifically, the model learned from Severance ICU is evaluated with MIMIC-III data), the device for predicting the risk of the present invention The AUC value according to the death prediction 6 hours ago by the risk sequence generation model used in the study is 0.835 and the AP value is 0.454.

That is, the risk sequence generation model using the reference feature data as additional learning data, consisting of the minimum, maximum, and average values for each of the biosignal data and the biological test data, is designed to predict not only the risk of death but also the risk of developing sepsis. can be used for

Although the embodiments of the present invention have been described in more detail with reference to the accompanying drawings, the present invention is not necessarily limited to these embodiments, and various modifications may be made within the scope without departing from the technical spirit of the present invention. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but to explain, and the scope of the technical spirit of the present invention is not limited by these embodiments. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. The protection scope of the present invention should be construed by the following claims, and all technical ideas within the equivalent range should be construed as being included in the scope of the present invention.

100: device for predicting the risk of death or sepsis
110: receiver
120: input unit
130: display unit
140: storage
150: processor
200: object
300: data
310: biosignal data
320: biological test data
330: age data
340: reference feature data
400: biosignal measuring device
500: medical staff device
600, 600': risk sequence generation model
610, 620: analysis module
630: Logit module
640: interpretable module
700: risk notification model
710: multiple notification models
710 (a): first notification model
710 (b): second notification model
710 (c): third notification model
720: fourth notification model
1000: risk prediction system

Claims (24)

A method for predicting risk of death or risk of sepsis implemented by a processor, the method comprising:
receiving biosignal data for the subject;
generating a risk sequence of the individual using a risk sequence generation model configured to generate a risk sequence based on the biosignal data;
predicting a risk of death or a risk of sepsis for the subject based on the risk sequence;
calculating a risk score based on the biosignal data using the risk sequence generation model;
determining whether to send a notification based on the risk score using a risk notification model based on a deep learning algorithm, configured to output whether to send a notification by inputting the risk score as an input;
providing a risk level notification to the entity according to whether the notification is sent, and
receiving biological test data for the subject;
The risk notification model is
A first alert model configured to output a vector value based on the risk score, a second alert model configured to output a vector value based on the biosignal data, and a second alert model configured to output a vector value based on the biological test data 3 at least one of the notification models, and
a fourth notification model configured to determine whether to provide a death risk notification based on the vector value output by the at least one model;
The at least one model includes a two-layer long short-term memory (LSTM) that outputs a vector value,
The fourth notification model includes a density layer that determines whether to provide an alarm by integrating the vector values output from the at least one model. According to claim 1,
Receiving the biosignal data includes:
Receiving the biosignal data of at least one of a group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, and mean blood pressure for the subject;
Predicting the risk of death or the risk of sepsis comprises:
Based on the risk sequence, the method of predicting the risk of death or sepsis comprising the step of predicting the risk of death. 3. The method of claim 2,
Receiving the biosignal data includes:
Receiving the biosignal data a plurality of times in a predetermined time unit,
The step of generating the risk sequence comprises:
generating a risk sequence of the predetermined time unit based on the biosignal data received a plurality of times in the predetermined time unit using the risk sequence generation model;
Predicting the risk of death or the risk of sepsis comprises:
A method of predicting a risk of death or sepsis, comprising predicting a risk of death based on the risk sequence of the predetermined time unit. According to claim 1,
Glasgow coma scale (GCS), arterial blood oxygen saturation, inspiratory oxygen concentration, bicarbonate ion concentration, bilirubin level, creatinine level, platelet count, total urine output, potassium concentration, sodium concentration, white blood cell count , receiving said biological test data for said subject at least one selected from the group consisting of lactate concentration, anterior pituitary hormone (APH) level and hematocrit level,
The step of generating the risk sequence comprises:
Using the risk sequence generation model, based on the biosignal data and the biological test data, further comprising the step of generating a sepsis risk sequence,
Predicting the risk of death or the risk of sepsis comprises:
Based on the sepsis risk sequence, the method of predicting the risk of death or sepsis further comprising the step of predicting the risk of sepsis. 5. The method of claim 4,
Receiving the biological test data comprises:
A method of predicting a risk of death or sepsis, comprising receiving a maximum value, a minimum value, and an average value of the biological test data measured a plurality of times in a predetermined time unit. 5. The method of claim 4,
Receiving the biosignal data includes:
A method of predicting the risk of death or sepsis, comprising receiving the biosignal data of at least one of the group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, and mean blood pressure for the subject. delete delete According to claim 1,
Performed before the step of providing the risk notification,
Further comprising the step of receiving a drug administration record or a notification sending record for the subject,
The notification model includes the first notification model and the fourth notification model,
The first notification model is
The method of predicting the risk of death or sepsis, further configured to output a vector value based on the risk score and the drug administration record or the notification sending record. According to claim 1,
The risk of death is
defined as the risk of death occurring before a predetermined time,
The sepsis risk is
A method of predicting a risk of death or sepsis, defined as a risk of developing sepsis before the predetermined time. According to claim 1,
The risk sequence generation model is
Receiving the biosignal data for learning acquired before a predetermined time before the risk level occurs with respect to the sample object;
generating a risk sequence for learning based on the biosignal data for learning; and
A method for predicting the risk of death or sepsis, which is a model trained by predicting the risk for the sample subject at any time before the risk occurs based on the learning risk sequence. 12. The method of claim 11,
The sample subject is a dead subject,
The step of estimating the risk for the sample entity,
Based on the learning risk sequence, the method of predicting the risk of death or sepsis, comprising the step of predicting the risk of death for the sample object. 12. The method of claim 11,
The sample subject is an individual with sepsis,
Further comprising the step of receiving the biological test data for learning obtained before a predetermined time from the onset of sepsis of the sample object,
The step of generating the risk sequence for learning comprises:
Generating a sepsis risk sequence for learning based on the biological test data for learning and the biosignal data for learning,
The step of estimating the risk for the sample entity,
A method of predicting death or sepsis risk, comprising predicting the risk of developing sepsis for the sample object based on the learning sepsis risk sequence. a receiver configured to receive biosignal data for the subject and biological test data for the subject; and
a processor coupled to communicate with the receiver;
The processor is
generating a risk sequence of the subject using a risk sequence generation model configured to generate a risk sequence based on the biosignal data, and predicting a risk of death or a risk of sepsis for the subject based on the risk sequence, and Using the risk sequence generation model, the risk score is calculated based on the biosignal data, and the risk using the deep learning algorithm-based risk notification model is configured to output whether or not to send a notification by inputting the risk score as an input. is configured to determine whether to send a notification based on the score, and to provide a risk level notification to the entity according to whether the notification is sent,
The risk notification model is
A first alert model configured to output a vector value based on the risk score, a second alert model configured to output a vector value based on the biosignal data, and a second alert model configured to output a vector value based on the biological test data 3 at least one of the notification models, and
a fourth notification model configured to determine whether to provide a death risk notification based on the vector value output by the at least one model;
The at least one model includes a two-layer long short-term memory (LSTM) that outputs a vector value,
The fourth notification model includes a density layer that determines whether to provide an alarm by integrating the vector values output from the at least one model, a device for predicting the risk of death or sepsis. 15. The method of claim 14,
the receiving unit
configured to receive the biosignal data of at least one of a group consisting of temperature, pulse, oxygen saturation, systolic blood pressure, diastolic blood pressure, and mean blood pressure for the subject,
The processor is
A device for predicting a risk of death or sepsis, further configured to predict a risk of death based on the risk sequence. 16. The method of claim 15,
the receiving unit
configured to receive the biosignal data a plurality of times in a predetermined time unit,
The processor is
Using the risk sequence generation model, a risk sequence of the predetermined time unit is generated based on the biosignal data received a plurality of times in the predetermined time unit, and death based on the risk sequence of the predetermined time unit A device for predicting risk of death or sepsis, further configured to predict risk. 15. The method of claim 14,
The receiving unit,
Glasgow Coma Scale, arterial blood oxygen saturation, inspiratory oxygen concentration, bicarbonate ion concentration, bilirubin level, creatinine level, platelet count, total urine output, potassium concentration, sodium concentration, white blood cell count, lactate concentration, APH level and hematocrit further configured to receive said biological test data for said subject at least one selected from a group;
The processor is
death or sepsis risk, further configured to generate a sepsis risk sequence based on the biosignal data and the biological test data using the risk sequence generation model, and predict the sepsis risk based on the sepsis risk sequence Predictive devices. delete delete According to claim 1,
Receiving the biosignal data includes:
Receiving the maximum value, minimum value, and average value of the biosignal data measured a plurality of times in a predetermined time unit,
The step of generating the risk sequence of the individual comprises:
Using the risk sequence generation model, based on the maximum value, the minimum value, and the average value of the biosignal data, comprising the step of generating a risk sequence of the individual, a method of predicting the risk of death or sepsis. According to claim 1,
Further comprising the step of receiving age data for the subject,
The generating the risk sequence of the subject further comprises generating the risk sequence of the subject based on the biosignal data and the age data using the risk sequence generation model. prediction method. 15. The method of claim 14,
The receiving unit,
further configured to receive a maximum value, a minimum value, and an average value of the biosignal data measured a plurality of times in a predetermined time unit,
The processor is further configured to generate a risk sequence of the individual based on a maximum value, a minimum value, and an average value of the biosignal data, using the risk sequence generation model. 18. The method of claim 17,
The receiving unit,
further configured to receive a maximum value, a minimum value, and an average value of the biological test data measured a plurality of times in a predetermined unit of time,
wherein the processor is further configured to generate the sepsis risk sequence based on the maximum, minimum and average values of the biological test data using the risk sequence generation model. 15. The method of claim 14,
The receiving unit,
further configured to receive age data for the subject,
The processor is
The device for predicting a risk of death or sepsis, further configured to generate a risk sequence of the individual based on the biosignal data and the age data using the risk sequence generation model.

Images