KR20100062439A - Bayesian network system for incident risk prediction of pressure ulcers - Google Patents

Abstract

PURPOSE: A Bayesian network system for incident risk prediction of bedsore generation is provided to use available clinical data of an EMRS(Electronic Medical Record System), thereby automatically predicting a bedsore of a patient. CONSTITUTION: A database(101) records clinical data of a patient. A bedsore predicting unit(102) defines variables which causes bedsores and the range of each variable. The bedsore predicting unit calculates bedsore generation probability based on the clinical data matched with the variables. The bedsore predicting unit builds a conditional probability table based on the calculated bedsore generation probability.

Description

BAESESIAN NETWORK SYSTEM FOR INCIDENT RISK PREDICTION OF PRESSURE ULCERS}

The present invention relates to a computerized system for predicting the risk of pressure sores through communication with the electronic medical record system.

The techniques that have been mainly used in health risk prediction algorithms in the healthcare field are merely logistic regression models using classical statistical techniques. This is how hypotheses are generated based on the modeler's knowledge and experience, variables are selected, and the final prediction results are divided into two categories: risk or non-risk. Therefore, it depends on the data availability of the patient cases applied and has as a basic assumption the number of relevant variables and the linear relationship between them.

However, due to the nature of the practice of clinical practice, it is difficult to assume that all the values of the necessary variables in the model are collected at one point in time. That is, there is a practical limitation in applying logistic regression model directly to the development of clinical decision support system / expert system that is applied in real time.

In addition, the Braden Scale, Norton Scale, Cubbin & Jackson tools, and modified tools based on these tools, have been developed in relation to the development of pressure sores. These tools are a form of patient assessment record that can be used as an input screen for a separate paper record or electronic medical record system (EMRS), in conjunction with previously collected data such as EMRS or test results. Not. Therefore, there is a need for a separate record or data input and there are practical problems such as reliability between users and user education related to the use of the tool.

The present invention provides a bedside risk prediction Bayesian network system that can support the decision and judgment of the clinician about the development of pressure sores and reduce the incidence of pressure sores.

The present invention provides a bed sores risk prediction Bayesian network system that can automatically predict the development of bedsores in patients by utilizing the available clinical data of the electronic medical record system (EMRS).

Bed pressure risk prediction Bayesian network system according to an embodiment of the present invention includes a database for recording the clinical data of the patient; And a pressure ulcer predictor for defining a variable which is a pressure ulcer causing factor and a category for each variable, and calculating a pressure ulcer for each pressure category based on the clinical data matched with the variable.

In this case, the pressure ulcer predictor may be configured to build a conditional probability table through the relationship between the variables by using the calculated pressure ulcer occurrence rate for each variable category.

In addition, the pressure ulcer predictor may be characterized by applying a single Bayesian network based on clinical data of a single time point.

The pressure ulcer predictor may be configured to apply a dynamic Bayesian network based on clinical data of multiple time points.

According to the present invention, computerization and automation of pressure sores risk prediction using an electronic medical record system (EMRS), it is possible to save human / physical resources required to determine pressure sores.

According to the present invention, it is possible to increase the consistency and accuracy of the risk prediction by reducing the variation between users for determining the risk prediction of pressure sores.

According to the present invention, it is possible to contribute to the quality improvement of services in the hospital by minimizing the frequency of bedsores in patients by actively determining the risk of bed sores.

Hereinafter, with reference to the accompanying drawings will be described in detail the Bayesian risk prediction Bayesian network system according to an embodiment of the present invention.

1 is a diagram showing the internal configuration of the bayescing risk prediction Bayesian network system of the present invention.

Referring to FIG. 1, the pressure ulceration risk prediction Bayesian network system of the present invention may include a database 101 and a pressure ulcer predictor 102.

Embodiments of the present invention may utilize clinical data of a patient to perform pressure sores risk prediction.

To this end, the database 101 may mean a storage means for recording clinical data of a patient, and may use an electronic medical record system (EMRS) including an electronic nursing record system (ENRS).

First, in order to select available data from the clinical data, a process of selecting a variable that is a risk factor of pressure sores is needed.

The selection of the parameters can be carried out by reviewing the literature or interviewing clinical practitioners.

For example, by using keywords related to pressure sores, relevant domestic and international papers can be searched through KERIS (academic research information service provided by the Korea Education & Research Information Service). From the searched papers, research on the development of tools and risk factors for pressure ulceration can be selected to extract related concepts presented as risk factors in each study. Using the concepts extracted through this process, it is possible to identify variables that cause pressure sores and the categories (levels) of each variable.

In addition, practitioners with work experience in surgical intensive care units other than the research team may be able to additionally identify any variables that need further consideration from their experience.

Through the above process, the selected variable can be arranged as shown in Table 1.

Next, as a process of identifying clinical data corresponding to the variable, a process of identifying clinical data matched by the variable is required.

The approach to matching the variables with clinical data can be proceeded according to the following principles and procedures.

First, review clinical data items in structured medical records. For example, the clinical observation record of a nurse is duplicated through a structured input screen such as an intensive care nursing record sheet and an unstructured input screen such as a nursing journal. Therefore, the data items of the structured medical record screen, which are considered to be flexible in clinical data use, are considered first.

Second, if there is no corresponding variable in the structured medical record screen, the core concept of the standard statement used in the nursing journal can be searched. For example, the clinical data items in Table 2 can be additionally considered.

In addition, a process of extracting clinical data matching the variable is required.

For this purpose, target patients are selected and clinical data of each target patient are extracted.

Patients can be selected by referring to clinical observations such as vital signs and patient status, and can be classified into pressure sores and non-beds.

For example, the selection criteria of the pressure ulcer developing group may correspond to a patient who has a hospital stay for three days or more, a patient who has stayed in a surgical intensive care unit (SICU) for two days or more, or a patient who has had a pressure ulcer (except for a patient with a pressure ulcer at the time of admission). The selection criteria of the non-depressive group may correspond to a patient who is 3 days or more in the hospital stay, a patient who has stayed in the SICU for 2 days or more, or a patient who does not have pressure beds.

Clinical data matching the selected variable in the EMRS records of the bedsores and non-bedsorees patients (this may include demographic and hospitalization information, such as the patient's sex, medical department, hospitalization path, hospitalization days) Can be extracted).

Subsequently, the extracted clinical data may have 27,056 to 889,000 input events according to the items as values repeatedly measured by a predetermined time or minute for each patient. Thus, the clinical data is organized through a "cleaning process" according to semantic validity and processing adequacy of the data format. Semantic semantics is about 'Atomic facts' is it possible to interpret the meaning of individual data ?, and the appropriateness of the data type is about 'Are the data types appropriate for processing large amounts of data?' will be.

The extracted processed clinical data consists of data records that are measured repeatedly several times a day. In order to construct the pressure sores prediction model of the present invention, it is necessary to integrate and reconstruct these clinical data meaningfully in a unit time zone.

Integration criteria for the clinical data are as follows.

First, an average value of a continuous variable among the variables is calculated.

Second, in the case of a variable that is a nominal or sequence variable among the variables and exists in one day, negative data values are applied first.

Third, for variables with a missing rate of 50% or more, determine whether it means 'NULL' or 'missing' according to the clinical meaning of the variable and treat it as a separate value.

The clinical data of the pressure ulcer development group and the pressure ulcer non-occurrence group that have undergone the above processes may be stored and maintained through the database 101 so as to be used to implement the pressure ulcer risk prediction model.

The pressure ulcer predictor 102 utilizes the predefined variable and the category of each variable and calculates the pressure ulcer for each category of the variable based on the clinical data stored in the database 101 to calculate the pressure ulcer risk prediction model. Can be implemented.

2 and 3 are diagrams for explaining the variables causing the pressure sores and the category of each variable. Variables for use in the pressure sores risk prediction model may be determined in consideration of the missing rate.

Among the variables, the variables with the highest rate of missing by category are in order of consciousness level, hemoglobin, and edema. The variables can be classified according to the missing rate as follows.

-Variables above 50% missing: consciousness level, hemoglobin, degree of edema

-Variables with a missing rate of more than 10% and less than 50%: body mass index, pulse, body temperature, ventilator mode, systolic blood pressure, diastolic blood pressure, cardiovascular status, cardiovascular severity, dietary type, self motor power, skin Condition, albumin, nursing severity (dose), nursing severity (respiratory system)

-Variables with a missing rate of less than 10%: sex, age, type of admission, risk of bed sores on the first day of ICU, number of days of admission, total length of stay, number of days of ICU stay, treatment department, muscle relaxants, frequency of urine measurement, frequency of stool measurement, storminess, drainage Number of measurements, indwelling catheterization, TPN, posture change, sedative administration, edema, peripheral circulation (S, M, C), anticoagulant administration, sedation administration, analgesic administration, inhibitor

2 and 3 include the results of performing a univariate analysis of the variable and group comparison between the pressure sores developing group and the pressure sore non-occurrence group. Among them, continuous variables are expressed as average (standard deviation) and categorical variables are expressed as frequency (%).

As a result of univariate analysis, some variables may be excluded from model building according to the p-value. For example, insignificant variables under the significance level of 5% can be excluded from model construction by cardiovascular status, peripheral circulation (C) and severity of care (dose).

4 is a diagram for explaining a single Bayesian network model.

The pressure ulcer predictor 102 may build a pressure ulcer risk prediction model by applying a single Bayesian network. The pressure ulcer predictor 102 performs a logic regression analysis for variable selection following the basic search for the clinical data. Variable selection in regression analysis involves designing the structure of a single Bayesian network for predicting the risk of pressure sores after forward and backward stepwise.

For example, the single Bayesian network model consists of 14 nodes, including dependent variables, and the dependencies between these nodes can be represented through 13 links. Accordingly, a total of 1,957 conditional probability tables can be used to construct the model.

Table 3 shows the sensitivity and specificity (large number of clinical data), which are performance indicators of the single Bayesian network model.

As a result of performing the ROC curve and sensitivity analysis to see the overall predictive power of the single Bayesian network model, the overall predictive power is about 85.0% as shown in FIGS. 5 and 4, and in the sensitivity analysis, the positive predictive value is generally low and the negative predictive value is generally low. It can be seen that high.

FIG. 6 illustrates a dynamic Bayesian network model, which illustrates a model composed of five time slices in units of 1 day.

The pressure ulcer predictor 102 may build a pressure ulcer risk prediction model by applying a dynamic Bayesian network. The dynamic Bayesian network model extends the single Bayesian network model and adds time slices based thereon. The clinical data increases in availability over time, but changes with respect to prescription orders such as 24 hours, 12 hours, etc., thereby establishing a Bayesian network model based on the time units.

Referring to FIG. 6, the dynamic Bayesian network model may be represented by 70 nodes and 69 links, and thus may include 16,913 conditional probability tables and 5 time slices.

The data preparation process for training and testing the dynamic Bayesian network model is as follows.

For example, in the clinical data of the database 101, 3,348 patients may be classified into 1st, 2nd, 3rd, 4th, and 5th Days according to the date of admission, and are required for modeling training data and model verification. Testing data can be classified using random methods. That is, according to the method used in constructing the predictive model after generating the random number table for each patient, 2/3 of the total patients can be divided into Train data and 1/3 as verification data.

At this time, train data were randomly divided into 2,348 people (70.1%) and test data by 1,000 people (29.9%), and as shown in Table 5, 14,940 trains 71.8% and test data 5,883 28.3 The percentage of occurrence group of train data is 14.9%, and the number of occurrence group of test data is 14.1%.

The following criteria may be used when testing the constructed dynamic Bayesian network model.

The first is discriminative ability: c statistics can be used as the ability to distinguish high and low risk subjects from pressure sores. In the present embodiment, an area under the receiver operating characteristic (AROC) curve and sensitivity may be used.

The second is calibration: this is the degree of correspondence between the predicted value and the actual observed value, which can be compared with the slop of the linear predictor (calibration slope) presented by Cox.

Third, the overall accuracy is: It is common to compare the Brier score (or average prediction error) with the D statistics calculated from the chi-square model, and Nagelkerke's R2, which represents the explanatory power, but in this embodiment the model that minimizes 1-Specificity. As the selection criteria.

As a result of fitting to the train data set, the incidence of pressure sores by time slices is gradually increased as follows.

As a result of applying the test data set of 1,000 people (29.9%) to the dynamic Bayesian model, the performance index for each time slice is shown in Table 6.

7 is a diagram illustrating AROC curves of 2nd Day and 3rd Day.

As can be seen in Table 6, the 1st, 4th Day shows a similar level of performance to a single Bayesian model, but as shown in Figure 7 shows a better performance than the single Bayesian network model on 2nd, 3rd Day.

Therefore, the present invention can be more accurately judged by directly using the actual clinical data of EMRS for the prediction of pressure sores, and computerized and automated bed sores risk prediction by applying the Bayesian network model based on probability theory. In addition, it can be applied to the implementation of clinical decision support system to prevent the development of bedsores in real-time or in a batch form through a communication between EMRS of medical institutions and various application program interfaces (APIs).

The pressure development risk prediction algorithm applying the Bayesian network model of the present invention may be implemented in the form of program instructions that can be executed by various computer means and recorded in a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the media may be those specially designed and constructed for the purposes of the present invention, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

As described above, the present invention has been described by way of limited embodiments and drawings, but the present invention is not limited to the above embodiments, and those skilled in the art to which the present invention pertains various modifications and variations from such descriptions. This is possible.

Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims below but also by the equivalents of the claims.

1 is a diagram illustrating a configuration of a Bayesian risk prediction Bayesian network system according to the present invention.

2 and 3 are diagrams for explaining the variables causing the pressure sores and the category of each variable.

4 is a diagram for explaining a single Bayesian network model.

5 is a diagram illustrating the area under the receiver operating characteristic (AROC) curve of the single Bayesian network model.

6 is a diagram for explaining a dynamic Bayesian network model.

7 illustrates an AROC curve of a dynamic Bayesian network model.

<Explanation of symbols for the main parts of the drawings>

101: database

102: bedsore prediction unit

Claims (9)

A database for recording clinical data of the patient; And, A pressure ulcer predictor that defines a variable that is a cause of pressure sores and a category for each variable, and calculates the pressure of pressure sores for each category of variables based on the clinical data matched with the variable. Bayesian outbreak risk prediction comprising a. The method of claim 1, The bedsore prediction unit, A conditional probability table is constructed through the relationship between the variables by using the calculated bedsore probability for each category of the variable. The method of claim 2, The bedsore prediction unit, A Bayesian Risk Prediction Bayesian Network System, applying a Bayesian network based on probability theory to the above variables. The method of claim 3, The bedsore prediction unit, Bayesette Risk Prediction Bayesian network system applying a single Bayesian network based on a single point of time clinical data. The method of claim 4, wherein The bedsore prediction unit, After receiving clinical data corresponding to the variable for the diagnosis subject, the clinical symptom is determined by determining the corresponding variable and the category of each variable to provide the probability of the pressure ulcer of the diagnosis subject through the single Bayesian network. Risk Prediction Bayesian Network System. The method of claim 3, The bedsore prediction unit, Bayesette risk prediction Bayesian network system applying a dynamic Bayesian network based on multi-point clinical data. The method of claim 6, After receiving the clinical data corresponding to the variable at regular intervals for the diagnosis target, the variable and the category of the variable corresponding to the clinical data are determined for each cycle, and the cycle of the diagnosis target is performed through the dynamic Bayesian network. A Bayesian Risk Prediction Bayesian network system that provides a probability of developing bedsores. The method of claim 1, The variable is Prediction of bed sores, including at least one of age, body mass index, length of stay, body temperature, ventilator mode, urine count, stool count, diet type, position change count, self-exercise, skin condition, hemoglobin, albumin Bayesian network system. The method of claim 8, The variable is Sex, type of admission, risk of bedsores, number of admissions, number of days spent in hospital, department of clinic, pulse, systolic blood pressure, diastolic blood pressure, cardiovascular status, cardiovascular status severity, muscle relaxant administration, stormy status, drainage measurement, indwelling catheterization, A Bayesian network system for predicting pressure sores, further comprising at least one of: PTN (complete parenteral nutrition), sedative administration, edema, peripheral circulation, anticoagulant use, sedation use, restraint use, and nursing severity.

Images