KR101141103B1 - Method of generating decision rule for clinical diagnosis - Google Patents

Abstract

In the method for generating a clinical diagnosis rule according to the present invention, the test data of a plurality of test items for each of a plurality of patients is input, classified by test items, and sorted according to the size of the value, and the clinical diagnosis results are marked on each test data. Making; Calculating cut-off values for determining fitness and clinical diagnosis for each test item; Generating a clinical diagnosis rule based on a cut-off value for each of the plurality of test items; Detecting the incidence of clinical diagnosis for the clinical diagnosis decision rule; And determining a part of the clinical diagnosis decision rules as the final clinical diagnosis decision rule according to the frequency of clinical diagnosis. The fitness indicates the degree to which the test item is suitable for the clinical diagnosis decision. Detecting the overlapping region of diagnostic results, calculating the separability indicating the degree of separation for each clinical diagnosis result for each of the plurality of test data included in the overlapping region, and calculating the sum of the separability. And the cut-off value is a weighted average value of values of inspection data in the overlapping area.

Medical data, clinical diagnosis decision

Description

Method of generating decision rule for clinical diagnosis

The present invention relates to medical data processing technology, and more specifically, to examine a certain test item for clinical diagnosis determination for a patient, and to determine a clinical diagnosis decision rule for determining a clinical diagnosis according to the test data according to the test. To generate.

Difficulty breathing is a subjective symptom of the patient and can be observed in the form of empty breathing, breathing breathing, cheynestokes breathing, or kussmaul breathing, and is one of the most common chief complaints seen in an emergency room.

Patients who visited the respiratory distress as the main station in the emergency room can be classified into cardiac disease and pulmonary disease. , Pneumonia and lung cancer are the main causes.

Since the causes of respiratory distress are difficult to diagnose in a short time, a clinical expert diagnoses by using a blood test or chest radiograph, and analyzes and discriminates important features from the results of the examined items. I invest a lot of time.

In general, the feature selection method selects a subset of related features from a given data or combines new features to deal with high-level problems by changing them to low dimensions. It is also used as a preprocessing process for problems such as pattern classification and decision making, because it can effectively solve computational complexity or curse of dimensionality as the number of features increases. The method used offers remarkable performance.

However, methods based on neural networks and genetic algorithms adjust several learning parameters (ie, learning rate, momentum, link weight pruning level for neural networks, parameters used in objective function for genetic algorithms, etc.) in selecting relevant features. And has the limitation that the interpretation of the relationships between the selected features is difficult or impossible.

Therefore, these machine learning algorithms are not suitable as a tool for selecting important features in multivariate or high-level clinical data, and the reliability of the selected features and cut-off values compared with the clinical criteria is not appropriate. It had to be further evaluated.

It is an object of the present invention to provide a method for generating a clinical diagnostic decision rule for examining a test item for a clinical diagnosis decision for a patient, and for determining the clinical diagnosis according to the test data according to the test.

In order to achieve the above object, the method for generating a clinical diagnosis rule according to the present invention includes receiving test data of a plurality of test items for each of a plurality of patients, sorting by test item, and sorting according to the size of each test item. Marking clinical diagnostic results in the data; Calculating cut-off values for determining fitness and clinical diagnosis for each test item; Generating a clinical diagnosis rule based on a cut-off value for each of the plurality of test items; Detecting the incidence of clinical diagnosis for the clinical diagnosis decision rule; And determining a part of the clinical diagnosis decision rules as the final clinical diagnosis decision rule according to the frequency of clinical diagnosis. The fitness indicates the degree to which the test item is suitable for the clinical diagnosis decision. Detecting the overlapping region of diagnostic results, calculating the separability indicating the degree of separation for each clinical diagnosis result for each of the plurality of test data included in the overlapping region, and calculating the sum of the separability. And the cut-off value is a weighted average value of values of inspection data in the overlapping area.

According to the present invention, it is possible to obtain reliable results in generating a clinical diagnosis decision rule for examining a test item for determining a clinical diagnosis for a patient and determining the clinical diagnosis according to the test data according to the test. It is effective.

<Objects and Data Collection>

The subjects of this study were the medical records of 1,129 patients who visited the D medical center in D metropolitan city between July 2006 and June 2007 as the main shelter for respiratory distress. Excluding the subject's personal information, data such as registration number, sex, age, date and time of emergency visit, medical result, hospitalization diagnosis, and initial test items were extracted from the data warehouse.

Initial tests included common blood cell & differential count (CBC & diff.count), prothrombin time (PT), activated partial thromboplastin time (aPTT), serum electrolyte (serum electrolytes), routine admission for patients, serum amylase, blood pH and gas, lipase, CK-MB, Troponin I, CK, lactate dehydrogenase (CRP), fibrinogen, ^{Ca 2 + (calcium), Mg} 2 + (magnesium), there is a Pro-BNP.

Among the collected data, patients who were transferred to another hospital, patients who died after DOA (death on arrival), CPR (Cardio-Pulmonary Resuscitation) or DNR (Do Not Resuscitate), their discharge or unknown other patients, medical records Data from a total of 668 patients (500 inpatients and 168 outpatients) were analyzed except for incomplete cases.

<How to create a diagnostic rule using spatial distribution of data>

A method for generating a clinical diagnosis decision rule according to a preferred embodiment of the present invention will be described with reference to the flowchart of FIG. 1. The processing process according to the method for generating a clinical diagnosis decision rule according to the present invention is implemented to be operated by a computing system. Here, the computing system is composed of a general server-client system. For example, a plurality of user terminals requesting the generation of a clinical diagnosis decision rule for a patient with a specific disease (eg, a patient having difficulty in breathing), a clinical diagnosis decision server connected to a network of the plurality of user terminals, and a specific one. It may consist of a clinical diagnostic database (DB) that stores test data for each patient of the disease (eg, patients with difficulty breathing). Accordingly, when the user requests the clinical diagnosis decision server to generate a clinical diagnosis decision rule for respiratory distress through any one of the plurality of terminals, the clinical diagnosis decision server receives test data from the clinical diagnosis database as follows. Similarly, the clinical diagnosis rule generation process is performed. The clinical diagnosis database (DB) stores the patient's personal information, registration number, gender, age, the date and time of the emergency room visit, the results of the examination, the diagnosis at the time of hospitalization, the initial examination items and related items. For example, test data from patients with dyspnea were the mainstay.

The clinical diagnosis decision server reads the patient's examination data from the clinical diagnosis database when there is a request for generating the clinical diagnosis decision rule of the user. The input data received by the clinical diagnosis determination server is composed of test data of a plurality of patients, the test data of each patient is composed of a plurality of test items, and the clinical diagnosis result of each patient is determined.
The clinical diagnosis determination server classifies and sorts the received test data of a plurality of patients by test items, and marks the result of clinical diagnosis on each test data (step 100). 2 illustrates this.

delete

The graph illustrated in FIG. 2 shows leukocyte counts for a number of patients, with circled data representing the leukocyte count of the discharged patient and squared data representing the leukocyte count of the hospitalized patient.

In this way, the test data of a plurality of test items for each of the plurality of patients is input, sorted by the test item, and sorted according to the size of the value, and each test data is marked for hospitalization or discharge of clinical diagnosis. Here, the hospitalization or discharge is a clinical diagnosis result when the test data of the patient is input, it is apparent from the present invention that the clinical diagnosis can be changed according to the type and purpose of the input data.

As described above, when the test data of a plurality of test items for each of the plurality of patients is input and classified according to the test items, sorted according to the size of the value, and the clinical diagnosis results are marked on each test data, the clinical diagnosis results overlap. There is a realm. That is, in the example of FIG. 2, the examination data of the hospitalized patient and the hospitalized patient is overlapped in the area A. FIG.

Based on the overlapping areas, the clinical diagnosis decision server evaluates the suitability for the clinical diagnosis decision. That is, for each test data included in the overlapping area, a separability degree indicating the degree of separation that can be separated for each clinical diagnosis result is calculated, and the sum of the separability degrees is calculated to determine the test item according to the test data. Determine fitness indicating how well suited to the clinical diagnosis decision.

Through this process, it is possible to calculate the goodness of fit of the clinical diagnosis for all the test items, and determine the rank of the test items for the corresponding clinical diagnosis based on the calculated goodness of fit.

When the fitness for each test item is determined, the clinical diagnosis determination server calculates a cutoff value for determining the clinical diagnosis result for each test item (step 102). The cut off value is a weighted average value of the values of the inspection data in the overlap region.

As described above, when the fitness for each test item and the cut-off value for each test item are determined, a rule pattern for determining a clinical diagnosis through the test items may be generated.

That is, after the clinical diagnosis decision server generates a rule pattern for clinical decision making for the plurality of test items (step 104), the clinical diagnosis decision for the rule pattern for clinical decision making for the plurality of test items is performed. The frequency of occurrence of C is detected (step 106), and a final rule pattern is determined by selecting some of the rule patterns for clinical decision making for the plurality of test items according to the frequency of occurrence (step 108).

The final rule pattern generation process will be described by way of example with reference to FIG. 3, which shows a rule space for white blood cell count and platelet count.

The horizontal space in the rule space of FIG. 3 is marked by sorting the test data of the WBC item among the test data of a plurality of patients according to the number and discharging or hospitalizing a clinical diagnosis to each test data (DISCHARGE OR). ADMISSION), and the vertical space indicates the number of platelet count (PLT) items in the test data of a large number of patients according to the number and marks the data. ADMISSION), and the dashed red line means the cut off value.

If a rule pattern for clinical decision making for the white blood cell count and the platelet count item is generated as follows.

1-1) Rule 1: WBC is Discharge and PLT is Discharge Then Patient is Admission

1-2) Rule 1: WBC is Discharge and PLT is Discharge Then Patient is Discharge

2-1) Rule 2: WBC is Admission and PLT is Discharge Then Patient is Admission

2-2) Rule 2: WBC is Admission and PLT is Discharge Then Patient is Discharge

3-1) Rule 3: WBC is Discharge and PLT is Admission Then Patient is Admission

3-2) Rule 3: WBC is Discharge and PLT is Admission Then Patient is Discharge

4-1) Rule 4: WBC is Admission and PLT is Admission The Patient is Admission

4-2) Rule 4: WBC is Admission and PLT is Admission The Patient is Discharge

When each of the rule patterns for determining the clinical diagnosis is applied to the example of FIG. 3, the frequency of clinical diagnosis is detected, as follows.

1-1) Rule 1: WBC is Discharge and PLT is Discharge Then Patient is Admission with freq. is 20

1-2) Rule 1: WBC is Discharge and PLT is Discharge Then Patient is Discharge with freq. is 10

2-1) Rule 2: WBC is Admission and PLT is Discharge Then Patient is Admission with freq. is 30

2-2) Rule 2: WBC is Admission and PLT is Discharge Then Patient is Discharge with freq. is 10

3-1) Rule 3: WBC is Discharge and PLT is Admission Then Patient is Admission with freq. is 15

3-2) Rule 3: WBC is Discharge and PLT is Admission Then Patient is Discharge with freq. is 20

4-1) Rule 4: WBC is Admission and PLT is Admission The Patient is Admission with freq. is 5

4-2) Rule 4: WBC is Admission and PLT is Admission The Patient is Discharge with freq. is 5

When the frequency of clinical diagnosis is detected, the rule pattern is finally determined according to the frequency of clinical diagnosis.

1) Rule 1: WBC is Discharge and PLT is Discharge Then Patient is Admission

2) Rule 2: WBC is Admission and PLT is Discharge Then Patient is Admission

3) Rule 3: WBC is Discharge and PLT is Admission Then Patient is Discharge

<Generalization>

Now, the process of generalizing the present invention through the equation will be described.

The input data X = {x _i | x _i = (x _i1 , x _i2 , ..., x _in ), i = 1,2, ..., s} is an n-dimensional vector containing s instances, The jth attribute (that is, the check item) is a _j = {(x _1j , x _2j , ..., x _sj }, (j = 1, ..., n) consists of s data pointers, and the output C = {c _k | k = 1,2, ..., m} is a set of m classes (clinical diagnostics).

<Step 1-Extracting Internal Sections of Attribute Values>

j th attribute domain of _{a j d j = [min (} a j), max (a j)] and the internal section of the property value of the j-th attribute corresponding to any Class I _jk = [I ^l _jk, I ^u _jk ], I _jk ∈d _j are extracted. Here, I ^l _jk and I ^u _jk represent the minimum and maximum values of the attribute values belonging to the k th class in the j th attribute.

<Step 2-Detecting overlapping regions between extracted internal sections>

The overlapping region O _j = [O ^lj _j , O ^uj _j ] is found between the extracted internal sections. Where [O ^lj _j , O ^uj _j ] represents the lower and upper bounds between all overlapping regions.

<Step 3-Evaluate fitness by calculating the frequency of the attribute values>

When different classes do not have the same attribute values in the overlapped region, the frequency of the attribute values is calculated according to Equation 1, and the fitness degree of the attribute is evaluated.

In Equation 1, t ^k _j denotes the frequency of unique attribute values included in the k th class in the j th attribute, and s denotes the number of instances. In this case, h ^k _j represents a relative degree of likelihood of separation of attribute values belonging to the k th class in the overlap region, and has a value of h ^k _j ∈ [0,1].

For example, suppose that the distribution of attribute values for each class as shown in Table 1 exists in the overlap region of the j-th attribute when the total number of instances is 20.

In Table 1, the goodness of fit of the j-th attribute may be expressed as H _j = h ¹ _j + h ² j = 0.2 + 0.2 = 0.4, which is the sum of the degree of separation between class 1 and class 2 according to Equation 1. Therefore, it can be seen that the maximum degree of separation of the j-th attribute is 0.4.

Step 4-Determine the cut-off value for each property

In order to determine a cut-off value of each attribute, a weighted average value, i.e., a center of gravity value of overlapping attribute values in the overlapping region is calculated.

In Equation 2, x _ij represents an attribute value of the i th instance in the j th attribute, and n _ij represents a frequency of duplicate attribute values. For example, in the example of step 3, the duplicated property values 1.1, 1.3, and 1.4 for two different classes are 3, 3, and 3, respectively, so the center of gravity calculated by Equation 2 is 11.4 / 9 = 1.2667. In addition, it has a relatively small misclassification (6) of 1.3756 as the center of gravity for the entire overlapping area (7).

The computing system then generates a rule pattern from the given data using the if-then structure (step 106).

Where a _j (j = 1,2, ..., n) is the jth attribute, A _ik (i = 1,2, ..., M; k = 1,2, ..., m) Denotes the interval of the k th class in the i th rule divided based on the cutoff value calculated from Equation 2, and freq ^k denotes the frequency of occurrence of the i th rule pattern in the k th class.

Based on the frequency of occurrence of such a rule for each class, the conflict problem of rules between classes is decomposed according to Equation (3).

R ^* _i in Equation 3 represents the output of the i-th rule pattern, and NA means a rule that cannot be defined by a rule pattern of any class.

1 is a flowchart of a method for generating a clinical diagnosis decision rule according to a preferred embodiment of the present invention.

2 illustrates data in accordance with a preferred embodiment of the present invention.

Figure 3 illustrates a clinical diagnostic decision rule according to a preferred embodiment of the present invention.

Claims (5)

A server system networked with a database storing test data of a plurality of test items for each of a plurality of user terminals and a plurality of patients, If the server has a request from the user terminal, the server, Receiving the test data from the database, classifying the test data into test items, sorting the test data according to the size of the test data, and marking a clinical diagnosis result on the test data; Calculating cut-off values for determining fitness and clinical diagnosis for each test item; Generating a clinical diagnosis rule based on a cut-off value for each of the plurality of test items; Detecting the incidence of clinical diagnosis for the clinical diagnosis decision rule; And determining a part of the clinical diagnosis decision rule as the final clinical diagnosis decision rule according to the frequency of clinical diagnosis and transmitting it to the user terminal. The goodness of fit indicates how well the test item is suitable for determining the clinical diagnosis, Detecting the overlapping area of clinical diagnosis results for each test item, calculating the separability indicating the degree of separation that can be separated for each clinical diagnosis result for each of the plurality of examination data included in the overlapping area, and the separation degree Obtained by calculating the sum of And the cut off value is a weighted average value of values of test data in the overlapping area. The method of claim 1, The clinical diagnosis decision rule, When test data of a plurality of test items is received for a patient, clinical diagnosis is performed according to whether the test data of each test item belongs to the divided clinical diagnosis section based on the cut-off value detected for each test item. Method for generating a clinical diagnostic rule, characterized in that the decision. The method of claim 1, And the cut-off value is calculated according to Equation 4.

In Equation 4, x _ij denotes the i-th data value in the j-th test item, and n _ij denotes the frequency of data in which clinical diagnosis results are duplicated. The method of claim 1, The goodness of fit, And when different clinical diagnosis results do not have the same data values in the overlapping area, calculating the frequency of the data values according to Equation 5, and obtaining the clinical diagnosis decision rule.

In Equation 5, t ^k _j means the frequency of the unique attribute values included in the k th clinical diagnosis in the j th test item, s indicates the number of data, and h ^k _j is the k th clinical diagnosis in the overlapping area. It indicates the degree of relative separation of the test items belonging to and has the value of h ^k _j ∈ [0,1]. A server system networked with a database storing test data of a plurality of test items for each of a plurality of user terminals and a plurality of patients, If the server requests a test item selection for clinical diagnosis determination from the user terminal, the server, Receiving test data of a plurality of test items for each of a plurality of patients from the database, classifying the test items by test items, sorting the test items according to the size of the test items, and marking clinical diagnosis results on each test data; And calculating and guiding the fitness for each inspection item. The goodness of fit indicates how well the test item is suitable for determining the clinical diagnosis, Detect areas where clinical diagnosis results overlap for each test item, The method for generating a clinical diagnosis rule, characterized in that the frequency obtained when different clinical diagnosis results do not have the same data value in the overlapping area is calculated and obtained according to Equation 6.

In Equation 6, t ^k _j denotes the frequency of unique attribute values included in the k th clinical diagnosis in the j th test item, s denotes the number of data, and h ^k _j denotes the k th clinical diagnosis in the overlapping region. It indicates the degree of relative separation of the test items belonging to and has the value of h ^k _j ∈ [0,1].

Images