JP2003296697A - Specimen mix-up determining method, and preparing method for map and learning data used in method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a specimen mix-up determining method capable of considering all inspection items. <P>SOLUTION: Unknown data is arranged on the map clustered into a cluster of specimen mix-up data and a cluster of normal data, and whether nor not a specimen is mixed up is determined by to which cluster the unknown data belongs. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for determining a sample error and a method for creating maps and learning data used in the method.

[0002]

2. Description of the Related Art In recent years, in the field of clinical examination, the development of an automatic analyzer for automatically analyzing a sample has been remarkable, and the sample processing capacity has been improved. For example, in blood tests, a blood analyzer (for example, XE-2100 (manufactured by Sysmex Corporation)) capable of measuring 150 samples per hour is known. On the other hand, the number of samples handled per day has increased with the improvement of the sample processing capacity, and the risk that the user of the automatic analyzer mixes the samples is also increasing. Mistaken the sample means that the sample of one subject A is mistaken for the sample of another subject B. When a doctor decides a treatment method, it is usual to refer to a test result of a sample, and therefore, a mistake of a sample leads to a serious medical error and must be absolutely avoided.

[0003]

Conventionally, in order to determine a sample error, only a specific test item is focused on, and if the test result largely deviates from the previous test result, the sample error occurs. I was re-examined as a possibility of. However, it is difficult to say that the conventional method of determining a sample difference is not a method that considers all test items and that the sample difference can be effectively determined. Therefore, it is an object of the present invention to provide an effective method for determining a sample difference, which can consider all test items.

[0004]

Means for Solving the Problems The method for determining a sample error of the present invention made to solve the above-mentioned problem is to provide unknown data on a map clustered into a cluster of sample error data and a cluster of normal data. Are arranged, and it is determined whether or not the sample is confused by which cluster the unknown data belongs to.

Further, the present invention is a method for creating a map used for determination of sample difference, in which learning data in which normal data and sample difference data are mixed is learned to create a learned map, and a learned map is created. There is provided a map creation method for clustering a cluster of normal data and a cluster of sample difference data.

Further, the present invention is a method for creating learning data used for creating a map used for determination of sample difference, in which sample difference data is created from samples of different subjects, and samples of the same subject are obtained. Create normal data,
Provided is a learning data creation method for mixing sample misplacement data and normal data.

[0007] The difference in the sample means that the sample of one subject A is mistaken for the sample of another subject B. Usually, in the case of a sample of the same subject, the test result has some correlation between the first test result (previous value data) and the second test result (current value data). On the other hand, there is no correlation in the test results even when the samples of different subjects are compared. The present invention provides a method for determining a sample error using this.

The sample difference data is data created from the test result of the sample of a subject A and the test result of the sample of another subject B, and preferably the absolute difference between each item of the test results of both subjects. It is data that vectorized the value. Normal data is data created from the first test result (previous value data) and the second test result (current value data) of a sample of subject A, and preferably for each item of the two test results. It is the data obtained by vectorizing the absolute value of the difference. This data is created in the same manner as the sample replacement data. The unknown data is data created from the test results of any two samples by the same method as the sample difference data and the normal data. This data is the data for determining whether or not the sample is wrong.

[0009]

The map used in the method of the present invention may be an SOM (self-organizing map). As a device for creating a map used in the method of the present invention,
A clustering device can be used, for example, based on a learning data reading unit in which “learning data” including “vector data” consisting of a set of numerical values is read, and a plurality of learning data read in the learning data reading unit. A map creation unit that creates a map in which position information is assigned to "cells" that are regularly arranged on a two-dimensional surface, a distance calculation unit that calculates a "distance" between each cell on the map, and a map A clustering device provided with a distance display unit for visually displaying the distance calculated by the distance calculation unit can be used. According to this clustering device, the learning data reading unit reads “learning data” including vector data consisting of a set of numerical values. The number of learning data to read depends on the number of dimensions of the vector data that constitutes the learning data and the number of cells on the map. Generally, the larger the number of learning data, the better the learning result, but it takes time for learning. Therefore, the appropriate number should be set in consideration of these. For example, in the case of learning data of three-dimensional vector data, it is sufficient to have several thousands.

The map creating section sequentially extracts the learning data from the learning data reading section. In the map creation section, for example,
By sequentially learning using the learning data extracted based on the self-organizing learning algorithm proposed by T. Kohonen, the same as “vector data” is obtained for each cell regularly arranged on the two-dimensional surface. A map, such as an SOM (self-organizing map), to which “reference data” having a meaning as position information composed of dimensional vector values is assigned is created.

The distance calculation unit calculates the distance between cells forming the map. The distance may be any index value as long as it can numerically represent the vectorial difference between the reference vectors of the two cells that are the objects of distance measurement.
For example, “Euclidean distance” described later is used. Also, instead of the “Euclidean distance”, the “inner product” of the vector may be used.

The distance display unit visually displays the distance on the map based on the distance between the cells calculated by the distance calculation unit. As a method of visually displaying the distance, it is preferable to display the distance between cells by the thickness of the boundary line. Further, the line type (line type) of the boundary line may be changed, or the color of the cell may be changed to express the distance.

Another example of a clustering device for creating a map used in the method of the present invention includes "vector data" consisting of a set of numerical values and "attribute data" which is information regarding the type of data. "Learning data"
The learning data reading unit that reads the data and the learning data read by the learning data reading unit performs learning based on the “SOM learning algorithm”, and thus the “cells regularly arranged on the two-dimensional surface “Reference data” consisting of vector data of the same dimension as the learning data is assigned to each cell, and one cell is defined as an “ignition cell” for each learning data and the learning data is assigned to this ignition cell. The SOM creating unit that creates the SOM to which the “firing information” associated with the attribute data included in the SOM is created, and for each cell on the SOM, the “cluster” of each cell based on the firing information allocated to the cell. A clustering device having a classifying unit for determining can be used.

According to this clustering device, the learning data reading unit reads "learning data" including vector data consisting of numerical values. The SOM creating unit sequentially extracts the learning data from the learning data reading unit and sequentially performs learning by using the learning data extracted based on the self-organizing learning algorithm (SOM learning algorithm) proposed by T. Kohonen. "Reference data" composed of vector values of the same dimension as the "vector data" is assigned to each of the cells regularly arranged on the two-dimensional surface.

The SOM creating unit searches for a cell having a reference vector having a vector value closest to the learning data for each learning data, and defines it as an ignition cell.
Then, the SOM creating unit allocates the firing information associated with the attribute data included in the learning data to the firing cell.
Depending on the cell, a plurality of learning data may result in a firing cell multiple times. The cell firing information in that case is accumulated. When learning is performed for all learning data, S
The OM creation unit creates an SOM in which reference data and ignition information are assigned to each cell.

For each cell on the SOM, the classifying section uses a "cluster" of each cell based on the firing information assigned to the cell, for example, using an automatic classification algorithm which will be described in detail later.
And SOM clustering is performed.

The clustering device may further include a comparison table display section for displaying a comparison table showing the relationship between cell firing information and cell clusters. The clustering device may further include a discriminant characteristic value display unit that displays a “discriminant characteristic value” calculated based on the number of cell firings. The clustering device further includes a cluster adjustment unit for adjusting the cluster of cells determined by the classification unit, and the discrimination characteristic value display unit updates the discrimination characteristic value in synchronization with the cluster adjustment by the cluster adjustment unit. You may do it.

As a method of clustering maps, particularly SOMs, a "distance" between cells calculated based on reference data assigned to each cell is calculated, and a "discrimination" calculated based on the number of firings of each cell is calculated. A method of calculating a “characteristic value”, determining the “cluster” of each cell based on either “distance” or “discrimination characteristic value” or both for each cell on the SOM, and performing clustering. May be. At this time, the determination characteristic value is "efficiency"
Any of “sensitivity”, “FNR”, “FPR”, and “specificity” may be included.

[0019]

Embodiments of the present invention will be described below with reference to the drawings. Here, description will be given using an example in which three different types of blood test data (MCV value, MCH value, MCHC value) are used as learning data targets. In addition, by performing learning based on the “SOM learning algorithm”, the map (S
The sample difference determination method using a clustering apparatus for creating OM) will be described as an example.

Configuration of Clustering Device FIG. 1 is a block configuration diagram showing the configuration of a clustering device used in the sample difference determination method according to an embodiment of the present invention. In the figure, 10 is a clustering device, 12 is an input device, 14 is an output device, and is a CPU, ROM, and R.
A hardware configuration is formed by a computer system including an AM, a hard disk, a keyboard, a mouse, a CRT and the like. Each part is connected by BUS.

The clustering device 10 includes a learning data reading unit 16, an SOM creating unit 18, a distance calculating unit 20, a distance displaying unit 22, a classifying unit 24, a comparison table displaying unit 26, a discriminant characteristic value displaying unit 28, and a cluster adjusting unit 30. , Unknown data reading unit 3
2. The evaluation determining unit 34 is used. In addition, the clustering device 10 includes a learning data DB (database) 36 that stores learning data necessary for creating an SOM,
Map data in the process of creating SOM or already completed SOM
Map DB 38 that stores the map data (including ignition information and reference data) of the unknown data, and the unknown data DB 40 that stores the unknown data that is evaluated using the created SOM.
have.

The learning data reading unit 16 is a learning data DB.
A learning data group necessary for creating the SOM is extracted from the learning data accumulated in 36 and read. Each of the learning data read here includes attribute data and three-dimensional vector data composed of a set of three numerical values. Of these, the three-dimensional vector data portion is used for assigning reference data to each cell in the learning process, and the attribute data portion is used when giving firing information to each cell in the learning process. Also,
The attribute data is data required when executing “automatic classification” described later after the reference data is assigned to each cell. Although it is convenient to load the learning data group from the learning data DB 36, the learning data group may be input from the input device 12. It should be noted that the creation of learning data will be further described later.

The SOM creating unit 18 includes the learning data reading unit 1
The learning algorithm for creating the "unclassified SOM" is executed by using the three-dimensional vector data portion of each learning data read by No. 6. "Unclassified SOM" means S
It refers to the SOM in a state where reference data is simply assigned to each cell on the OM and a group (cluster) of similar cells is not divided by a boundary line. The learning algorithm executed here is the SO proposed by T. Kohonen.
An M learning algorithm (self-organizing learning algorithm) is used.

By executing the SOM learning algorithm, the SOM creating unit 18 allocates reference data having the same dimension as the learning data to each cell regularly arranged in the map on the two-dimensional surface, thereby performing clustering. An unclassified SOM that has not been created is created and displayed on the output device 14. Here, the SOM learning algorithm will be described using a general example. Cells 2 (also called neurons, which are divided by regularly arranged squares (not only squares but also hexagons or the like)) are arranged on a map 1 including a two-dimensional region as shown in FIG. Each cell 2
Is specified by a Cartesian coordinate system consisting of x and y coordinates,
In the example of FIG. 3, 32 cells in the x direction and 32 cells in the y direction are arranged side by side, and 32 × 32 = 1024 cells 2 are arranged.
Are arranged. Each cell 2 can be specified by a coordinate system in which the cell at the upper left corner is (0, 0) and the cell at the lower right corner is (31, 31) and the horizontal direction is the x coordinate and the vertical direction is the y coordinate.

Multi-dimensional vector data consisting of a set of numerical values is assigned to each cell 2. The example shown in the center column of FIG. 4 is a case of three-dimensional data, and (a _i , b _i , c _i )
As shown in (a _i , b _i ,
c _i represents an arbitrary numerical value). This data is called "reference data". The initial value of the reference data is generally given by a random number, and is formal data having no meaning or correlation between the respective data.

For this initial "reference data", learning is performed using multidimensional vector data called "learning data" having the same number of dimensions as the above-mentioned reference data. For example, taking a blood test as an example, as shown in FIG.
Blood test data values such as MCH and MCHC values are 0 to 1
Three-dimensional vector data consisting of a set of values normalized to the value of is given as learning data.

SOM learning is performed as follows. First, the first learning data (MCV = 0.6 in the example of FIG. 5)
811, MCH = 0.8084, MCHC = 0.623
When 7) is given, the reference data closest to the learning data is searched from the reference vector group (formal vector group assigned by using random numbers) assigned to 1024 cells. The cell with the closest reference data is called the "firing cell". Selection as the cell having the closest reference data is called firing.

Then, for a plurality of cells in the vicinity of this firing cell (for example, within 2 cells from the firing cell), reference data (vector value) of each cell is set according to a predetermined rule (for example, the vector value is fired at a fixed rate). The reference data is updated by slightly resembling the learning data (approaching the vector value) by a rule such as approaching the cell). With the above, one learning is completed.

After that, when the same learning is repeated using the learning data one after another (for example, about several thousand times), the reference data of each cell is determined as shown in the right column of FIG. At this time, cells having similar reference data are gathered in the vicinity of arbitrary cells. As used herein, "similar reference data" means a vector distance between reference data (for example,
It means that the Euclidean distance described later) is close.

FIG. 10A shows an example of an unclassified SOM created by the clustering device used in this embodiment. In this SOM, the ignition cell, that is, the cell selected as the cell having the reference data closest to the learning data is marked with a mark (ignition information) indicating that the ignition is visually indicated. To This mark is given a different mark for each attribute data included in the learning data so that the firing cell and the attribute data included in the learning data to which the firing cell is assigned can be visually recognized. To do.

There are cases where one cell is ignited by a plurality of learning data. In that case, the number of firings is cumulatively counted. Also, when there is a fire due to two or more learning data having different attribute data in one cell,
Make sure that the cell is known to have fired due to both attributes.

In FIG. 10A, "+""-""±"
Although there is a mark (ignition information) of "0", cells marked with "+" indicate that only learning data having attribute data of "+" ignited that cell, and were marked with "-". The cell indicates that only the learning data having the attribute data of "-" has fired in the cell (the number of firings of the cell may be plural times). The cells with "±" indicate that the learning data having the attribute data of "+" and the learning data having the attribute data of "-" are both fired at least once in the cell. The cells marked with "0" indicate that the learning data having any attribute data has never fired. As will be described in detail later, there are two types of attribute data included in the learning data of the present embodiment, "+" and "-".

The distance calculator 20 calculates the distance between the cells. Here, various values can be used for the distance between the cells depending on how they are defined. In the present embodiment, the following "Euclidean distance" is used as the distance. For the Euclidean distance between cells, the reference vectors of two adjacent cells are (X ₁ , X ₂ , ..., Xi ,.
Xn), (Y ₁ , Y ₂ , ..., Yi, ... Yn), the Euclidean distance is defined as D = √ (Σ (Xi-Yi) ² ).

The Euclidean distance (D) between adjacent cells is calculated for all cells on the SOM. The Euclidean distance D may be divided by its maximum value (Dmax) for normalization (all distances are within the range of 0 to 1).

When the inner product is used as the distance, the distance defined by the following equation is used. D = ΣXiYi

The distance display unit 22 visually displays the distance between cells calculated by the distance calculation unit 20. In this embodiment, the distance is converted into the thickness of the boundary line between cells and displayed. That is, cells with a short distance are separated by a thin boundary line, and cells with a long distance are separated by a thick boundary line so that the distance can be determined by checking the thickness of the line.

More specifically, for example, when assigning four types of line thicknesses according to distances, when the normalized distance (Ds) between cells is 0≤Ds <0.25, "thickness" 1 ", when 0.25≤Ds <0.5," thickness 2 ",
When 0.5 ≦ Ds <0.75, “thickness 3”, 0.75
When ≦ Ds ≦ 1, the thickness of the boundary line between the cells is selected as “thickness 4”.

The method of normalizing the distance is not necessarily limited to dividing by the maximum value. You may standardize by dividing by an average value or dividing by an intermediate value.

FIG. 11A is a diagram showing the state of the SOM when the distance between cells is visually displayed by the thickness of the boundary line. The thickness of the boundary line makes it possible to grasp the distance between the cells on the SOM.

The classification unit 24 performs statistical numerical calculation (calculation of efficiency described later) on the unclassified SOM created by the SOM creating unit 18 using the attribute data to automatically perform clustering. The output device 1 is configured such that the clustering result is partitioned as a boundary on the SOM (color-coded. A lump partitioned by the boundary is referred to as a cluster).
Display on 4. The automatic classification algorithm executed here will be described later.

The comparison table display unit 26 has the number of cells for each cluster when the SOM is clustered by the classification unit 24 (or the classification adjustment unit 30 described later) and the learning data assigned to the cells of each cluster. Display a comparison table that summarizes the total number of firings for each attribute data in tabular form. FIG. 8 is an explanatory diagram showing the configuration of the comparison table. In this embodiment, there are two types of clusters, "class 1" and "class 2". In addition to class 1 and class 2, the number of "undefined" cells for which the type of cluster is undetermined and is not defined, and the number of firings for each attribute data of "undefined" cells I am trying to display it.

The calculation of the total number of firings is performed separately for each attribute data, for example, "+" and "-" in this embodiment. For example, if one cell defined in class 1 is fired into two learning data having attribute data “+” and three learning data having attribute data “−”, “+” of class 1 Is counted as "2", and "-" in Class 1 is counted as 3 counts. This calculation is performed for all cells, and the total number is displayed in the control table.

The discriminant characteristic value display unit 28, based on the ignition information, includes efficiency, sensitivity, specificity, FNR, FPR, which will be described later.
Calculate and display the discrimination characteristic value such as. These discriminant characteristic values are statistical information that can be referred to when adjusting the boundaries of the clusters.

The classification adjusting unit 30 is calculated by the discriminating characteristic value display unit 28 in conjunction with this when manually changing the boundary of the cluster arbitrarily for the SOM after the clustering by the classifying unit 24. Recalculate and display the discriminant characteristic value. In other words, the boundary of the cluster can be arbitrarily changed with reference to the calculation result of the discrimination characteristic value. Then, the SOM after the cluster adjustment is displayed.

In this way, the SOM creating unit 18 and the classifying unit 2
4. In the cluster adjusting unit 30, the contents S are different from each other.
Since the OM is created, when it is necessary to distinguish the SOMs in the following description, the SOM in the state created by the SOM creation unit 18 (unclassified SOM without clustering) is referred to as “primary SOM”. For the sake of convenience, the SOM that is only clustered by the unit 24 is called a “secondary SOM”, and the SOM after the cluster boundary is arbitrarily set by the cluster adjusting unit 30 is called a “third SOM”. To do.

If these SOMs are sequentially stored in the map DB 38, they can be taken out when necessary.

The unknown data reading section 32 reads unknown data to be evaluated. This unknown data is unknown data D
It may be input from B40 or input from the input device 12.

The evaluation deciding unit 34 includes the unknown data reading unit 32.
For the unknown data read from, the firing cell is obtained using the secondary SOM or the tertiary SOM to evaluate and determine in which cluster the unknown data is defined.

Outline of Flow of Specimen Differences Judgment Next, an outline of the flow of processing of the sample mistakes determination of the present embodiment will be described. FIG. 2 is a flow chart for explaining a typical process flow performed by the clustering device 10.

(St101) First, the learning data reading unit 1
6 reads the learning data consisting of the attribute data and the three-dimensional vector data, and proceeds to st102. (St102) The SOM creating unit 18 extracts the three-dimensional vector data one by one from the learning data reading unit 16, and S
The OM learning algorithm is repeatedly executed, and the process proceeds to st103. (St103) The SOM creation unit 18 allocates reference data to each cell based on the result of executing the SOM learning algorithm on all the three-dimensional vector data, and refers to the attribute data to indicate the firing information to each cell. Create a marked primary SOM (unclassified SOM), s
Proceed to t104. (St104) The distance calculation unit 20 calculates the distance between adjacent cells using the reference data assigned to each cell, and proceeds to st105. (St105) The distance display unit 22 selects the thickness of the boundary line between cells based on the distance calculation result by the distance calculation unit 20, and displays it on the primary SOM. Then, when clustering is started, the process proceeds to st106. (St106) The classification unit 24 executes an automatic classification algorithm described later, and proceeds to st107. (St107) The classification unit 24 clusters each cell based on the result of the automatic classification algorithm, and makes it possible to identify each cluster (by color coding or the like) secondary SO.
Create M. The discriminant characteristic value display unit 28 calculates the discriminant characteristic value and displays it on the output device 14. Then, when performing cluster adjustment, it progresses to st108. When performing evaluation on unknown data, the process proceeds to st110. (St108) The cluster adjustment unit 30 receives the adjustment of the boundary of the cluster with respect to the secondary SOM. The discriminant characteristic value display unit 28 recalculates the discriminant characteristic value when the boundary of the cluster is adjusted, and displays it on the output device 14. (St109) The cluster adjustment unit 30 creates a third-order SOM in which the boundaries of clusters are arbitrarily set based on the discrimination characteristic value. Then, when evaluating unknown data, st110
Proceed to. (St110) The unknown data reading unit 32 reads unknown data and proceeds to st111. (St111) The evaluation determining unit 34 searches the firing cell for the unknown data to be evaluated using the secondary SOM or the tertiary SOM, determines the firing cell, and proceeds to st112. (St112) The evaluation determination unit 34 determines the attribute of the unknown data based on the cluster including the firing cell.

Details of Process Flow of Specimen Differences Judgment Next, the details of the process flow of sample wrongness determinations of this embodiment will be sequentially described by using an example in which blood test data is used as learning data.

Example of Creating Learning Data (Example of Creating Sample Conflict Data and Normal Data) FIG. 7 shows an example of a learning data group used in the clustering apparatus 10. As you can see in the figure, the learning data group is
One unit is a pair of "attribute data" representing the type of data, and "three-dimensional vector data" consisting of a set of three numerical values (MCV value, MCH value, MCHC value standardized to a value of 0 to 1). A large number of data to be collected (in this embodiment, 45
00 data). FIG. 7 shows a part thereof.

In the present embodiment, since cells on the primary SOM are defined in two clusters, attribute data ("+" and "-") of two different types are given in the attribute data column. When defining cells on the SOM into three clusters, attribute data of three different types (for example, “A”, “B”, and “C”) are given in the attribute data column. This attribute data is generally given as attribute data examined by another known method. As described above, the three-dimensional vector data is converted into SO by the SOM learning algorithm.
It is used to obtain reference data to be assigned to each cell of M.

In the sample difference determination of the present embodiment, the blood test data of 3000 patients having both the previous value data which is the previous test result and the current value data which is the current test result is used. As shown in FIG.
The inspection data numbers of 0 persons are respectively S ₁ to S ₃₀₀₀ ,
S _1a the previous value data for the S _1, the current value data S
_1b , likewise S _2a , S _{2 b} , ... S _ma , S _mb , ...
_Called S _na , S _nb , ..., S _3000a , S _3000b . Data S _1a , S _1b , S _ma , S _mb , etc. are MCV (mean red blood cell volume), MCH (mean red blood cell hemoglobin amount), MC
It is three-dimensional vector data composed of a set of values obtained by normalizing the numerical values of three test items consisting of HC (mean red blood cell hemoglobin concentration) to values of 0 to 1.

Then, for each of S ₁ to S ₃₀₀₀ , the difference between the present value and the previous value, S _1b -S _1a , S _mb
-S _ma, by calculating the S _3000b -S _3000a, creating a total of 3000 data (three-dimensional vector data). However, each component in the vector data takes an absolute value. The 3000 pieces of data thus created are treated as being included in the attribute of “normal data” in which the samples are not mixed up. "-" Attribute data is combined with this normal data.

Subsequently, 1500 pairs are formed from 3000 pieces of data of S ₁ to S ₃₀₀₀ . For example, it is assumed that the inspection data Sm and Sn are paired. In this case, using the current value of the previous value and the Sn of Sm, the S _nb -S _ma, using the previous value of the present value and Sn of Sm, calculates the S _mb -S _na.
The same calculation is performed for the remaining 1499 pairs. In this way, 3000 pieces of data (three-dimensional data) are created.

These 3000 pieces of data are treated as being included in the attribute of "mistake data" in which an error has occurred. Attribute data of "+" is combined with this mistake data.

Subsequently, the above-mentioned 3000 normal data and the after-mentioned 3000 misconfiguration data are mixed to form an aggregate of 6000 data. And 6
Randomly 4500 data from 000 data (2250 normal data, 2250 misplaced data)
Select. The 4500 pieces of data are used as learning data for creating the primary SOM (the remaining 1500 pieces of data are unknown data for evaluation and the clustering device 10).
Will be used for performance confirmation).

The learning data created in this way is
Learning for artificial judgment by artificially creating attribute data that indicates whether it is mistaking data (“+”) or normal data (“−”) and attaching the created attribute data to the attribute data field. This is data.

The learning data of FIG. 7 described above shows a part of the 4500 pieces of data created in this way, and “+” in the attribute data column of FIG. "Indicates normal data. The numerical values of the MCV value, MCH value, and MCHC value seen in FIG. 7 are standardized so that the maximum value in each item is 1.

Execution of SOM Learning Algorithm The SOM creating section 18 executes the SOM learning algorithm. Since the SOM learning algorithm has already been described, the description is omitted. As the three-dimensional vector data, the three-dimensional vector data (MCV value, MCH value, MCHC value), a part of which is shown in FIG. 7, is used.

When the execution of the SOM learning algorithm is completed, 1024 cells are fired 4500 times in total, so that some cells are fired a plurality of times. In each cell, as the ignition information, the number of times the cell is fired is stored together with the attribute data for each type of attribute data.

Display of Primary SOM (Unclassified SOM) The primary SOM created by executing the above SOM learning algorithm is shown in FIG. One of “+”, “−”, “±”, and “0” is attached to all cells as ignition information.

A control table at this time is shown in FIG. 10 (b).
Since no cluster has been defined for any cell yet,
All cells (1024) are “undefined”. The 1024 undefined cells indicate that the learning data having "+" attribute data has fired 2250 times and the learning data having "-" attribute data has fired 2250 times.

Visualization of inter-cell distance FIG. 11 is a diagram showing the state of the primary SOM when the inter-cell distance is visually displayed by the thickness of the boundary line. The thickness of the boundary line allows the distance between the cells on the SOM to be grasped. The boundary display is performed by the distance display unit 22. The visualization of the inter-cell distance may be similarly displayed in the secondary SOM and the tertiary SOM.

Generation of Secondary SOM by Executing Automatic Classification Algorithm Next, clustering will be described. Clustering is
The classification unit 24 performs this. FIG. 9 is a flowchart illustrating the operation of clustering. Further, FIG. 12 to FIG. 22 are SOMs, comparison tables, which explain states in the middle of clustering operation,
It is a figure explaining the display screen of a discriminating characteristic value. Although the comparison table and the discrimination characteristic value are always displayed at the same time as the SOM, only the necessary states are shown in the figure for convenience of explanation.

In this embodiment, "+" (mistake data)
Since the learning data having the two types of attribute data, "-" and "-" (normal data), is used, the SOM only needs to be able to define two clusters by clustering. Now, the two clusters created by clustering are "class 1" (shown in red on the screen),
We will call it "Class 2" (shown in blue on the screen),
Class 1 is a cluster where learning data having “+” attribute data (that is, mistake data) is originally collected, and class 2 is a cluster where learning data having “−” attribute data (that is, normal data) is originally gathered. To do.

First, one cluster is selected. Here, class 1 is selected (st201). All cells with undefined clusters (initially all cells are undefined)
Is temporarily set as belonging to class 1 (red) (st20
2). FIGS. 12A and 12B show the SOM and the control table in a state where all the cells are set to “class 1” (all cells turn red).

For all cells belonging to class 1,
The number of times a neighboring cell of that cell fired learning data having "+" attribute data and the number of firing of learning data having "+" and "-" attribute data (in this embodiment, the total number of firings of neighboring cells). ) Is calculated, and the ratio (called selection ratio) is calculated. Here, the neighboring cell can be defined as a cell that is included within one square around the one cell, for example (st203).

The cell with the smallest selection ratio is removed from class 1 (st204). FIG. 13A shows a state in which the cell at the coordinate (23, 28) in the SOM deviates from “class 1” (red) and becomes “undefined” (white). FIG.
(B) shows a control table at this time. It can be seen that the number of firings of the "-" attribute was 5 and the number of firings of the "+" attribute was 0 in the cell of the undefined coordinate (23, 28).

Next, "efficiency" defined by the following equation (1)
(Example of discrimination characteristic value) is calculated and stored together with the state of the cell at that time (st205). Efficiency will mean the percentage that is correctly clustered. Efficiency: (TP + TN) / (TP + TN + FP + FN) (1) Formula TP (True Positive): The cells belonging to the selected cluster (class 1 here) are selected The number of firings by learning data (attribute data here is “+”) having attribute data originally possessed by the cluster. TN (True Ne
gative): Learning data in which cells belonging to unselected clusters (here, "class 2" and "undefined") have attribute data that the selected cluster does not originally have (attribute data here is "-"). ). FP (fault Positive): Learning data in which cells belonging to unselected clusters (here, “class 2” and “undefined”) have attribute data originally possessed by the selected clusters (the attribute data here is “ + ") The number of fires. FN (fault Negative): A cell belonging to a selected cluster (here, "class 1") is fired by learning data (attribute data here is "-") having attribute data that the selected cluster does not originally have. The number of times you did it.

The steps from st203 to st205 are repeated until all cells belonging to class 1 are exhausted (st2
06). 14 and 15 show how cells are sequentially changed from "class 1" to "undefined".

Further, FIG. 16A shows a state in which all cells are changed from "class 1" to "undefined".
FIG. 16B shows a comparison table at this time,
All of the 1024 cells are “undefined”, and the undefined cells have 2250 firings due to the learning data having “+” attribute data and 2250 firings due to the learning data having “−” attribute data. It shows that it is a time.

Next, the stored "efficiency" data is searched for the data with the maximum "efficiency", and when the "efficiency" is the maximum, the cell defined in "class 1" is determined as "class 1". (St207, st208).

FIG. 17 (a) is a diagram showing the SOM when the "class 1" is determined by obtaining the maximum "efficiency" state. The cells determined to be "class 1" are painted red, and the remaining cells are "undefined" so they are white. FIG. 17
(B) shows the control table at this time. In this example, 597 cells are defined as "class 1". In the cells included in “class 1”, the number of firings by the learning data having the attribute data “+” is 1997, and the number of firings by the learning data having the attribute data “−” is 424. The remaining 427 not defined in "Class 1"
Cells are "undefined".

An example of calculating the efficiency at this time is shown below. TP: 1997 TN: 1826 FP: 253 FN: 424 Efficiency: (TP + TN) / (TP + TN + FP + FN) = (1997 + 18
26) / (1997 + 1826 + 253 + 424) =
0.8496

Next, it is confirmed whether all clusters are defined (st209). Exit if all clusters are defined. In the present example, class 2 is undefined, so the process returns to st202 and the same calculation is repeated. This time, class 2 is selected (st210). All the remaining 427 undefined cells (that is, cells other than the 597 cells determined as "class 1") are provisionally set as belonging to the cluster of class 2 (blue) (2
Order st202).

Hereinafter, the operations of st203 to st206 are repeated. 18A and 18B are SOM screens in which all 427 cells are set to “class 2” (597 cells of “class 1” are red and the remaining 427 cells are blue) And the control table is shown.

FIGS. 19A and 19B show a state in which the cell having the smallest selection ratio is excluded from “class 2” (the second order st204). In other words, the cell at the coordinate (13, 16) in the SOM is out of the cluster of “class 2” (blue) and is “undefined” (white). FIG. 19B shows a control table at this time. The number of firings by the learning data having the attribute data of "-" is 0 in the cell of the undefined coordinate (13, 16),
It can be seen that the number of firings by the learning data having the attribute data of “+” was once.

FIG. 20 shows a scene in which the cluster of cells sequentially changes from "class 2" to "undefined". Further, FIG. 21A shows a state in which the cluster of 427 cells has changed from “class 2” to “undefined”. FIG. 21B shows a comparison table at this time, in which 427 cells are “undefined”, and the undefined cells have a firing number of 253 due to learning data having “+” attribute data. Times, the number of firings by the learning data having the attribute data of "-" is 1826 times.

FIG. 22 (a) is a diagram showing the SOM when the "class 2" is determined by obtaining the maximum "efficiency" state. The cells that have already been determined to be "class 1" are painted red, and the cells that have been determined to be "class 2" this time are painted blue. FIG. 22B shows a control table at this time. In this example, 427 cells are defined as "class 2". The cells included in “Class 2” have 253 firings due to the learning data having the attribute data of “+”,
The number of firings by the learning data having the attribute data of "-" is 1826. In this example, all cells on the SOM were defined in one of the clusters because the efficiency was maximized when 427 cells were defined in "class 2". Is not always maximal if is defined in either cluster.
In this case, undefined cells will remain. The undefined cells are defined in any of the clusters by executing the automatic classification algorithm again or by the cluster adjustment described later.

By the above processing, the SOM as shown in FIG. 22A, that is, the SOM that is clustered on the primary SOM by the automatic classification algorithm under the statistical support based on the firing information is created. It

Creation of Third-Order SOM by Cluster Adjustment and Display of Discriminant Characteristic Value Next, regarding the secondary SOM whose cluster is determined by the automatic classification algorithm, it is obtained by arbitrarily adjusting the cluster with reference to the “discriminant characteristic value”. Tertiary SO
M will be described. The classification adjustment unit 30 displays the “discrimination characteristic value”.

As one of the discrimination characteristic values used here, the above-mentioned "efficiency" is used.

In addition, "sensitivity (True Positive Rati
o) ”,“ Singularity (True Negative Ratio) ”,“ FP
R (Fault Positive Ratio) ”,“ FNR (Fault Neg
The four values of “ative ratio” can also be used as the discrimination characteristic value.

Here, the sensitivity, specificity, FPR and FNR are defined by the following equations. Sensitivity = (number of times a cell defined in "class 1" is fired by learning data having attribute data of "+") / (number of firings of all cells defined in "class 1") (2) Specificity = (number of times cells defined by “class 2” are fired by learning data having attribute data of “−”) / (number of firing times of all cells defined by “class 2”) (3 ) FPR = (number of times a cell defined in “class 1” is fired by learning data having attribute data “−”) / (number of firings of all cells defined in “class 1”) (4 ) FNR = (number of firings of cells defined by "class 2" by learning data having attribute data of "+") / (number of firings of all cells defined by "class 2") (5 )

Designate a cell on the created secondary SOM,
By giving an instruction to change the cluster using the input device 12, the cluster of the cell is changed, and a third-order SOM that is arbitrarily clustered is obtained. For example, the cluster defined by "class 1" (red) can be changed to "class 2" by clustering by executing the automatic classification algorithm. At this time, the above-mentioned efficiency, sensitivity, specificity, FPR and FNR are recalculated and displayed on the screen.

This state will be described with reference to FIGS. 22 and 23. FIG. 22A shows the secondary SOM created by the automatic classification algorithm as described above, and FIG.
Is a third-order S that is arbitrarily cluster adjusted based on the second-order SOM
OM. 22 (b) and 23 (b) are corresponding comparison tables, and FIG. 22 (c) and FIG. 23 (c) are discrimination characteristic values. By adjusting the cluster as shown in the figure, the values in the comparison table change in conjunction,
The discrimination characteristic value also changes.

The operator of the clustering device arbitrarily adjusts the cluster by referring to the mark indicating the ignition information displayed on the SOM, the discriminating characteristic value, and the distance display created by the distance display section. To go. In this way, the third-order SOM whose cluster is finally determined is completed.

Creation and Evaluation of Unknown Data Next, a procedure for creating unknown data and determining which cluster the unknown data belongs to by using the created secondary SOM or tertiary SOM will be described. First, two arbitrary blood test data are prepared. In the present embodiment, it is possible to determine whether these two blood test data belong to the same subject or different subjects. Then, “three-dimensional vector data” is created from the two blood test data by a method similar to the method shown in the example of creating learning data. This "3
"Dimensional vector data" is unknown data. Next, when the unknown data is read by the unknown data reading unit 32, the reference data is searched for a cell having the reference data closest to the unknown data.

FIG. 24 is a diagram for explaining the state at this time. As a result of the search, as shown in the figure, the coordinates (6, 9) are set as the cells having the reference data most similar to the unknown data.
Suppose that the cell of fired. At this time, it is checked which of the cell clusters at the coordinates (6, 9) is. In this case, the cell with coordinates (6, 9) is defined as "class 1", so it can be seen that the unknown data belongs to "class 1".

Since "class 1" is a cluster in which the wrong data is originally collected, it can be determined that this unknown data is likely to be wrong data.

From another point of view, the present invention may be embodied as a computer program for realizing the above-mentioned sample misplacement determination method of the present invention or a medium recording the computer program. In this case, the sample replacement method of the present invention can be realized by reading the computer program from the input device 12.

[0094]

As described above, according to the present invention, it is possible to provide an effective method for determining a sample error by considering all test items.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a clustering device that is an embodiment of the present invention.

FIG. 2 is a flowchart showing an example of a flow of processing executed by a clustering device according to an embodiment of the present invention.

FIG. 3 is a diagram illustrating a learning process by an SOM learning algorithm.

FIG. 4 is a diagram illustrating reference data assigned to cells on an SOM.

FIG. 5 is a diagram illustrating learning data used in the SOM learning algorithm.

FIG. 6 is a diagram illustrating an example of creating learning data (learning data for error determination) used in the clustering device that is an embodiment of the present invention.

FIG. 7 is a diagram showing an example of learning data used in a clustering device according to an embodiment of the present invention.

FIG. 8 is a diagram illustrating a comparison table.

FIG. 9 is a flowchart showing an example of the processing flow of an automatic classification algorithm.

FIG. 10 is a diagram showing an example of SOM display (unsorted SO
M).

FIG. 11 is a diagram showing an example of SOM (a state in which distance display is added to unclassified SOM).

FIG. 12 is a view showing a display example of SOM (a state in which all cells are provisionally set to “class 1” for automatic classification).

FIG. 13 is a view showing a display example of SOM (a state in which one cell is undefined from FIG. 12).

FIG. 14 is a view showing a display example of SOM (a state in which undefined cells are further increased from FIG. 13).

FIG. 15 is a diagram showing a display example of SOM (a state in which undefined cells are further increased from FIG. 14).

FIG. 16 is a diagram showing a display example of SOM (a state in which all cells are undefined cells).

FIG. 17 is a diagram showing a display example of SOM (state in which cells belonging to “class 1” are determined).

FIG. 18 is a diagram showing a display example of SOM (a state in which all remaining cells are provisionally set to “class 2”).

FIG. 19 is a view showing a display example of SOM (a state in which one cell is undefined from FIG. 18).

FIG. 20 is a diagram showing a display example of SOM (a state in which undefined cells are further increased from FIG. 19).

FIG. 21 is a view showing a display example of SOM (a state where all remaining cells are undefined cells).

FIG. 22 is a view showing a display example of SOM (state in which classification of all cells is determined by automatic classification).

FIG. 23 is a view showing a display example of SOM (a state in which cell classification is adjusted from FIG. 22).

FIG. 24 is a view showing a display example of SOM (state in which unknown data is fired in FIG. 23).

[Explanation of symbols]

1: SOM (self-organizing map) 2: cell 12: Input device 10: Clustering device 12: Input device 14: Output device 16: Learning data reading section 18: SOM creation section 20: Distance calculation unit 22: Distance display section 24: Classification section 26: Control display section 28: Discrimination characteristic value display section 30: Cluster adjustment unit 32: Reading unknown data 34: Evaluation determination unit 36: Learning data DB (database) 38: Map DB (database) 40: Unknown data DB (database)

Continued front page    (72) Inventor Kazuyuki Kanai             1-5-1, Wakihama Kaigan Dori, Chuo-ku, Kobe-shi             Inside Sysmex Corporation F-term (reference) 2G058 GD01 GD05 GD07

Claims (6)

[Claims] 1. Arrangement of unknown data in a map clustered into a cluster in which sample difference data is collected and a cluster in which normal data is collected, and it is determined whether or not the sample is mixed depending on which cluster the unknown data belongs to A method for determining the sample mix-up. 2. The method for determining a sample difference according to claim 1, wherein the map is SOM. 3. A method for creating a map to be used for determination of sample difference, wherein learned data is created by learning learning data in which normal data and sample difference data are mixed, and the learned map is created as normal data. A method of creating a map for clustering a cluster of data and a cluster of data of sample difference. 4. The map creating method according to claim 3, wherein the map is an SOM. 5. A method for creating learning data to be used for creating a map used for determination of sample difference, wherein sample difference data is created from samples of different subjects, and normal data is obtained from samples of the same subject. A learning data creation method that creates and mixes sample difference data and normal data. 6. The learning data creating method according to claim 5, wherein the map is an SOM.