JP7421703B2 - Classification program, classification device and classification method - Google Patents

Description

The present invention relates to a classification program, a classification device, and a classification method.

Parkinson's disease is an incurable neurological disease that exhibits characteristic symptoms, and motor symptoms of Parkinson's disease include, for example, tremor, bradykinesia, and micrographia. Parkinson's disease is generally difficult to distinguish from other diseases, and it is known that diagnosis by a doctor requires time and money.

Therefore, in recent years, research has been conducted to diagnose Parkinson's disease from the analysis results of handwritten characters obtained using a pen tablet or the like. Specifically, in research using such handwritten characters, the presence or absence of tremors, etc. is analyzed based on information such as writing pressure and writing speed obtained from a pen tablet etc. (Patent Document 1) ).

Japanese Patent Application Publication No. 2010-131280

However, Parkinson's disease includes various other motor symptoms in addition to the motor symptoms described above. Therefore, the analysis using pen pressure, writing speed, etc. as described above may not be able to provide a highly accurate diagnosis of the specific progress of Parkinson's disease. Therefore, in the medical field, there is a need for a method that can diagnose Parkinson's disease with higher accuracy.

SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to provide a classification program, a classification device, and a classification method that make it possible to diagnose Parkinson's disease with higher accuracy.

The classification program according to the present invention for achieving the above object includes, for each of a plurality of subjects included in a first group and a second group, a first feature value indicating each of a plurality of characteristics of handwritten characters by each subject; is obtained, and for each of the plurality of subjects, performs DP (Dynamic Programming) matching between the first feature value corresponding to each subject and a second feature value indicating each of the plurality of features of the reference character. and based on the result of the DP matching, the first feature value is determined between the plurality of subjects included in the first group and the plurality of subjects included in the second group among the plurality of features. the first feature value indicating each of the one or more characteristics for each subject for each of the plurality of subjects included in the first group and the second group; and information indicating whether each subject is included in the first group or the second group, and learning by using the plurality of learning data. It is characterized by generating a model and causing a computer to execute processing.

Further, in one aspect of the classification program according to the present invention for achieving the above object, the plurality of features are based on the writing pressure and coordinates of the tip of the pen at each predetermined time when writing the handwritten character. , an altitude angle and an azimuth angle of the rear end of the pen.

Further, in one aspect, the classification program according to the present invention for achieving the above object performs a t-test on a value indicating the result of the DP matching, and performs a t-test on a value indicating the result of the DP matching, and selects among the plurality of features calculated by the t-test. The feature is that a feature whose p value is less than or equal to a predetermined value is specified.

Further, in one aspect of the classification program according to the present invention for achieving the above object, the plurality of subjects included in the first group are subjects who are not suffering from a predetermined disease, and the second group It is characterized in that the plurality of subjects included in the above are subjects who are suffering from the predetermined disease.

Further, in one aspect, the classification program according to the present invention for achieving the above object includes, for each of the plurality of features, a plurality of objects included in the first group among the first feature values indicating each feature. The present invention is characterized by causing a computer to execute a process of acquiring an average value of feature values corresponding to a person, and specifying a character corresponding to each of the average values as the reference character.

Further, in one aspect, the classification program according to the present invention for achieving the above object uses the first feature value for each of the plurality of subjects included in the first group and the second group. By calculating a third feature value indicating each of a plurality of other features that are different from the plurality of features, and using the plurality of third feature values, it is possible to determine which of the plurality of other features are included in the first group. identifying one or more other characteristics in which the tendency of the third feature value differs between the plurality of subjects included in the plurality of subjects included in the second group and the plurality of subjects included in the second group; For each of the plurality of subjects included, at least a part of the first feature value indicating each of the one or more characteristics for each subject, and the first characteristic value indicating each of the one or more characteristics for each subject. The present invention is characterized in that a plurality of learning data including at least some of the three characteristic values and information indicating which of the first group and the second group each subject is included in is generated.

Further, in one aspect of the classification program according to the present invention for achieving the above object, the plurality of other features include the predetermined time period of the altitude angle of the rear end of the pen when writing the handwritten character. a second rate of change of the first rate of change in the predetermined time period; a third rate of change of the azimuth angle of the rear end of the pen in the predetermined time period; and a fourth rate of change over a predetermined period of time.

Further, in one aspect of the classification program according to the present invention for achieving the above object, the plurality of other features are set at a first coordinate among the coordinates at each predetermined time when writing the handwritten character. and a straight line passing through the first coordinate and the second coordinate at the coordinates of three points where the distance between the second coordinate and the second coordinate and the distance between the second coordinate and the third coordinate are predetermined lengths. and a straight line passing through the second coordinate and the third coordinate.

Further, in one aspect, the classification program according to the present invention for achieving the above object receives new feature values indicating each of the plurality of features about a new subject, and calculates the new feature values for the learning model. The computer is characterized by causing a computer to execute a process of outputting information indicating which of the first group and the second group the new target person is included in as the information output in response to the input. do.

In addition, the classification program according to the present invention for achieving the above-mentioned object includes a classification program for each of the plurality of subjects included in the first group and the second group. By acquiring a first feature value and using the first feature value for each of a plurality of subjects included in the first group and the second group, each of a plurality of other features different from the plurality of features is determined. By calculating the third feature value shown and using the plurality of third feature values, among the plurality of other characteristics, a plurality of subjects included in the first group and a plurality of subjects included in the second group identify one or more other characteristics in which the tendency of the third feature value differs between the subjects and the subjects, and for each of the plurality of subjects included in the first group and the second group, A plurality of learning data including at least a portion of the first feature values indicating each of the one or more features and information indicating whether each subject is included in the first group or the second group. and generate a learning model by using the plurality of learning data, and the plurality of other features are the pen pressure and coordinates of the tip of the pen at each predetermined time when writing the handwritten character, The altitude angle and azimuth angle of the rear end of the pen, the distance between the first coordinate and the second coordinate among the coordinates at each predetermined time, and the distance between the second coordinate and the third coordinate. at least an angle formed by a straight line passing through the first coordinate and the second coordinate and a straight line passing through the second coordinate and the third coordinate in the coordinates of the three points where and have a predetermined length. It is characterized by containing one.

In addition, the classification device according to the present invention for achieving the above object includes, for each of the plurality of subjects included in the first group and the second group, a first a feature value acquisition unit that obtains a feature value; and a feature value acquisition unit that obtains, for each of the plurality of subjects, the first feature value corresponding to each subject and a second feature value indicating each of the plurality of features of the reference character. A matching unit that performs DP matching, and based on the result of the DP matching, among the plurality of characteristics, between a plurality of subjects included in the first group and a plurality of subjects included in the second group. a feature identifying unit that identifies one or more features having different trends in the first feature value; Data for generating a plurality of learning data including at least a part of the first feature values indicating each of the characteristics and information indicating which of the first group and the second group each subject is included in. The present invention is characterized in that it includes a generation section and a model generation section that generates a learning model by using the plurality of learning data.

In addition, the classification device according to the present invention for achieving the above-mentioned object is configured to classify, for each of the plurality of subjects included in the first group and the second group, a plurality of classification devices each indicating a plurality of characteristics of handwritten characters by each subject. A feature value acquisition unit that obtains a first feature value, and a plurality of other features different from the plurality of features by using the first feature value for each of the plurality of subjects included in the first group and the second group. a feature value calculation unit that calculates a third feature value indicating each of the features, and a feature value calculation unit that calculates a third feature value indicating each of the features of a feature identifying unit that identifies one or more other features in which the tendency of the third feature value differs between the subject and the plurality of subjects included in the second group; For each of the plurality of subjects included, at least a part of the first feature value indicating each of the one or more characteristics for each subject, and whether each subject belongs to the first group or the second group. a data generating section that generates a plurality of learning data including information indicating whether the learning data is included in the learning data, and a model generating section that generates a learning model by using the plurality of learning data, The characteristics include the writing pressure and coordinates of the tip of the pen at each predetermined time when writing the handwritten character, the altitude angle and azimuth of the rear end of the pen, and the coordinates at each predetermined time. The first coordinate and the second coordinate in the coordinates of three points where the distance between the first coordinate and the second coordinate and the distance between the second coordinate and the third coordinate are predetermined lengths. and an angle formed by a straight line passing through the second coordinate and the third coordinate.

Furthermore, the classification method according to the present invention for achieving the above object includes, for each of the plurality of subjects included in the first group and the second group, a first obtaining feature values, and performing DP matching for each of the plurality of subjects between the first feature value corresponding to each subject and a second feature value indicating each of the plurality of features of the reference character; Based on the result of the DP matching, among the plurality of features, a tendency of the first feature value between the plurality of subjects included in the first group and the plurality of subjects included in the second group is determined. at least one of the first feature values indicating each of the one or more characteristics for each subject for each of the plurality of subjects included in the first group and the second group; and information indicating whether each subject is included in the first group or the second group, and by using the plurality of learning data, a learning model is generated. It is characterized by causing a computer to generate and execute the processing.

Furthermore, the classification method according to the present invention to achieve the above object includes, for each of the plurality of subjects included in the first group and the second group, By acquiring a first feature value and using the first feature value for each of a plurality of subjects included in the first group and the second group, each of a plurality of other features different from the plurality of features is determined. By calculating the third feature value shown and using the plurality of third feature values, among the plurality of other characteristics, a plurality of subjects included in the first group and a plurality of subjects included in the second group identify one or more other characteristics in which the tendency of the third feature value differs between the subjects and the subjects, and for each of the plurality of subjects included in the first group and the second group, A plurality of learning data including at least a portion of the first feature values indicating each of the one or more features and information indicating whether each subject is included in the first group or the second group. and generate a learning model by using the plurality of learning data, and the plurality of other features are the pen pressure and coordinates of the tip of the pen at each predetermined time when writing the handwritten character, The altitude angle and azimuth angle of the rear end of the pen, the distance between the first coordinate and the second coordinate among the coordinates at each predetermined time, and the distance between the second coordinate and the third coordinate. at least an angle formed by a straight line passing through the first coordinate and the second coordinate and a straight line passing through the second coordinate and the third coordinate in the coordinates of the three points where and have a predetermined length. It is characterized by containing one.

According to the classification program, classification device, and classification method of the present invention, it becomes possible to diagnose Parkinson's disease with higher accuracy.

FIG. 1 is a diagram showing a configuration example of a classification device 1 according to an embodiment of the present invention. FIG. 2 is a diagram illustrating the classification process in the first embodiment. FIG. 3 is a flowchart illustrating details of the classification process in the first embodiment. FIG. 4 is a flowchart illustrating details of the classification process in the first embodiment. FIG. 5 is a specific example of handwritten characters written by the subject. FIG. 6 is a specific example of handwritten characters written by a subject. FIG. 7 is a diagram illustrating a specific example of first feature values indicating each feature acquired from handwritten characters. FIG. 8 is a diagram illustrating a specific example of first feature values indicating each feature acquired from handwritten characters. FIG. 9 is a diagram illustrating a specific example of first feature values indicating each feature acquired from handwritten characters. FIG. 10 is a diagram illustrating a specific example of first feature values indicating each feature acquired from handwritten characters. FIG. 11 is a diagram illustrating a specific example of the result of DP matching. FIG. 12 is a diagram explaining "Angle 1 mm". FIG. 13 is a diagram illustrating a specific example of the results of the t-test. FIG. 14 is a diagram illustrating a specific example of the results of the t-test. FIG. 15 is a diagram illustrating a specific example of the results of the t-test. FIG. 16 is a diagram illustrating the classification accuracy of the learning model generated in the process of S25. FIG. 17 is a diagram illustrating the classification accuracy of the learning model generated in the process of S25.

Embodiments of the present invention will be described below with reference to the drawings. However, these embodiments do not limit the technical scope of the present invention.

FIG. 1 is a diagram showing a configuration example of a classification device 1 according to an embodiment of the present invention. The classification device 1 is a computer device, and may be, for example, a general-purpose PC (Personal Computer). Furthermore, the classification device 1 may be of any type, such as a stationary type, a notebook type, or a tablet type.

The classification device 1 has a hardware configuration of a general-purpose computer device, and includes, for example, a CPU 101 as a processor, a memory 102, a network interface 103, and a storage medium 104, as shown in FIG. Each part is connected to each other via a bus 105.

The storage medium 104 has, for example, a program storage area (not shown) that stores a program (not shown) for performing processing for classifying Parkinson's disease patients and healthy individuals (hereinafter also referred to as classification processing). Furthermore, the storage medium 104 includes a storage area 110 that stores information used when performing classification processing, for example. Note that the storage medium 104 may be, for example, an HDD (Hard Disk Drive) or an SSD (Solid State Drive).

The CPU 101 executes a program loaded into the memory 102 from the storage medium 104 (storage area 110) to perform classification processing.

Further, the network interface 103 communicates with an operating terminal 5 operated by, for example, a doctor who diagnoses Parkinson's disease.

[Outline of the first embodiment]
Next, an outline of the first embodiment will be described. FIG. 2 is a diagram illustrating the classification process in the first embodiment.

First, as shown in FIG. 2, the feature value acquisition unit 11 of the classification device 1 collects data written by each subject for each of a plurality of subjects (hereinafter also simply referred to as a plurality of subjects) including Parkinson's disease patients and healthy subjects. A feature value (hereinafter also referred to as a first feature value) indicating each of a plurality of features of the handwritten character is acquired.

The matching unit 12 then calculates, for each of the plurality of subjects, the first feature value corresponding to each subject and a plurality of characters (hereinafter also referred to as reference characters) that serve as a reference for handwritten characters written by each subject. DP matching is performed with feature values (hereinafter also referred to as second feature values) representing each of the features.

Subsequently, the feature identifying unit 13 identifies one or more features among the plurality of features that have different tendencies in the first feature value between Parkinson's disease patients and healthy individuals, based on the results of the DP matching. Specifically, the feature specifying unit 13 uses, for example, a t-test to specify one or more features in which the tendency of the first feature value differs between Parkinson's disease patients and healthy subjects.

Thereafter, the data generation unit 14 generates, for each of the plurality of subjects, at least a portion of the first feature value indicating each of the one or more features (the one or more features identified by the feature identification unit 13) about each subject, and each A plurality of learning data including information indicating whether the subject is a Parkinson's disease patient (or not a healthy person) is generated.

The model generation unit 15 then generates a learning model by using the plurality of learning data. Specifically, the model generation unit 15 generates a learning model that can classify patients with Parkinson's disease and healthy individuals.

Note that the reference characters used when the matching unit 12 performs DP matching are acquired (specified) by the character specifying unit 16, for example, as shown in FIG.

That is, the classification device 1 in this embodiment uses DP matching to identify features that can be determined to be effective to include as part of the learning data. Then, the classification device 1 generates a learning model by using a plurality of learning data including feature values (first feature values) corresponding to the identified features.

This allows the classification device 1 to generate a learning model that can classify Parkinson's disease patients and healthy individuals with high accuracy. Therefore, by using the generated learning model, a doctor who diagnoses Parkinson's disease can diagnose Parkinson's disease with higher accuracy.

[Details of the first embodiment]
Next, details of the first embodiment will be described. 3 and 4 are flowcharts illustrating details of the classification process in the first embodiment. Further, FIGS. 5 to 17 are diagrams illustrating details of the classification process in the first embodiment.

As shown in FIG. 3, the feature value acquisition unit 11 obtains, for each of a plurality of subjects, first feature values indicating a plurality of features of handwritten characters written by each subject (S11). Hereinafter, specific examples of handwritten characters will be explained.

[Specific examples of handwritten characters]
5 and 6 are specific examples of handwritten characters written by the subject. Specifically, FIGS. 5 and 6 are specific examples showing the state on the tablet when the subject writes handwritten characters.

Figures 5(A) and 5(B) are handwritten characters drawn by tracing an Archimedean spiral (tracing from the inside to the outside), and Figures 5(C) and 5(D) are handwritten characters of a sine curve. These are handwritten characters made by tracing (tracing from the left side to the right side), and FIGS. 5(E) and 5(F) are handwritten characters made by imitation writing of triple circles (pattern writing of the triple circles on the left).

In addition, Figures 6(A) and 6(B) are handwritten characters drawn by tracing three lines, and Figures 6(C) and 6(D) are free riding of the katakana "ku" (on the left side). 6(E) and 6(F) are handwritten characters obtained by free riding the katakana "se".

In the examples shown in FIGS. 5 and 6, the handwritten characters in FIGS. 5(A), 5(C), 5(E), 6(A), 6(C), and 6(E) are written by a Parkinson's disease patient, and in FIGS. 5(B), 5(D), 5(F), 6(B), 6(D) and 6(F). Handwritten characters are written by a healthy person.

[Specific examples of characteristics of handwritten characters]
Next, a specific example of the first feature value of each feature acquired from each handwritten character described in FIGS. 5 and 6 will be described. FIGS. 7 to 10 are diagrams illustrating specific examples of first feature values of each feature acquired from handwritten characters. That is, FIGS. 7 to 10 are diagrams for explaining various information acquired via the tablet when the subject writes handwritten characters.

Hereinafter, the features acquired from handwritten characters include the writing pressure at the tip of the pen when writing handwritten characters (hereinafter also simply referred to as writing pressure), and the X coordinate of the tip of the pen on the tablet (hereinafter referred to as The Y coordinate of the tip of the pen on the tablet (hereinafter simply referred to as the Y coordinate), the azimuth angle of the rear end of the pen (hereinafter simply referred to as the azimuth angle), and the The explanation will be given assuming that the altitude angle of the rear end portion (hereinafter also simply referred to as altitude angle) is included.

Note that the azimuth here is, for example, as shown in FIG. This is the value indicated by the slope AZ. Further, the altitude angle here is, for example, a value indicating the inclination AL of the rear end portion of the pen PN with respect to the vertical direction of the tablet TB, as shown in FIG.

9(A), FIG. 9(C), FIG. 9(E), FIG. 10(A), and FIG. 10(C) are each based on the tracing of the Archimedean spiral explained in FIG. 5(A). It is a graph showing time-series data of first feature values corresponding to pen pressure, X coordinate, Y coordinate, azimuth angle, and altitude angle when writing handwritten characters. Moreover, each of FIG. 9(B), FIG. 9(D), FIG. 9(F), FIG. 10(B), and FIG. 10(D) is based on the tracing of the Archimedean spiral explained in FIG. 5(B). It is a graph showing time-series data of first feature values corresponding to pen pressure, X coordinate, Y coordinate, azimuth angle, and altitude angle when writing handwritten characters. Note that in each graph shown in FIGS. 9 and 10, the horizontal axis indicates time, and the vertical axis indicates pen pressure, etc.

Returning to FIG. 3, the character specifying unit 16 specifies, for each of the plurality of features, the average value of the feature values corresponding to a healthy person among the first feature values indicating each feature (S12).

Then, the character identifying unit 16 identifies characters corresponding to each of the average values obtained in the process of S12 as reference characters (S13). That is, the character identifying unit 16 identifies the reference character by averaging each handwritten character written by each able-bodied person.

Next, the matching unit 12 compares, for each of the plurality of subjects, the first feature value obtained in the process of S11 with the special second feature value indicating each of the plurality of features of the reference character identified in the process of S13. DP matching is performed (S14).

Specifically, the matching unit 12 performs DP matching for each of the pen pressure, the X coordinate, the Y coordinate, the azimuth angle, and the altitude angle acquired in the process of S11, for example. Further, the matching unit 12 performs DP matching for the position corresponding to the X coordinate and Y coordinate obtained in the process of S11, for example. Below, a description will be given of the processing when performing DP matching.

[Specific example of processing when performing DP matching]
For example, time series data of pen pressure obtained from handwritten characters T = t ₁ , t ₂ , ..., t _n , and time series data of pen pressure obtained from reference characters R = r ₁ , r ₂ , ..., r _m , the matching unit 12 first optimizes the deviation in each time series data (the number of elements included in each time series data) by using the working path W=w ₁ , w ₂ , ..., w _l . (absorption of the difference). Each w included in the working path W is data corresponding to t _i and r _j , and can be expressed as “w=(i, j)”.

In this case, the distance d(i, j) between t _i and r _j is defined by the following equation (1), and the result of DP matching is defined by the following equation (2).

Further, the cumulative minimum distance g(i, j) in each element is expressed by the following equation (3).

The DP matching result becomes lower as the time series data T and time series data R become more similar. A specific example of the DP matching results will be described below.

[Specific example of DP matching results]
FIG. 11 is a diagram illustrating a specific example of the result of DP matching. Specifically, FIG. 11 is a diagram illustrating a specific example of the result of DP matching between the time-series data T of pen pressure obtained from handwritten characters and the time-series data R of pen pressure obtained from the reference character.

In the example shown in FIG. 11A, character WD1a corresponds to time series data T, and character WD1b corresponds to time series data R. Further, the example shown in FIG. 11(A) is an example where the result of DP matching is "3.8".

Further, in the example shown in FIG. 11(B), the character WD2a corresponds to the time series data T, and the character WD2b corresponds to the time series data R. Further, the example shown in FIG. 11(B) is an example where the result of DP matching is "32.3".

That is, the size of the difference between the time series data T and the time series data R in the example shown in FIG. 11(A) (the colored part in FIG. 11(A)) is the same as the difference in the example shown in FIG. 11(B). It is smaller than the size of the portion (the colored portion in FIG. 11(B)). Therefore, the result of DP matching in the example of FIG. 11(A) is smaller than the result of DP matching in the example of FIG. 11(B).

Returning to FIG. 3, the feature specifying unit 13 performs a t-test on each value indicating the result of the process in S14 (S15).

Then, the feature identifying unit 13 identifies a feature whose p value calculated in the process of S15 is less than or equal to a predetermined value from among the multiple features of the handwritten character (S16).

Next, as shown in FIG. 4, the feature value acquisition unit 11 obtains feature values ( (hereinafter also referred to as the third feature value) is calculated (S21). A specific example of the process in S21 will be described below.

[Specific example of processing in S21]
The feature value acquisition unit 11 determines the tip of the pen on the tablet by using the position acquired from the handwritten character (the position corresponding to the X coordinate and Y coordinate acquired from the handwritten character) as a plurality of other features. "Writing Velocity" indicates the rate of change (moving speed) per unit time of the position of "Writing Velocity", "Writing Acceleration" indicates the rate of change per unit time of "Writing Velocity", and the change per unit time of "Writing Acceleration""WritingJerk" indicating the rate is calculated for each unit time to obtain time series data of "Writing Velocity", "Writing Acceleration" and "Writing Jerk", respectively.

Specifically, the feature value acquisition unit 11 calculates "Writing Velocity", "Writing Acceleration", and "Writing Jerk" for each unit time by following each of the following equations (4) to (6). Note that in equations (4) to (6), x _i indicates the X coordinate of the tip of the pen at time t _i , y _i indicates the Y coordinate of the tip of the pen at time t _i , and d _i represents the moving distance of the tip of the pen from time t _i to time t _i+1 , and v _i represents the moving speed of the tip of the pen from time t _i to time t _i+1 (i.e., “Writing "Velocity"), a _i indicates the rate of change of v _i from time t _i to time t _i+1 (that is, "Writing Acceleration"), and j _i indicates the rate of change of v i from time t _i to time t _{i+1 (} i.e., "Writing Acceleration"); It shows the rate of change of a _i between (ie, "Writing Jerk").

In addition, the feature value acquisition unit 11 uses the pen pressure acquired from the handwritten characters to obtain "Pressure 1st Derivative" indicating the first derivative value of the pen pressure, and a second derivative of the pen pressure, as a plurality of other features. "Pressure 2nd Derivative" indicating the differential value is calculated for each unit time to obtain time series data of "Pressure 1st Derivative" and "Pressure 2nd Derivative", respectively.

Specifically, the feature value acquisition unit 11 calculates "Pressure 1st Derivative" and "Pressure 2nd Derivative" by following each of the following equations (7) and (8). In equations (7) and (8), p _i indicates the writing pressure at the tip of the pen at time t _i , and p1 _i is the differential of p _i from time t _i to time t _i+1 . p2 _i indicates the differential value of p1 _i from time t _i to time t _i+1 (ie, “Pressure 2nd Derivative”).

In addition, the feature value acquisition unit 11 uses the azimuth acquired from the handwritten characters to obtain "Azimuth 1st Derivative" indicating the first derivative of the azimuth, and the second derivative of the azimuth, as a plurality of other features. "Azimuth 2nd Derivative" indicating a differential value is calculated for each unit time to obtain time series data of "Azimuth 1st Derivative" and "Azimuth 2nd Derivative", respectively.

Specifically, the feature value acquisition unit 11 calculates "Azimuth 1st Derivative" and "Azimuth 2nd Derivative" by following each of the following equations (9) and (10). In equations (9) and (10), az _i represents the azimuth at the tip of the pen at time t _i , and az1 _i represents the azimuth at the tip of the pen from time t _i to time t _i+1. indicates the differential value of the azimuth angle (i.e., “Azimuth 1st Derivative”), and az2 _i indicates the differential value of az1 _i from time t _i to time t _i+1 (i.e., “Azimuth 2nd Derivative”) .

In addition, the feature value acquisition unit 11 uses the altitude angle acquired from the handwritten characters as a plurality of other features to obtain "Altitude 1st Derivative" indicating the first derivative value of the altitude angle, and the second derivative of the altitude angle. "Altitude 2nd Derivative" indicating a differential value is calculated for each unit time to obtain time series data of "Altitude 1st Derivative" and "Altitude 2nd Derivative", respectively.

Specifically, the feature value acquisition unit 11 calculates "Altitude 1st Derivative" and "Altitude 2nd Derivative" by following each of the following equations (11) and (12). In equations (11) and (12), al _i represents the altitude angle at the tip of the pen at time t _i , and al1 _i represents the altitude angle at the tip of the pen from time t _i to time t _i+1. indicates the differential value of the altitude angle (i.e., "Altitude 1st Derivative"), and al2 _i indicates the differential value of al1 _i from time t _i to time t _i+1 (i.e., "Altitude 2nd Derivative") .

In addition, the feature value acquisition unit 11 uses the positions acquired from handwritten characters as the plurality of other features to obtain the first A position whose distance from the second position is "1 (mm)" and which was acquired before the second distance (hereinafter also referred to as the first position) and a position whose distance from the second position is "1 (mm)". mm)" that was acquired after the second distance (hereinafter also referred to as the third position), and a straight line passing through the first and second positions and a straight line between the second and second positions. "Angle 1 mm" indicating the angle formed by the straight line passing through the three positions is calculated for each unit time, and time series data of "Angle 1 mm" is obtained.

Specifically, as shown in FIG. 12, the feature value acquisition unit 11 selects the position _Pk from among the positions of the tip of the pen (the shaded part in the figure) when writing the handwritten character WD. The first position is specified, the position P _i is specified as the second position, and the position P _j is specified as the third position. Then, the feature value acquisition unit 11 specifies AN, which is the angle formed by the straight line passing through the positions P _k and P _i and the straight line passing through the positions P _i and P _j .

Further, the feature value acquisition unit 11 uses the positions acquired from the handwritten characters as the plurality of other features, so that for each of the second positions acquired every unit time, the distance from the second position is " 3 (mm)'' and a first position acquired before the second distance; and a position where the distance between the second position is ``3 (mm)'' and is less than the second distance. The third position acquired later is specified, and "Angle 3 mm" indicating the angle between the straight line passing through the first position and the second position and the straight line passing through the second position and the third position is calculated every unit time. Calculate and obtain time series data of "Angle 3mm".

Furthermore, the feature value acquisition unit 11 uses positions acquired from handwritten characters as a plurality of other features, so that for each of the second positions acquired every unit time, the distance from the second position is " 5 (mm)'' and a first position acquired before the second distance; and a position where the distance between the second position is ``5 (mm)'' and is less than the second distance. The third position acquired later is specified, and "Angle 5 mm" indicating the angle between the straight line passing through the first position and the second position and the straight line passing through the second position and the third position is calculated every unit time. Calculate and obtain time series data of "Angle 5mm".

Specifically, the feature value acquisition unit 11 calculates "Angle 1 mm", "Angle 3 mm", and "Angle 5 mm" according to the following equation (13).

Here, in equation (13), i, j, and k are calculated (specified) according to equation (14) below.

Further, the cumulative distance l is calculated according to the following equation (15).

That is, when straight lines included in handwritten characters are written, the values of "Angle 1 mm", "Angle 3 mm", and "Angle 5 mm" each approach "180" normally. Therefore, for example, if a value such as "Angle 1 mm" is far from "180", it can be determined that the subject who wrote the handwritten character may have Alzheimer's disease.

Therefore, as will be described later, the classification device 1 improves the classification of Alzheimer's disease patients and healthy individuals by calculating and including the above-mentioned "Angle 1mm," "Angle 3mm," and "Angle 5mm" in the learning data. It becomes possible to generate a learning model that can be performed accurately.

Returning to FIG. 4, the feature specifying unit 13 performs a t-test on the third feature value calculated in the process of S21 (S22).

Then, the feature identifying unit 13 identifies a feature whose p value calculated in the process of S15 is less than or equal to a predetermined value from among the plurality of other features (S23). Hereinafter, a specific example of the results of the t-test performed in the processes of S15 and S22 will be explained.

[Specific example of t-test results]
FIGS. 13 to 15 are diagrams illustrating specific examples of t-test results. Specifically, FIGS. 13 and 14 are diagrams for explaining a specific example of the t-test performed in the process of S22, and FIG. 15 is a diagram for explaining a specific example of the t-test performed in the process of S15. .

The table shown in Figure 13 etc. contains the "item number" that identifies each data included in the table, the "feature value" that indicates the target of the t-test, and the t-test for the handwritten characters tracing Archimedes' spiral. "Task 1" shows the p-value calculated by performing a t-test on the handwritten character tracing the sine curve, and "Task2" shows the p-value calculated by performing the t-test on the handwritten character tracing the sine curve, and the handwritten character imitating the triple circle. "Task 3" shows the p-value calculated by performing a t-test on three-line handwritten characters, "Task 4" shows the p-value calculated by performing a t-test on three-line handwritten characters, and "Task 5" shows the p-value calculated by performing a t-test on the free-riding handwritten character of "ku" and the p-value calculated by performing a t-test on the free-riding handwritten character of katakana "se". ``Task6'' indicating the p-value.

Note that the "feature value" includes "Writing Velocity Mean" and "Writing Acceleration Mean", which are the average values for each subject of "Writing Velocity", "Writing Acceleration", and "Writing Jerk". and “Writing Jerk Mean” is set. be done.

In addition, "feature value" includes "Writing Velocity Std", "Writing Acceleration Std" and "W riting Jerk Std” is set. be done.

Similarly, the "feature value" includes "Pressure 1st Derivative Mean", "Pressure 2nd Derivative Mean", "Pressure 1st Derivative Std", "Pressure 2nd Derivative Mean" e Std”, “Azimuth 1st Derivative Mean”, “Azimuth 2nd Derivative Mean” ”, “Azimuth 1st Derivative Std”, “Azimuth 2nd Derivative Std”, “Altitude 1st Derivative Mean”, “Altitude 2nd Derivative Mea” n”, “Altitude 1st Derivative Std”, “Altitude 2nd Derivative Std”, “Angle 1mm Mean”, "Angle 3mm Mean", "Angle 5mm Mean", "Angle 1mm Std", "Angle 3mm Std" and "Angle 5mm Std" are set.

In addition, "feature value" includes the results of matching for each of the X coordinate, Y coordinate, pen pressure, azimuth angle, altitude angle, and position (position corresponding to the X coordinate and Y coordinate) performed in the process of S14. "DP Matching (X coordinate)", "DP Matching (Y coordinate)", "DP Matching (Pressure)", "DP Matching (Azimuth)", "DP Matching (Al position)” and “DP Matching (XY "coordinate)" is set.

Specifically, in the table shown in FIG. 13, for the data whose "item number" is "1", "Writing Velocity Mean" is set as the "feature value", and "0.10560" is set as "Task1". "0.05703" is set as "Task2", "0.00092" is set as "Task3", "0.09379" is set as "Task4", and "0.00001" is set as "Task5". is set, and "0.00001" is set as "Task6".

Furthermore, in the table shown in FIG. 13, for the data whose "item number" is "2", "Writing Velocity Std" is set as the "feature value", "0.35079" is set as "Task1", "0.24773" is set as "Task2", "0.01310" is set as "Task3", "0.51000" is set as "Task4", and "0.00002" is set as "Task5". and "0.00150" is set as "Task6". Descriptions of other data included in FIGS. 13 to 15 will be omitted.

Then, in the processing of S16 and S23, the feature specifying unit 13 determines that, for example, all of the values set in "Task1", "Task2", "Task3", "Task4", "Task5", and "Task6" are set to predetermined values. Identify the features that correspond to the "feature values" of the data below.

Returning to FIG. 4, the data generation unit 14 generates, for each of the plurality of subjects, a first feature value indicating the feature identified in the processing of S16, a third feature value indicating the feature identified in the processing of S23, and a third feature value indicating the feature identified in the processing of S23. A plurality of learning data including information indicating whether the patient is a Parkinson's disease patient or a healthy person is generated (S24).

Specifically, the data generation unit 14 generates learning data for each type of handwritten character for each of a plurality of subjects, for example.

Then, the model generation unit 15 generates a learning model by using the plurality of learning data generated in the process of S24 (S25).

Specifically, the model generation unit 15 generates a learning model for each type of handwritten character, for example, by using learning data generated for each type of handwritten character.

After that, for example, when a new subject writes a handwritten character, the classification device 1 identifies the first feature value and the third feature value corresponding to the handwritten character. Then, the classification device 1 inputs the identified first feature value and third feature value into a learning model (the learning model generated in the process of S25), and indicates whether the new subject is a Parkinson's disease patient. Obtain an output value (eg, an output value indicating the probability of having Parkinson's disease).

Thereby, a doctor diagnosing Parkinson's disease can easily diagnose whether a new subject is a Parkinson's disease patient by referring to the output value from the learning model. The classification accuracy of the learning model generated in the process of S25 will be explained below.

[Classification accuracy of learning model]
FIGS. 16 and 17 are diagrams illustrating the classification accuracy of the learning model generated in the process of S25. FIG. 16 is a diagram illustrating the classification accuracy of the learning model generated in the process of S25. On the other hand, FIG. 17 shows learning data that does not include the first feature value indicating the feature identified in the process of S16 (the third feature value indicating the feature identified in the process of S23, and that each subject is a Parkinson's disease patient and a healthy FIG. 4 is a diagram illustrating the classification accuracy of a learning model generated using learning data (including information indicating which one is a person and a person).

Specifically, in the example shown in FIG. 16, the classification accuracy for each of "Task1", "Task2", "Task3", "Task4", "Task5", and "Task6" is "92.83 (%)", "97.17(%)", "99.50(%)", "99.75(%)", "95.92(%)" and "94.17(%)" This value is higher than each classification accuracy in the example shown in FIG.

In other words, FIGS. 16 and 17 show that the learning model generated by the learning data including the first feature value indicating the feature identified in the process of S16 is more The results show that the learning model can classify Parkinson's disease patients and healthy people with higher accuracy than a learning model generated using training data that does not contain this data.

This allows a doctor who diagnoses Parkinson's disease to diagnose Parkinson's disease with higher accuracy by using the generated learning model.

1: Classification device 5: Operation terminal 101: CPU
102: Memory 103: Network interface 104: Storage medium 105: Bus

Claims (14)

For each of the plurality of subjects included in the first group and the second group, obtain a first feature value indicating each of the plurality of features of handwritten characters by each subject,
For each of the plurality of subjects, performing DP (Dynamic Programming) matching between the first feature value corresponding to each subject and a second feature value indicating each of the plurality of features of the reference character,
Based on the result of the DP matching, among the plurality of features, a tendency of the first feature value between the plurality of subjects included in the first group and the plurality of subjects included in the second group is determined. identify one or more characteristics that differ,
For each of the plurality of subjects included in the first group and the second group, at least a part of the first feature value indicating each of the one or more characteristics of each subject, and Generating a plurality of learning data including information indicating which of the first group and the second group is included,
generating a learning model by using the plurality of learning data;
A classification program that causes a computer to perform processing. In claim 1,
The plurality of characteristics include at least one of the writing pressure and coordinates of the tip of the pen at predetermined time intervals when writing the handwritten character, and the altitude angle and azimuth of the rear end of the pen. ,
A classification program characterized by: In claim 1,
In the process of identifying the one or more characteristics,
Perform a t-test on the value indicating the result of the DP matching,
out of the plurality of features, specifying a feature whose p value calculated by the t-test is less than or equal to a predetermined value;
A classification program characterized by: In claim 1,
The plurality of subjects included in the first group are subjects who are not affected by a predetermined disease,
The plurality of subjects included in the second group are subjects who are suffering from the predetermined disease.
A classification program characterized by: In claim 4, further:
For each of the plurality of features, among the first feature values indicating each feature, obtain an average value of feature values corresponding to the plurality of subjects included in the first group,
identifying characters corresponding to each of the average values as the reference characters;
A classification program that causes a computer to perform processing. In claim 2, further:
For each of the plurality of subjects included in the first group and the second group, by using the first feature value, a third feature value indicating each of a plurality of other features different from the plurality of features is calculated. death,
By using the third characteristic value, among the plurality of other characteristics, the third characteristic is determined between the plurality of subjects included in the first group and the plurality of subjects included in the second group. identifying one or more other features whose values have different trends;
Let the computer carry out the process,
In the process of generating the plurality of learning data,
For each of the plurality of subjects included in the first group and the second group, at least a part of the first feature value indicating each of the one or more characteristics for each subject; A plurality of learning data including at least a part of the third feature value indicating each of one or more features and information indicating which of the first group and the second group each subject is included in. generate,
A classification program characterized by: In claim 6,
The plurality of other features include a first rate of change of the altitude angle of the rear end of the pen during the predetermined time period when writing the handwritten character, and a second rate of change of the first rate of change during the predetermined time period. and at least one of a third rate of change of the azimuth angle of the rear end of the pen in the predetermined time period, and a fourth rate of change of the third rate of change in the predetermined time period.
A classification program characterized by: In claim 6,
The plurality of other features include a distance between a first coordinate and a second coordinate among the coordinates at each predetermined time when writing the handwritten character, and a distance between the second coordinate and the third coordinate. includes an angle between a straight line passing through the first coordinate and the second coordinate and a straight line passing through the second coordinate and the third coordinate at the coordinates of three points where the distance is a predetermined length;
A classification program characterized by: In claim 1, further:
accepting new feature values indicating each of the plurality of features about the new subject;
outputting information indicating which of the first group and the second group the new target person is included in as information output in response to input of the new feature value to the learning model;
A classification program that causes a computer to perform processing. For each of the plurality of subjects included in the first group and the second group, obtain a plurality of first feature values indicating each of the plurality of features of handwritten characters by each subject,
For each of the plurality of subjects included in the first group and the second group, by using the first feature value, a third feature value indicating each of a plurality of other features different from the plurality of features is calculated. ,
By using the third characteristic value, among the plurality of other characteristics, the third characteristic is determined between the plurality of subjects included in the first group and the plurality of subjects included in the second group. identify one or more other features whose values have different trends;
For each of the plurality of subjects included in the first group and the second group, at least a portion of the first feature value indicating each of the one or more other characteristics for each subject, and each subject Generating a plurality of learning data including information indicating which of the first group and the second group the learning data is included in;
generating a learning model by using the plurality of learning data;
Let the computer carry out the process,
The plurality of other characteristics include the writing pressure and coordinates of the tip of the pen at each predetermined time when writing the handwritten character, the altitude angle and azimuth of the rear end of the pen, and the coordinates of the tip of the pen at each predetermined time. Among the coordinates, the first coordinate and the coordinate of three points have a distance between the first coordinate and the second coordinate and a distance between the second coordinate and the third coordinate of a predetermined length. at least one of an angle formed by a straight line passing through the second coordinate and a straight line passing through the second coordinate and the third coordinate;
A classification program characterized by: a feature value acquisition unit that acquires, for each of the plurality of subjects included in the first group and the second group, a first feature value indicating each of the plurality of characteristics of handwritten characters by each subject;
a matching unit that performs DP matching for each of the plurality of subjects, between the first feature value corresponding to each subject and a second feature value indicating each of the plurality of features of the reference character;
Based on the result of the DP matching, among the plurality of features, a tendency of the first feature value between the plurality of subjects included in the first group and the plurality of subjects included in the second group is determined. a feature identification unit that identifies one or more features that differ in
For each of the plurality of subjects included in the first group and the second group, at least a part of the first feature value indicating each of the one or more characteristics of each subject, and a data generation unit that generates a plurality of learning data including information indicating which of the first group and the second group the learning data is included in;
a model generation unit that generates a learning model by using the plurality of learning data;
A classification device characterized by: a feature value acquisition unit that acquires, for each of the plurality of subjects included in the first group and the second group, a plurality of first feature values indicating each of the plurality of characteristics of handwritten characters by each subject;
For each of the plurality of subjects included in the first group and the second group, by using the first feature value, a third feature value indicating each of a plurality of other features different from the plurality of features is calculated. a feature value calculation unit;
By using the third characteristic value, among the plurality of other characteristics, the third characteristic is determined between the plurality of subjects included in the first group and the plurality of subjects included in the second group. a feature identification unit that identifies one or more other features with different value trends;
For each of the plurality of subjects included in the first group and the second group, at least a portion of the first feature value indicating each of the one or more other characteristics for each subject, and each subject a data generation unit that generates a plurality of learning data including information indicating which of the first group and the second group the learning data is included in;
a model generation unit that generates a learning model by using the plurality of learning data,
The plurality of other characteristics include the writing pressure and coordinates of the tip of the pen at each predetermined time when writing the handwritten character, the altitude angle and azimuth of the rear end of the pen, and the coordinates of the tip of the pen at each predetermined time. Among the coordinates, the first coordinate and the coordinate of three points have a distance between the first coordinate and the second coordinate and a distance between the second coordinate and the third coordinate of a predetermined length. at least one of an angle formed by a straight line passing through the second coordinate and a straight line passing through the second coordinate and the third coordinate;
A classification device characterized by: For each of the plurality of subjects included in the first group and the second group, obtain a first feature value indicating each of the plurality of features of handwritten characters by each subject,
Performing DP matching for each of the plurality of subjects, the first feature value corresponding to each subject and a second feature value indicating each of the plurality of features of the reference character,
Based on the result of the DP matching, among the plurality of features, a tendency of the first feature value between the plurality of subjects included in the first group and the plurality of subjects included in the second group is determined. identify one or more characteristics that differ,
For each of the plurality of subjects included in the first group and the second group, at least a part of the first feature value indicating each of the one or more characteristics of each subject, and Generating a plurality of learning data including information indicating which of the first group and the second group is included,
generating a learning model by using the plurality of learning data;
A classification method characterized by having a computer perform the processing. For each of the plurality of subjects included in the first group and the second group, obtain a plurality of first feature values indicating each of the plurality of features of handwritten characters by each subject,
For each of the plurality of subjects included in the first group and the second group, by using the first feature value, a third feature value indicating each of a plurality of other features different from the plurality of features is calculated. ,
By using the third characteristic value, among the plurality of other characteristics, the third characteristic is determined between the plurality of subjects included in the first group and the plurality of subjects included in the second group. identify one or more other features whose values have different trends;
For each of the plurality of subjects included in the first group and the second group, at least a portion of the first feature value indicating each of the one or more other characteristics for each subject, and each subject Generating a plurality of learning data including information indicating which of the first group and the second group the learning data is included in;
generating a learning model by using the plurality of learning data;
Let the computer carry out the process,
The plurality of other characteristics include the writing pressure and coordinates of the tip of the pen at each predetermined time when writing the handwritten character, the altitude angle and azimuth of the rear end of the pen, and the coordinates of the tip of the pen at each predetermined time. Among the coordinates, the first coordinate and the coordinate of three points have a distance between the first coordinate and the second coordinate and a distance between the second coordinate and the third coordinate of a predetermined length. at least one of an angle formed by a straight line passing through the second coordinate and a straight line passing through the second coordinate and the third coordinate;
A classification method characterized by

Images