JP2021047548A - Biological data processing apparatus, biological data processing method, and program - Google Patents

Abstract

To provide a technique which, when missing occurs in biological data, can complement a missing part with higher accuracy compared to complementing the missing part by a prescribed value, for example, an average value.SOLUTION: A biological data processing apparatus 30 comprises a neural network which uses a pair of data having some missing part and data having no missing parts and is learned by pix2pix so as to, in response to input of the data having some missing part, output data having the missing part complemented. Biological data having some missing part is inputted to the neural network, and the biological data having the missing part complemented is outputted to calculate a stress index or the like.SELECTED DRAWING: Figure 4

Description

The present invention relates to a biometric data processing apparatus, a biometric data processing method and a program.

In recent years, attention has been paid to the working environment, such as the realization of health-friendly working styles. Along with this, as an objective monitoring method for internal conditions such as health indicators, the use of biological data acquisition devices worn on living organisms, including wristband type devices, is being studied. For example, the heartbeat interval data calculated from the electrocardiographic data and the volume pulse wave data is said to reflect the state of the autonomic nerve, and an index indicating an internal state such as a stress index is calculated from the heartbeat interval data.

Patent Document 1 describes a machine learning device that predicts markers useful for improving output accuracy and disease prediction from clinical test data when performing machine learning by a neural network using a data set with many missing values. Has been described. The machine learning device shapes a neural network unit including an input layer including a plurality of input units and an output layer including one or more output units, and predetermined input data including a plurality of values, and inputs the predetermined input data to the input layer. It is provided with a preprocessing unit that outputs well-formed input data. In addition, the preprocessing unit generates information indicating whether or not the plurality of values are missing values, or alternative values used when the plurality of values are missing values, and whether or not the plurality of values are missing values. The information indicating the above or the alternative value is output as the formatted input data.

Patent Document 2 describes a recognition device capable of acquiring a target model and an appropriate internal representation, and learning / recognizing with incomplete sensor information. Missing value estimation / learning means are associated with each observation target, and missing value estimation / learning means corresponding to a plurality of observation targets are learned from sensor information obtained by observing the observation target from various viewpoints. The missing value estimation / learning means includes an hourglass model, and the missing value of incomplete sensor information is estimated and the model is acquired by the iteration method and error back propagation. After learning the missing value estimation / learning means, the sensor information of the observation target observed by the sensor is given to a plurality of missing value estimation / learning means, and the missing value estimation / learning means that minimizes the error by the minimum value error selection means. Is selected to recognize the observation target observed by the sensor.

JP-A-2018-08752 Japanese Unexamined Patent Publication No. 08-212184

When a device equipped with a sensor such as a wristband type device is attached to a living body and the biological data is acquired from the sensor, the biological data acquired by the sensor may be lost due to body movement, and the living body is in a continuous state. It becomes difficult to grasp.

An object of the present invention is to provide a technique capable of complementing a defective portion with higher accuracy than in the case of complementing the defective portion with a predetermined value, for example, an average value when a defect occurs in biological data.

The invention according to claim 1 has been learned to use a pair of data having a defective portion and data having no defective portion, and to output data complementing the defective portion when data having the defective portion is input. Biological data processing including a neural network, an input unit for inputting biological data having a defective portion into the neural network, and an output unit for outputting the output from the neural network as biological data in which the defective portion is complemented. It is a device.

The invention according to claim 2 is the biological data processing apparatus according to claim 1, wherein the neural network is a pix2pix neural network.

The invention according to claim 3 is the biometric data processing apparatus according to claim 1, wherein the input unit inputs to the neural network only when the ratio occupied by the defective portion is within a predetermined range. is there.

The invention according to claim 4 is the biometric data processing apparatus according to any one of claims 1 to 3, further comprising a calculation unit for calculating a stress index using biometric data in which the defective portion is complemented.

The invention according to claim 5 is a step in which a computer processor generates biometric data having the defective portion from biological data having no defective portion, and biological data having the defective portion and data having no defective portion. This is a biometric data processing method that executes a step of learning a neural network in order to output biometric data that complements the defective portion when biometric data having the defective portion is input using the pair of.

The invention according to claim 6 further executes a step of inputting biological data having a defective portion into the learned neural network and outputting biological data complementing the defective portion from the neural network portion. The biometric data processing method described.

The invention according to claim 7 is a step of generating biometric data having the defective portion from biometric data having no defective portion in a computer processor, and biological data having the defective portion and data having no defective portion. This is a program that executes a step of learning a neural network in order to output biometric data that complements the defective portion when biometric data having the defective portion is input using the pair of.

The invention according to claim 8 is the seventh aspect of claim 7, wherein the step of inputting the biological data having a defective portion into the learned neural network and outputting the biological data complementing the defective portion from the neural network is further executed. Program.

According to the inventions of claims 1, 5, 6, 7, and 8, when a defect occurs in biological data, the defect is deleted with higher accuracy than when the defect is supplemented with a predetermined value, for example, an average value. Can complement the department.

According to the second aspect of the present invention, the defective portion can be complemented with high accuracy by the neural network of pix2pix.

According to the invention of claim 3, the accuracy of complementing the defective portion can be further ensured.

According to the invention of claim 4, the stress index can be calculated with high accuracy.

It is the basic principle explanatory drawing (the 1) of embodiment. It is the basic principle explanatory drawing (the 2) of embodiment. It is the basic principle explanatory drawing (the 3) of embodiment. It is a block diagram of the structure of embodiment. It is a functional block diagram of an embodiment. It is a processing flowchart of an embodiment. It is the generation explanatory diagram of the learning data of embodiment. It is a graph figure (the 1) which shows the defective part complementary data of an Example. It is a graph figure (No. 2) which shows the defective part complementary data of an Example. It is a graph figure (No. 3) which shows the defective part complementary data of an Example. It is a graph figure (the 4) which shows the defective part complementary data of an Example. It is a graph figure (No. 5) which shows the defective part complementary data of an Example. It is a graph figure (No. 6) which shows the defective part complementary data of an Example. It is a partially enlarged view of FIG. 8F. It is a table figure which shows the comparison result of the model generation data of an Example, and the data complemented by the average value. It is a graph which shows the comparison result (count) of the model generation data and the data supplemented with the average value with respect to the data loss rate of an Example. It is a graph which shows the comparison result (the average value of the difference) of the model generation data and the data supplemented with the average value with respect to the data loss rate of an Example. It is a histogram figure of the test data of an Example.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

<Basic principle>
First, the basic principle of biometric data processing in this embodiment will be described.

When a device equipped with a sensor such as a wristband type device is attached to a living body and the biological data is acquired from the sensor, if the biological data acquired by the sensor is lost due to body movement, the continuous state of the living body is grasped. It becomes difficult to do. Specifically, for example, when heartbeat interval data that is considered to reflect the state of the autonomic nerve is acquired from biometric data and an index indicating an internal state such as a stress index is calculated from the heartbeat interval data, the biometric data is missing. When it occurs, the heart rate interval data becomes inaccurate data, and the index such as the stress index inevitably becomes inaccurate.

When a defect occurs in the biometric data, there are countermeasures such as supplementing the defective part with a fixed value or complementing with an average value, but the accuracy of the heartbeat interval data after complementation is not guaranteed, which is sufficient. It cannot be said that it is a coping method.

Therefore, in this embodiment, we focus on a form called pix2pix, which is an application of a hostile generation network (GAN: Generative Adversarial Network), which is a kind of neural network (NN) called a generation model in deep learning technology. A model is constructed by learning a pair of biometric data having a defective portion and biometric data having no defective portion using this fix2pix. Then, using the trained pix2pix model, the deficient portion of the data having the deficient portion is complemented and output as data without the deficient portion.

pix2pix is a kind of GAN, and is based on GAN as a basic structure. The GAN is composed of two NNs, a generator and a discriminator. The generator receives an input of noise (for example, a value between 0 and 1 in 100 dimensions) and outputs fake image data. When true data or false data is input, the discriminator outputs the probability that the input data is true. The generator learns to generate more true fake image data, and the discriminator learns to improve the accuracy of distinguishing between true and false. This allows the generator to generate data that is closer to true as the learning progresses.

Regarding GAN, for example,
"Generative Adversarial Nets", Ian J. Goodfellow et.al, D'epartement d'informatique et de recherche op'erationnelle Universit'e de Montr'eal Montr'eal, QC H3C 3J7
It is described in.

In pix2pix, one of the pair data (hereinafter, this is referred to as "original data") is input to the generator 12 instead of the noise in GAN.

FIG. 1 shows a conceptual configuration diagram of pix2pix. The pix2pix includes two NNs, a generator 12 and a discriminator 20, similar to the GAN.

The original data 10, which is one of the pair data, is input to the generator 12. For example, when creating a map from an aerial photograph, the aerial photograph and the map are a pair, and the aerial photograph is the original data 10. When the original data 10 is input, the generator 12 generates a fake 14 from the original data 10 and outputs it. For example, input an aerial photograph and output fake map data.

In the hostile generation network, true data and false data generated by the generator 12 are individually input to the discriminator 20, and the discriminator 20 determines the authenticity. Either the data-true data pair 16 or the fake data pair 18 generated by the original data-generator 12 is input, and the discriminator 20 determines the authenticity of these paired data. As a result, the generator 12 will generate fake data that is paired with the original data and is closer to the true data.

FIG. 2 shows an example of the data generated and output by the generator 12 in which the data is input to the generator 12 constructed by the pix2pix. When the aerial photograph 100 is input as input data, the generator 12 generates and outputs a map 102 paired with the aerial photograph 100. In this way, the generator 12 whose model is constructed by pix2pix can output similar data paired with the input data. Therefore, in the present embodiment, paying attention to such a technique, the generator 12 is learned by using pix2pix as a pair of biological data having a defective portion and biological data having no defective portion, and the learned generator 12 has a defective portion. When the biometric data having the defect portion is input, the biometric data having no defective portion is output so as to complement the defective portion of the biometric data having the defective portion.

FIG. 3 shows a conceptual configuration diagram of the present embodiment. Similar to FIG. 2, data is input to the generator 12, and data paired with the input data is output. The generator 12 is an NN learned as a pair of biological data having a defective portion and biological data having no defective portion. When the data 200 having the defective portion is input to the generator 12, the biological data 202 in which the defective portion is complemented is output. The biometric data 202 in which the defective portion is complemented is biometric data close to the true value, and is biometric data closer to the true value than in the case of simply complementing with a constant value, for example, an average value.

That is, the trained generator 12 can create a fake that cannot be distinguished from the real one by the discriminator 20, and the discriminator 20 at this time can create a fake created by the generator 12. It was learned so that it could be distinguished from the real thing. It can be said that the generator 12 creates biometric data close to biometric data without defects, which is indistinguishable from biometric data without defects.

The generator 12 and the discriminator 20 are specifically composed of a CNN (Convolutional Neural Network), particularly a convolution-Batch Norm-ReLu. For fix2pix, for example
"Image-to-Image Translation with Conditional Adversarial Networks", Phillip Isola et.al, Berkeley AI Research (BAIR) Laboratory, UC Berkeley
It is described in.

In the present embodiment, the defective portion of the biological data having the defective portion is complemented by the generator learned by the pix2pix, and the biological data having no defective portion is output. The biological data uses a sensor such as a wristband type device. It contains arbitrary data that can be obtained from the living body by the equipped device. The biometric data is specifically, but is not limited to, an electrocardiogram (ECG) or a volume pulse wave (BVP). The biological data also includes heartbeat interval data calculated from an electrocardiogram or the like.

<Structure>
Next, a specific configuration of the present embodiment will be described.

FIG. 4 is a block diagram of the biological data processing device 30 according to the present embodiment. The biological data processing device 30 is composed of a computer, and specifically includes a control unit 32, a biological data processing unit 34, a communication unit 36, an operation unit 38, a display unit 40, and a storage unit 42.

The control unit 32 is composed of a processor such as a CPU (Central Processing Unit), and controls each part of the device to realize various functions.

The biometric data processing unit 34 is composed of a processor such as a GPU (Graphics Processing Unit) or a dedicated circuit, and executes biometric data processing in response to a control command from the control unit 32.

The operation of the processors in the control unit 32 and the biometric data processing unit 34 may be performed not only by one processor but also by a plurality of processors existing at physically separated positions in cooperation with each other. ..

The communication unit 36 is a communication module that realizes a communication connection to a communication line such as a public line or a dedicated line.

The operation unit 38 is a user interface such as a keyboard and a mouse. The operation unit 38 is not limited to physical buttons and the like, and may be software buttons such as touch buttons displayed on the display unit 40.

The display unit 40 is a liquid crystal display, an organic EL display, or the like, and displays various data by a control command from the control unit 32. The various data include biological data having a defective portion, biological data in which the defective portion is complemented, and the like.

The storage unit 42 is composed of a hard disk, a flash memory, or the like. The storage unit 42 stores the biological data processing program 42a and the CNN library 42b that exerts a function as a CNN. Although not shown, the storage unit 42 stores information defining the NN, parameter information including the weighting coefficient of each layer in the learned CNN, and the like.

In such a configuration, the biometric data processing unit 34 functions as a CNN using a memory based on the CNN library 42b, definition data, and parameter information stored in the storage unit 42.

Further, the control unit 32 executes the biometric data processing based on the biometric data processing program 42a stored in the storage unit 42. The biometric data processing includes a process of creating learning data to be supplied to the CNN, a process of supplying the learning data to learn the CNN, and a process of supplying biometric data having a defect to the trained CNN and outputting the living body. Includes the process of retrieving data.

FIG. 5 shows the functions of the control unit 32, the biological data processing unit 34, and the storage unit 42. The control unit 32 reads the data stored in the storage unit 42 and supplies the data stored in the storage unit 42 to the biometric data processing unit 34 according to the biometric data processing program 42a stored in the storage unit 42, and the NN realized by the biometric data processing unit 34. To learn. The CNN realized by the biological data processing unit 34 has a network configuration corresponding to pix2pix, and specifically, as shown in FIG. 1, functions as a first CNN that functions as a generator 12 and a discriminator 20. It has a second CNN. The discriminator 20 learns to determine the authenticity of the original data and the false data generated by the generator 12, while the generator 12 learns to trick the discriminator 20. The control unit 32 generates and supplies a pair of biological data having a defective portion and biological data having no defective portion as learning data. The parameter information of the first CNN that functions as the generator 12 and the parameter information of the second CNN that functions as the discriminator 20 by learning are stored in the storage unit 42.

After the learning is completed, the control unit 32 supplies the biological data having the defective portion as input data to the first CNN. The first CNN functioning as the generator 12 complements the defective portion of the biological data having the defective portion and outputs the biological data having no defective portion. The control unit 32 acquires the biometric data output from the first CNN that does not have a defective portion. When the control unit 32 acquires the biological data in which the defective portion is complemented, the control unit 32 calculates a stress index based on the acquired biological data and displays it on the display unit 40.

As the CNN library 42b, any deep learning library can be used, for example, keras can be used. Keras is an open source NN library written in Python that can be run on TensorFlow. The NN layer, loss function, optimization, initialization, activation function, and normalization are all provided as separate, combinable modules for creating new models and support CNN.

Further, when the control unit 32 acquires the biometric data in which the defective portion is complemented, the control unit 32 calculates the stress index based on the acquired biometric data and displays it on the display unit 40, but displays the biometric data in which the defective portion is complemented on the display unit 40. Alternatively, the biometric data in which the defective portion is complemented may be supplied to another processor, and the stress index may be calculated by the other processor. That is, the processing for complementing the defective portion and the processing for calculating the stress index may be executed by different processors. The processor that executes the complement processing of the defective portion and the processor that calculates the stress index do not have to be incorporated in the same device, and may be connected by a communication network.

Further, the biometric data processing device 30 may be configured as a server-client system. In this case, for example, the terminal device on the client side collects biological data from the subject and transmits it to the server device. In the server device, the learned CNN is used to supplement the missing part of the biological data, and the stress index is calculated and transmitted to the terminal device. The terminal device can also receive the stress index and display it on the display unit.

<Processing>
FIG. 6 is a processing flowchart of the biological data processing device 30 of the present embodiment. The processes S101 to S107 are processes for creating learning data, the processes S108 and S109 are learning processes, and S110 is a complementary process using a trained model.

First, the control unit 32 collects biological data as learning data (S101). Biological data can be obtained by having a plurality of subjects wear a device equipped with a sensor such as a wristband type device. The collected biological data is stored in the storage unit 42.

Next, the control unit 32 extracts a non-defective portion without a defective portion from the collected biological data (S102). Specifically, the control unit 32 extracts a portion having a length exceeding a predetermined time without any loss from the collected whole biological data. The predetermined time can be set arbitrarily, but may be, for example, 30 seconds. Further, "no loss" means that the difference between the acquisition times of adjacent data is equal to or less than the threshold time. The threshold time can be set arbitrarily, but can be, for example, 1.5 seconds.

After extracting the non-defective portion from all the data, the control unit 32 resamples the non-defective portion and creates true value data by creating fixed-length data having a predetermined length (S103). Resampling is performed at a sampling frequency of, for example, 4 Hz, and the predetermined length can be, for example, 31.75 seconds (128 at 4 Hz).

After creating the true value data, next, biometric data having a defect is created based on the true value data.

That is, first, the number of missing parts is set (S104). The number of defects can be set arbitrarily, but for example, it is set to 2 or 3 places.

Next, the index of the missing portion is randomly set (S105). The index of the missing part indicates the starting position of the missing part. When there are two missing points, two randomly set points are set as missing points with respect to the true value data having a data length of 128.

Further, the length of the defective portion is randomly set for each index of the defective portion (S106). When setting at random, the upper limit value may be set and may be set randomly within the range of 0 to the upper limit value. The upper limit value can be set to, for example, 20 or the like. Since the index of the missing part is set randomly, the missing part may be continuous as a result, and in this case, the length of the missing part may exceed the upper limit value.

The number of missing parts, the index, and the length of the missing parts may be incorporated in the biometric data processing program 42a in advance, or may be input by the user by operating the operation unit 38.

After setting the number of missing points, the index, and the length of the missing points, data is lost for all true value data using this information, and the missing points are filled with a predetermined value, for example, 0. Create data (S107).

In this way, missing data is created from the true value data. The true value data is assumed to be x1, x2, x3, x4, ..., X127, x128.
Then, from this true value data, as missing data with two missing parts,
x1, x2,0,0, ..., x60, x61,0,0, ..., x127, x128
Etc. are created.

The control unit 32 gives the created true value data and missing data as a pair to the CNN as learning data. The CNN learns using the given learning data (S108), and the parameter information of the first CNN that functions as the generator 12 and the parameter information of the second CNN that functions as the discriminator 20 are stored in the storage unit 42. The model is created by being stored in (S109). When missing data is input as input data by this model, false data that can be determined as a true value by the discriminator 20, that is, true value data in which the missing portion is complemented is output.

When the learning of the CNN is completed, the control unit 32 then reads the biometric data collected from the evaluator to be stress-evaluated from the storage unit 42 and supplies it to the trained model. The first CNN functioning as the generator 12 in the trained model complements the defective portion of the input biological data having the defective portion and outputs the biological data (S110).

FIG. 7 schematically shows a process of creating learning data, that is, a process of creating missing data from true value data.

The upper part of FIG. 7 shows an example of true value data 300 as a time series waveform. The true value data 300 is obtained by extracting a non-defective portion from the collected data and resampling the non-defective portion to create fixed-length data having a predetermined length. Specifically, the true value data 300 is heartbeat interval data obtained by calculating the heartbeat interval from an electrocardiogram obtained by a wristband type device or the like. The middle part of FIG. 7 shows a case where the number of missing parts, the index, and the length of the missing parts are set and these information are applied to the true value data 300. The number of defective parts is set to 2, and the index and length at each part are set at random. The set locations are shown as broken lines 300a and 300b in the figure. The sizes of the broken line 300a and the broken line 300b are generally not the same because they are randomly set within the range of 0 to 20. The lower part of FIG. 7 shows the missing data 302 in which the true value data is lost at the missing points indicated by the broken lines 300a and 300b. The missing portion is actually filled with 0. The learning data is composed of the true value data 300 shown in the upper row and the missing data 302 shown in the lower row as a pair.

It was implemented and tested using Keras as the CNN library 42b.

First, as a simple test, sine waves of various frequencies were created, and data having a defect was created by omitting a part of the sine wave, and pair data with no defect and pair data with a defect were created. Then, using these pair data, the CNN of the biometric data processing unit 34 was learned with pix2pix, and then a sine wave with a defect was input, and it was confirmed that a sine wave with the defective portion complemented was output. ..

Next, heartbeat interval data (IBI) was collected as biometric data by having six subjects wear wristband type devices for about one month.

From all the data, a part having a length slightly over 30 seconds without any loss was extracted. No loss means that the difference in acquisition time of adjacent data is less than 1.5 seconds. A total of 10684 non-defective data were extracted.

Then, the non-defective data was resampled at 4 Hz, and fixed-length data having a length of 31.75 seconds (128 at 4 Hz) was created and used as true value data.

Next, for the true value data, the number of missing parts was set to three, the index of the missing part was randomly selected with respect to the data length of 128, and the length of each missing part was randomly selected from 0 to 20. Then, the missing part was filled with zero (0) to create the missing data.

The true value data and the missing data were paired, and 9000 training pair data and 1000 test data were randomly selected from a total of 10684 pair data. The CNN of the biological data processing unit 34 was trained using 9000 pair data for learning, and after the learning was completed, 1000 test data were given as input data and the output was obtained.

8A to 8F show input waveforms and output waveforms of different test data. In each figure, the horizontal axis is time, the vertical axis is the heartbeat interval time value in the heartbeat interval data, and a, b, and c are a: true value data b: missing data c: complementary data output from the model. .. Further, FIG. 8G shows a partially enlarged view of FIG. 8F. If the complementary data output from the model indicated by c and the true value data indicated by a match, it indicates that the missing portion has been complemented with high accuracy. As shown in FIGS. 8A to 8G, the complementary data output from the model is close to the true value data in many missing parts. However, there are some places where the complementary data output from the model does not match the true value data. For example, focusing on FIG. 8G, in the second defect portion, the complementary data (c) and the true value data (a) output from the model are in good agreement, but in the first defect portion, both are in agreement. Absent.

FIG. 9 shows the ratio of the number closer to the true value than the other, the ratio of the number closer to the true value than the other, and the true value, comparing the complementary data output from the model and the data obtained by complementing the missing part with the average value. The average value of the difference from and is shown in the table. It is the result of calculating the stress index (LF / HF) from the heartbeat interval data and comparing using the stress index. It should be noted that the stress evaluation (LF / HF) of the true value data is 20 or more, excluding the data that is not common sense, and is the result of evaluating 868 test data.

Stress evaluation (LF / HF) is known, and from heartbeat interval data, high-frequency fluctuation component (HF component) corresponding to respiratory fluctuation and low-frequency component (LF component) corresponding to Mayer wave, which is blood pressure fluctuation, are known. Is extracted and the sizes of the two are compared. The magnitudes of the HF component and the LF component can be calculated arbitrarily, for example, in the LF component region (0.05 Hz to 0.15 Hz) and the HF component region (0.15 Hz to 0.40 Hz) of the power spectrum. The total value (integral value) of the intensities can be used.

There are 494 complementary data output from the model that are closer to the true value than the data complemented by the mean value, and conversely, the data complemented by the average value is closer to the true value than the complementary data output from the model. There are 374 numbers. In addition, the ratio of the complement data output from the model closer to the true value than the data complemented by the mean value is 56.90%, and conversely, the data complemented by the mean value is better than the complement data output from the model. The ratio close to the true value is 43.10%. Further, the average value of the difference between the complementary data output from the model and the true value is 1.19, while the average value of the difference between the data complemented by the average value and the true value is 1.5.

Therefore, the complementary data output from the model is often closer to the true value than the case of simply complementing with the average value, and the difference from the true value is also small.

FIG. 10 graphically shows the number of complementary data (c) output from the model and data (d) simply complemented by the average value for each data loss rate, which is closer to the true value. In the figure, the number of counts on the vertical axis is the number of which is closer to the true value, and the larger the value is, the larger the number is closer to the true value. As shown in the figure, the complementary data (c) output from the model over the entire region where the data loss rate is 0.05 to 0.40 is more than the data (d) supplemented by the average value. A value close to the true value was obtained.

FIG. 11 graphically shows the average value of the differences from the true values of the complementary data (c) output from the model and the data (d) simply supplemented with the average value for each data loss rate. When the data loss rate is 0.30 or less, the complementary data (c) output from the model has a smaller average value of the difference from the true value than the data (d) supplemented with the average value, and becomes a more true value. A close value was obtained. On the other hand, when the data loss rate exceeds 0.30, the difference between the complementary data (c) output from the model and the data (d) simply complemented by the average value is not so much seen and is about the same. From this, the model of the present embodiment was particularly effective when the data loss rate was within a predetermined range, and specifically when the data loss rate was 30% or less.

FIG. 12 shows a histogram of the data loss rate of 1000 test data used in this embodiment. There is a peak at the data loss rate of about 20%, and the number of data loss rates exceeding 30% is small. Therefore, even if the application range of the model is limited to the case where the data loss rate is 30% or less, it can be applied to most test data to supplement the missing part.

In the embodiment, since the training data and the test data are created by the same method, the histogram of the training data shows the same tendency as the histogram of the test data shown in FIG. Therefore, as a result of learning using the learning data showing such a histogram, it can be considered that the model of the present embodiment is particularly effective when the data loss rate is 30% or less. Therefore, by increasing and learning the training data with a high data loss rate of 30% or more, it is possible to complement the test data with a high data loss rate of 30% or more, which is close to the true value. obtain.

12 generator, 20 discriminator, 30 biometric data processing device, 32 control unit, 34 biometric data processing unit, 36 communication unit, 38 operation unit, 40 display unit, 42 storage unit, 42a biometric data processing program, 42b NN library.

Claims (8)

A neural network learned to output data that complements the defective portion when data having the defective portion is input using a pair of data having a defective portion and data having no defective portion, and a neural network.
An input unit for inputting biological data having a defect into the neural network,
An output unit that outputs the output from the neural network as biological data in which the defect portion is complemented, and an output unit.
A biological data processing device comprising. The biological data processing apparatus according to claim 1, wherein the neural network is a pix2pix neural network. The biometric data processing apparatus according to any one of claims 1 and 2, wherein the input unit inputs to the neural network only when the ratio occupied by the defective portion is within a predetermined range. The biometric data processing apparatus according to any one of claims 1 to 3, further comprising a calculation unit for calculating a stress index using biometric data in which the defective portion is complemented. The computer processor
The step of generating the biological data having the defective portion from the biological data having no defective portion, and
Using a pair of biometric data having the defective portion and data not having the defective portion, when the biometric data having the defective portion is input, the step of learning the neural network to output the biological data complementing the defective portion, and
Biodata processing method to perform. The biometric data processing method according to claim 5, further performing a step of inputting biometric data having a defective portion into the learned neural network and outputting biometric data complementing the defective portion from the neural network. To the processor of the computer
The step of generating the biological data having the defective portion from the biological data having no defective portion, and
Using a pair of biometric data having the defective portion and data not having the defective portion, when the biometric data having the defective portion is input, the step of learning the neural network to output the biological data complementing the defective portion, and
A program that executes. The program according to claim 7, wherein the step of inputting the biological data having a defective portion into the learned neural network and outputting the biological data complementing the defective portion from the neural network is further executed.

Images