JP7557773B2 - HEART RATE VARIABILITY INDEX ESTIMATION DEVICE, HEART RATE VARIABILITY INDEX ESTIMATION METHOD, AND HEART RATE VARIABILITY INDEX ESTIMATION PROGRAM - Google Patents

Description

The present invention relates to a heart rate variability index estimation device, a heart rate variability index estimation method, and a heart rate variability index estimation program that estimate heart rate variability based on the heart rate measured by a wearable device.

The heart regularly repeats contraction and expansion according to the rhythm of the electrical signal. When the electrical signal suddenly contracts to pump blood out of the heart, it generates an R wave, which shows the sharpest peak. The interval at which the heart beats (beat interval) is generally measured using the R wave. The beat interval (RR interval) is longer when at rest and shorter when exercising, but it also fluctuates periodically while at rest. The periodic fluctuation of the beat interval (heart rate fluctuation) includes high frequency components (HF: high frequency) and low frequency components (LF: low frequency). For example, HF is relatively high compared to LF when in a relaxed state, and LF is relatively high compared to HF when in a tense state. Therefore, psychological stress can be estimated based on HF or LF/HF calculated from heart rate fluctuation (see Non-Patent Document 1 and Non-Patent Document 2).

For example, Patent Document 1 discloses a technique for interpolating beat interval data acquired by a sensor and calculating HF or LF/HF based on the interpolated data when the heart rate cannot be measured from a sensor signal of the sensor that measures the heart rate within an interval time determined based on past heart rate data.

JP 2019-103614 A

"The Science of Stress and the Autonomic Nerve" (http://hclab.sakura.ne.jp/stress_novice_hartrate.html) "The Science of Stress and the Autonomic Nerve" (http://hclab.sakura.ne.jp/stress_novice_LFHF.html)

Many wearable devices, such as smartwatches, are equipped with a heart rate measurement function. The heart rate measurement function of wearable devices is generally used for health management, such as evaluating training intensity and monitoring sleep-wake rhythms. Recently, it has also come to be used for relaxation and stress evaluation, emotion estimation, and sleep-wake determination.

Figure 10 compares data measured synchronously using an electrocardiograph (DL-5000, S&ME, electrocardiogram heart rate sensor DL-310) and a heart rate monitor on a wristwatch-type wearable device (A370, Polar). The vertical axis indicates the beat interval (ms), and the horizontal axis indicates time (sec). The wearable device measures pulse waves (changes in blood pressure, pulse rate), but does not directly measure heart rate. In contrast, an electrocardiograph measures the electrical signals generated by the heart, which regularly repeats contraction and expansion, and directly measures heart rate. For example, a wearable device measures pulse waves at a cycle of 1 to 1.5 Hz, while an electrocardiogram measures heart rate at a cycle of 1 kHz. Therefore, wearable devices have lower accuracy in measuring heart rate than electrocardiographs, and there is a lot of data loss regarding the actual movement of the heart. Also, as shown in P11 and P12 in FIG. 10, the heart rate time series data WHR from the wearable device sometimes had a time lag of several tens of seconds compared to the beat interval time series data ECG based on the electrocardiogram.

A technology for interpolating missing data is disclosed in Patent Document 1. However, when missing data in the heart rate time series data of a wearable device is interpolated, the missing data is smoothed. As a result, for example, in the portions shown in P11 and P12 of FIG. 10, there is a large error between the interpolated heart rate time series data and the beat interval time series data ECG.

Figure 11 is a graph showing the correlation between HF calculated from the beat interval time series data ECG of Figure 10 and HF calculated from the heart rate time series data WHR of Figure 10. Figure 12 is a graph showing the correlation between LF calculated from the beat interval time series data ECG of Figure 10 and LF calculated from the heart rate time series data WHR of Figure 10. The horizontal axes of Figures 11 and 12 indicate values based on the beat interval time series data ECG (HF_ECG, LF_ECG), and the vertical axes indicate values based on the heart rate time series data WHR from the wearable device (HF_WHR, LF_WHR). The LF and HF values (LF_WHR, HF_WHR) obtained from the heart rate time series data of the heart rate monitor of the wearable device were smaller than the LF and HF values (LF_ECG, HF_ECG) obtained from the beat interval time series data ECG of the electrocardiogram, and the contribution rate (r*r=0.352) was also low.

The present invention has been made to solve the above problems, and aims to provide a technology for constructing a model that can estimate a heart rate variability index with high accuracy from a heart rate measured by a wearable device.

One aspect of the present invention is a method for measuring the heart rate of a subject by using a wearable device that is attached to a part of the subject's body and has a heart rate measurement function for measuring the heart rate of the subject. A first acquisition unit acquires, for each of a plurality of wearable devices that are attached to a part of the subject's body and have a heart rate measurement function for measuring the heart rate of the subject, heart rate time series data indicating the heart rate of the subject measured by the wearable device in a time series manner. A second acquisition unit acquires, for each of a plurality of electrocardiographs that are attached to the subject and measure the electrical signals generated by the subject's heart, electrocardiogram data indicating the heart rate fluctuation of the subject measured by the electrocardiograph. An explanatory variable generation unit converts the heart rate time series data acquired by the first acquisition unit into beat interval time series data, calculates a plurality of feature amounts from the beat interval time series data, and sets the calculated feature amounts as explanatory variables. An explanatory variable combination unit combines the plurality of explanatory variables generated by the explanatory variable generation unit. The system is characterized by having a heart rate variability index actual value calculation unit that calculates an actual value of a heart rate variability index from electrocardiogram data, a learning unit that selects a plurality of regression models using the actual value of the heart rate variability index as a response variable, and for each selected regression model, applies a regression model algorithm to the combination data of explanatory variables combined by the explanatory variable combination unit to learn the regression models, an evaluation index calculation unit that calculates an evaluation index for each selected regression model based on the actual value of the heart rate variability index calculated by the response variable generation unit and an estimated value of the heart rate variability index calculated from the trained model trained by the learning unit, and an optimal model selection unit that selects, as an optimal model, a model consisting of a combination of a regression model and explanatory variables with the best evaluation index calculated by the evaluation index calculation unit.

The heart rate variability index estimation device configured as above converts the heart rate time series data of the wearable device into beat interval time series data, calculates multiple feature amounts from the beat interval time series data to set them as explanatory variables, and applies a regression model algorithm to the data combining the explanatory variables to learn. Then, the model with the best evaluation index calculated based on the actual value of the heart rate variability index calculated from the electrocardiogram data and the estimated value of the heart rate variability index calculated from the learned model is selected as the optimal model, so that the estimated value of the heart rate variability index calculated by the selected optimal model approximates the actual value of the heart rate variability index calculated from the electrocardiogram data. Therefore, the heart rate variability index estimation device configured as above can construct a model that can estimate the heart rate variability index with high accuracy from the heart rate measured by the wearable device.

The control method, system, and program realized by the above device, as well as the computer-readable storage medium storing the program, are also novel and useful.

The invention having the above configuration realizes technology for constructing a model that can estimate a heart rate variability index with high accuracy from a heart rate measured by a wearable device.

1 is a schematic diagram illustrating the configuration of a heart rate variability index estimation device according to an embodiment of the present invention. FIG. 2 is a block diagram of functions executed by a heart rate variability index estimation device. 13 is a flowchart illustrating an example of a control procedure for a model construction process. 1 is a graph showing the average value of the coefficient of determination when HF was maximum in each regression model in Table 5. The average value of the coefficient of determination at which LF was maximized in each regression model in Table 5 is shown. 13 is a graph showing heart rate time series data measured simultaneously using a first and second wearable device. 13 is a graph showing the results of HF estimation. FIG. 1 is a diagram illustrating a heartbeat. FIG. 1 is a diagram illustrating heart rate variability. 11 is a graph comparing a time series of heart rate variability intervals measured by a wearable device with a time series of heart rate variability intervals (RR intervals) based on an electrocardiogram. 11 is a graph showing the correlation between HF calculated from electrocardiogram beat interval time series data and HF calculated from heart rate time series data of a wearable device. 11 is a graph showing the correlation between LF calculated from electrocardiogram beat interval time series data and LF calculated from heart rate time series data of a wearable device.

Below, one embodiment of the present invention will be described with reference to the drawings. This embodiment discloses a device that estimates a heart rate variability index from a heart rate measured by a wearable device.

<About heart rate variability index>
First, we will explain heart rate variability indexes. The heart has four chambers, the right ventricle, the right atrium, the left ventricle, and the left atrium, and four valves, the tricuspid valve, the pulmonary valve, the mitral valve, and the aortic valve, and it regularly repeats contraction and expansion according to the rhythm of electrical signals. The electrical signals are caused by changes in the distribution of electricity that occur within the heart when it contracts, and are detected as electrical potentials from the body surface. Electrocardiogram data measured by an electrocardiograph is a record of such electrical activity of the heart.

For example, as shown in FIG. 8(a), the peaks and valleys of the electrical signal measured by an electrocardiograph are labeled with names such as P, Q, R, and T. The sharpest peak, "R," is the electrical signal that occurs when the ventricles of the heart suddenly contract and pump blood out of the heart. In other words, it is the signal that occurs when the heart "thumps." Beat intervals are generally measured using R waves. When the heart is beating slowly, such as when at rest, the beat interval T is long, as shown in FIG. 8(b). On the other hand, when the heart is beating quickly, the beat interval T is short, as shown in FIG. 8(c) (see, for example, Non-Patent Document 1).

The beat interval T is constantly fluctuating. For example, as shown in FIG. 9, the beat interval fluctuates periodically even when at rest. In this specification, this periodic fluctuation of the beat interval T is referred to as "heart rate fluctuation." Heart rate fluctuation includes high frequency components (HF) and low frequency components (LF). HF indicates the total amount of the power spectrum in the high frequency band. LF indicates the total amount of the power spectrum in the low frequency band.

The magnitude of the HF and LF fluctuation waves in heart rate fluctuations varies depending on the balance of tension between the sympathetic and parasympathetic nerves. In a relaxed state, that is, when the parasympathetic nerves are activated, LF appears and HF decreases. In contrast, in a tense state, that is, when the sympathetic nerves are activated, HF appears and LF decreases. Thus, in a relaxed state, HF becomes larger relative to LF, and LF/HF becomes larger, but in a tense state, LF becomes larger relative to HF, and LF/HF becomes smaller. Due to these characteristics, HF and LF/HF calculated from the time series of heart rate fluctuations are used as indices that can be used to examine psychological states, such as stress indicators (see, for example, Non-Patent Document 2).

In this specification, the heart rate variability index is defined as a value calculated from fluctuations in the beat interval (R-R interval). HF, LF, and LF/HF are examples of heart rate variability indexes.

<Configuration of the heart rate variability index estimation device>
1 is an electrical block diagram of a heart rate variability index estimation device 10. The heart rate variability index estimation device 10 is, for example, an information processing device such as a personal computer, and has a controller 11 equipped with a CPU 21, a ROM 22, a RAM 23, and an NVRAM 24. A communication unit 12, an operation unit 13, and a display unit 14 are connected to the controller 11.

The communication unit 12 has hardware for communicating with external devices such as a wearable terminal 50 and an electrocardiograph 70. The heart rate variability index estimation device 10 is communicatively connected via the communication unit 12 to, for example, a plurality of wearable terminals 50 (50A, 50B, 50C...) and a plurality of electrocardiographs 70 (70A, 70B, 70C...) worn by a subject X (X1, X2, X3...). The communication method of the communication unit 12 may be wireless or wired.

The wearable device 50 is a portable device that is worn on a part of the body and has a communication function. The wearable device 50 also has a heart rate measurement unit 51 that measures the heart rate based on the pulse wave. The heart rate measurement unit 51 is, for example, an optical heart rate sensor (PPG sensor). The wearable device 50 transmits heart rate time series data indicating the heart rate measured by the heart rate measurement unit 51 in a time series to the heart rate variability index estimation device 10.

The electrocardiograph 70 is a device that measures electrical signals generated by the heart and records an electrocardiogram. Thus, the wearable device 50 and the electrocardiograph 70 have different methods of measuring heartbeats. The electrocardiograph 70 has a wireless communication function and transmits electrocardiogram data indicating an electrocardiogram to the heartbeat variability index estimation device 10. The heartbeat variability index estimation device 10 is communicatively connected to a plurality of electrocardiographs (70A, 70B, 70C...). The electrocardiographs 70 (70A, 70B, 70C...) are each worn by a subject X (X1, X2, X3...) wearing a wearable device 50 (50A, 50B, 50C...), and measure the heartbeat in synchronization with the wearable device 50 (50A, 50B, 50C...).

The heart rate time series data and the electrocardiogram data may be transmitted from the wearable terminal 50 and the electrocardiograph 70 periodically, at a time specified by the heart rate variability index estimation device 10 or at regular intervals, or may be transmitted from the wearable terminal 50 and the electrocardiograph 70 in response to a request from the heart rate variability index estimation device 10. The heart rate time series data and the electrocardiogram data are accompanied by user information identifying the subject, terminal information regarding the wearable terminal 50, electrocardiograph information indicating the electrocardiograph 70, etc., so that the heart rate variability index estimation device 10 can identify the subject and the wearable terminal 50 and electrocardiograph 70 used for measurement.

The operation unit 13 is, for example, a keyboard or a mouse, and accepts input operations such as information. The display unit 14 is, for example, a liquid crystal display, and displays information. Note that the operation unit 13 and the display unit 14 may be configured as components that combine operation functions and display functions, such as a touch panel.

ROM 22 and NVRAM 24 are non-volatile memories in which various programs and data are stored. RAM 23 is a volatile memory, and is used, for example, to temporarily store data when executing a program stored in ROM 22 or NVRAM 24.

The NVRAM 24 incorporates a first application 25 and a second application 26. The first application 25 is an application program for constructing an optimal model for estimating a heart rate variability index from the heart rate time series data of the wearable device 50. The second application 26 is an analysis application program for calculating an actual value of the heart rate variability index from electrocardiogram data. In this specification, the heart rate variability index calculated from the electrocardiogram data is defined as an "actual value of the heart rate variability index." Also, in this specification, the heart rate variability index calculated from the heart rate time series data acquired by the wearable device is defined as an "estimated value of the heart rate variability index." For example, the heart rate variability index calculated by the second application 26 is the "actual value of the heart rate variability index." And the heart rate variability index calculated using the optimal model selected by the first application 25 is an example of an "estimated value of the heart rate variability index."

The first app 25 and the second app 26 may be pre-installed in the heart rate variability index estimation device 10, or may be installed in the heart rate variability index estimation device 10 by downloading them from an application store or the website of the manufacturer of the heart rate variability index estimation device 10. The first app 25 is an example of a "heart rate variability index estimation program."

NVRAM 24 includes a first database (hereinafter referred to as "first DB") 27 that stores data used by first application 25, a second database (hereinafter referred to as "second DB") 28, and a third database (hereinafter referred to as "third DB") 29.

The first DB 27 stores the heart rate time series data received from the wearable device 50 in association with the user information, device information, etc. attached to the heart rate time series data. The second DB 28 stores the electrocardiogram data received from the electrocardiograph 70 in association with the user information, electrocardiograph information, etc. attached to the electrocardiogram data. The third DB 29 stores the trained model in association with model information indicating the model of the wearable device 50.

The CPU 21 executes various processes by executing programs stored in the ROM 22 and the NVRAM 24. Both the first application 25 and the second application 26 are executed by the CPU 21. The CPU 21 is an example of a "computer." The controller 11 may also be an example of a "computer."

Figure 2 is a functional block diagram of the heart rate variability index estimation device 10. The heart rate variability index estimation device 10 includes a first acquisition unit 101, a second acquisition unit 102, a heart rate variability index actual value calculation unit 108, an explanatory variable generation unit 103, an explanatory variable combination unit 104, a learning unit 105, an evaluation index calculation unit 106, and an optimal model selection unit 107.

The first acquisition unit 101 has a function of storing and acquiring, for each of the multiple wearable devices 50, heart rate time series data indicating the heart rate of subject X measured by the wearable device 50 in a time series in the first DB 27. The second acquisition unit 102 has a function of storing and acquiring, for each of the multiple electrocardiographs 70, electrocardiogram data indicating the heart rate fluctuations of subject X measured by the electrocardiograph 70 in the second DB 28.

The explanatory variable generation unit 103 has a function of converting the heart rate time series data acquired by the first acquisition unit 101 into beat interval time series data, calculating multiple feature amounts from the beat interval time series data, and setting them as explanatory variables. The explanatory variable combination unit 104 has a function of combining the explanatory variables generated by the explanatory variable generation unit 103. The heart rate variability index actual value calculation unit 108 has a function of calculating the actual values of LF and HF, which are heart rate variability indexes, from the electrocardiogram data acquired by the second acquisition unit 102.

The learning unit 105 has a function of selecting multiple regression models with HF and LF, which are heart rate variability indices, as objective variables, and for each selected regression model, applying a regression model algorithm to the combined data of explanatory variables combined by the explanatory variable combination unit 104 to learn. The evaluation index calculation unit 106 has a function of comparing the actual HF value and the actual LF value calculated by the heart rate variability index actual value calculation unit 108 with the estimated HF value and the estimated LF value estimated by the learning unit 105 for each selected regression model to calculate an evaluation index. The optimal model selection unit 107 has a function of selecting a model consisting of a regression model and a combination of explanatory variables with which the evaluation index calculated by the evaluation index calculation unit 106 is the best, as the optimal model.

Next, the operation of the heart rate variability index estimation device 10 will be described with reference to FIG. 3. FIG. 3 is a flowchart showing an example of a control procedure for a model construction process. When the heart rate variability index estimation device 10 receives a model construction instruction to construct an optimal model, the CPU 21 starts up the first application 25 and starts executing the model construction process shown in FIG. 3.

In this embodiment, the first DB 27 and the second DB 28 each store 456 pieces of data during cooking in a standing position, 27 pieces of data during lying down, 25 pieces of data during standing, and 162 pieces of data during viewing slides in a standing position, for a total of 670 pieces of data. All of this data is from healthy university students. In addition, this data is measured using the same model of wearable device 50 (A370 manufactured by Polar) and the same model of electrocardiograph 70 (electrocardiogram heart rate sensor DL-310 of DL-5000 manufactured by S&ME).

The heart rate variability index estimation device 10 receives heart rate time series data of the heart rate measured by each wearable device 50 from the multiple wearable devices 50 via the communication unit 12, and stores the data in the first DB 27. The first DB 27 is an example of a "first storage unit". The heart rate variability index estimation device 10 also receives electrocardiogram data measured by each electrocardiograph 70 from the multiple electrocardiographs 70 via the communication unit 12, and stores the data in the second DB 28. Receiving the heart rate time series data and storing it in the first DB 27 is an example of a "first acquisition step", and receiving the electrocardiogram data and storing it in the second DB 28 is an example of a "second acquisition step".

First, the CPU 21 selects multiple regression models (S1). Each regression model has its own characteristics, and multiple regression models are selected in order to select the most accurate regression model. For example, the CPU 21 displays a list of selectable regression models on the display unit 14, and allows the operator to select from multiple regression models via the operation unit 13. The objective variables of each regression model are the LF and HF of heart rate variability.

The CPU 21 acquires the actual value of the heart rate variability index (S2). That is, the CPU 21 uses the second application 26 to calculate the actual LF value and the actual HF value from the electrocardiogram data for each of the multiple electrocardiogram data stored in the second DB 28. The CPU 21 then stores the calculated actual LF value and actual HF value in the RAM 23 or the NVRAM 24. The RAM 23 or the NVRAM 24 is an example of a "second storage unit." Also, S2 is an example of a "heart rate variability index actual value calculation step."

The CPU 21 selects the model of the wearable device 50 (S3). The measurement method of the heart rate measurement unit 51 of the wearable device 50 differs depending on the model, and the accuracy of the heart rate measurement differs. In order to suppress the variation in the heart rate time series data depending on the model of the wearable device 50, one model of the wearable device 50 that measures the heart rate that is the source of the learning data is selected. In this description, the heart rate time series data is acquired by the same model of wearable device 50 (A370 manufactured by Polar Inc.), so in S3, the A370 manufactured by Polar Inc. is selected as the model of the wearable device 50. For example, the CPU 21 may display a list of models of the wearable device 50 on the display unit 14 based on the device information associated with the heart rate time series data stored in the first DB 27, and accept an operation to select the model of the wearable device 50 via the operation unit 13.

When multiple regression models and the model of the wearable device 50 are selected, the CPU 21 converts the heart rate time series data stored in the first DB 27 into beat interval time series data in order to calculate the heart rate variability index (S5). For example, the CPU 21 converts the extracted heart rate time series data into beat interval time series data using the following formula 1. In formula 1, the heart rate time series data is [hr(t)|t=1,...,n], and the beat interval time series data is [beat_i(t)|t=1,...,n].

In this description, all of the heart rate time series data stored in the first DB 27 is converted into beat interval time series data in S5, but only a portion of the heart rate time series data used as learning data may be converted into beat interval time series data. If the first DB 27 stores heart rate time series data for multiple models of wearable devices 50, the heart rate time series data associated with the model of wearable device 50 selected in S3 may be extracted and converted.

The CPU 21 calculates explanatory variables (feature amounts) from the beat interval time series data converted in S5 (S7). S7 is an example of an "explanatory variable generation process" or an "explanatory variable generation step." Table 1 below shows an example of explanatory variables.

The explanatory variable LF_w shown as number 1 in Table 1 is the power spectrum value of LF from 0.04 to 0.15 Hz obtained by Fourier transform of the beat interval time series data beat_i(t). The explanatory variable LF_w is an index that reflects the activity of both the sympathetic and parasympathetic nerves.

The explanatory variable HF_w shown in number 2 in Table 1 is the power spectrum value of HF from 0.15 to 0.4 Hz obtained by Fourier transform of the beat interval time series data beat_i(t). The explanatory variable HF_w is an index reflecting the activity of the parasympathetic nervous system.

The explanatory variable avg shown in number 3 in Table 1 is the average value of the beat interval time series data beat_i(t). In other words, the explanatory variable avg is one of the time domain heart rate variability index data, and is the average value of the beat interval within a period. The explanatory variable avg indicates an increase or decrease in the amount of activity.

The explanatory variable sd shown in number 4 in Table 1 is the standard deviation of the beat interval time series data beat_i(t). In other words, the explanatory variable sd is one of the time domain heart rate variability index data, and indicates the standard deviation of the beat interval within a period. In other words, the explanatory variable sd indicates the variation in the beat interval, and is an index of the intensity of tension.

The explanatory variable cv shown in number 5 in Table 1 is the coefficient of variation of the beat interval time series data beat_i(t) calculated by the following formula 2. In other words, the explanatory variable cv is one of the time domain heart rate variability index data, and indicates the coefficient of variation of the beat interval within a period. The explanatory variable cv is a non-dimensional version of the explanatory variable sd, and is treated in the same way as the explanatory variable sd.

The explanatory variable pp shown in number 6 in Table 1 is the difference between the maximum and minimum values of the beat interval time series data beat_i(t). In other words, the explanatory variable pp indicates the maximum fluctuation range of the beat interval within the period.

The explanatory variable rmssd shown in number 7 in Table 1 is the square root of the mean square of the difference between the adjacent beat interval time series data beat_i(t) shown in Formula 1 (see Formula 3 below). In other words, the explanatory variable rmssd is one of the time domain heart rate variability index data, and is used as an index of vagal tone strength.

The explanatory variable dif_avg shown as number 8 in Table 1 is the average value of the differential time series beat_dif(t). Here, beat_dif(t) indicates the difference, which is the value obtained by subtracting the previous term beat_i(t-1) from the next term beat_i(t) in the beat interval time series data. In other words, the explanatory variable dif_avg represents the fluctuation with the baseline beat interval removed.

The explanatory variable dif_min shown as number 9 in Table 1 is the minimum value of the differential time series beat_dif(t). In other words, the explanatory variable dif_min indicates the minimum value of the fluctuation with the baseline beat interval within the period removed.

The explanatory variable dif_max shown in number 10 in Table 1 is the maximum value of the differential time series beat_dif(t). In other words, the explanatory variable dif_max indicates the maximum value of the fluctuation with the baseline beat interval within the period removed.

The explanatory variable avg_30 shown in number 11 in Table 1 is the value obtained by dividing the average value of the beat interval time series data beat_i(t) for 30 seconds (t = 31 to 60) by the average value of the average value of the beat interval time series data beat_i(t) for the previous 30 seconds (t = 1 to 30). In other words, the explanatory variable avg_30 divides all the data into 30-second intervals and calculates the ratio of the average beat interval value for each 30-second interval to the average beat interval value for the first 30 seconds. The explanatory variable avg_30 indicates a tendency for large fluctuations from the start point.

The explanatory variable avg_60 shown in number 12 in Table 1 is the value obtained by dividing the average value of the beat interval time series data beat_i(t) for 30 seconds (t = 61 to 120) by the average value of the average value of the beat interval time series data beat_i(t) for the previous 30 seconds (t = 1 to 60). In other words, the explanatory variable avg_60 divides all the data into 60-second intervals and calculates the ratio of the average beat interval value for each 60-second interval to the average beat interval value for the first 60 seconds. The explanatory variable avg_60 indicates a tendency for large fluctuations from the start point.

The explanatory variable avg_90 shown in number 13 in Table 1 is the value obtained by dividing the average value of the beat interval time series data beat_i(t) for 30 seconds (t = 91 to 180) by the average value of the average value of the beat interval time series data beat_i(t) for the previous 30 seconds (t = 1 to 90). In other words, the explanatory variable avg_90 divides all the data into 90-second intervals and calculates the ratio of the average beat interval value for each 90-second interval to the average beat interval value for the first 90 seconds. The explanatory variable avg_90 indicates a tendency for large fluctuations from the start point.

As shown in FIG. 3, after calculating the explanatory variables in S7, the CPU 21 arbitrarily combines the calculated explanatory variables (S9). S9 is an example of an "explanatory variable combination process" or "explanatory variable combination step". Combining explanatory variables makes it easier to eliminate explanatory variables that may reduce the estimation rate in advance. For example, the CPU 21 selects multiple arbitrary combinations from the combinations of explanatory variables shown by numbers 1 to 13 in Table 1. The number of combinations of explanatory variables as is is r out of n explanatory variables, and the total number of combinations is nCr. From these combinations, a combination is selected in accordance with the selection criteria. In this embodiment, the explanatory variables are combined as shown in Table 2 below.

The combination shown in combination number 1 in Table 2 is a combination of explanatory variable LF_w shown in number 1 and explanatory variable HF_w shown in number 2 in Table 1. The explanatory variables LF_w and HF_w are actual values obtained by Fourier transforming the heart rate time series data of the wearable device.

The combination shown in combination number 2 in Table 2 is a combination of explanatory variables number 3 to number 13 in Table 1. In other words, the heart rate variability index data calculated from the heart rate time series data of the wearable device 50 is aggregated (consolidated).

The combination shown in combination number 3 in Table 2 is a combination of explanatory variables among the combinations shown in combination number 2 above, in which the absolute value of the correlation with the objective variable is 0.5 or more. In other words, for the same subject X (X1), the correlation coefficient is calculated between the time domain heart rate variability index data obtained from the heart rate time series data of the wearable device 50 (50A) stored in the first DB 27 and the LF and HF obtained from the electrocardiograph 70 (70A) stored in the second DB 28, and explanatory variables with relatively high correlations are aggregated (consolidated).

The combination shown in combination number 4 in Table 2 is a combination of explanatory variables among the combinations shown in combination number 2 above, in which the absolute value of the correlation with the objective variable is less than 0.5. In other words, for the same subject X (X1), the correlation coefficient is calculated between the time domain heart rate variability index data obtained from the heart rate time series data of the wearable device 50 (50A) stored in the first DB 27 and the LF and HF obtained from the electrocardiograph 70 (70A) stored in the second DB 28, and explanatory variables with relatively low correlation are aggregated (consolidated). The combination shown in combination number 4 is used to verify the combination shown in combination number 3.

The combination shown in combination number 5 in Table 2 is a principal component that has an eigenvalue of 1 or more when principal component analysis is performed on the 11 explanatory variables shown in numbers 3 to 13 in Table 1. Principal component analysis is performed to reduce the number of explanatory variables (features) 11 from numbers 3 to 13, and they are replaced with the extracted principal component. The data for the principal component is the principal component score.

We will now explain combination numbers 3 and 4 in Table 2 in detail. Table 3 shows the classification of explanatory variables for combination numbers 3 and 4 in Table 2.

As shown in combination number 3 in Table 3, the correlation of LF was high with the explanatory variable avg shown in number 3 in Table 1, the explanatory variable sd shown in number 4 in Table 1, the explanatory variable pp shown in number 6 in Table 1, the explanatory variable rmssd shown in number 7 in Table 1, and the explanatory variable dif_max shown in number 10 in Table 1. On the other hand, the correlation of HF was high with the explanatory variable avg shown in number 3 in Table 1, the explanatory variable sd shown in number 4 in Table 1, the explanatory variable pp shown in number 6 in Table 1, the explanatory variable rmssd shown in number 7 in Table 1, the explanatory variable dif_min shown in number 9 in Table 1, and the explanatory variable dif_max shown in number 10 in Table 1.

On the other hand, as shown in combination number 4 in Table 3, the correlation of LF was low with explanatory variable cv shown in number 5 in Table 1, explanatory variable dif_avg shown in number 8 in Table 1, explanatory variable dif_min shown in number 9 in Table 1, explanatory variable avg_30 shown in number 11 in Table 1, explanatory variable avg_60 shown in number 12 in Table 1, and explanatory variable avg_90 shown in number 13 in Table 1. The correlation of HF was low with explanatory variable cv shown in number 5 in Table 1, explanatory variable dif_avg shown in number 8 in Table 1, explanatory variable avg_30 shown in number 11 in Table 1, explanatory variable avg_60 shown in number 12 in Table 1, and explanatory variable avg_90 shown in number 13 in Table 1.

Therefore, explanatory variables with a correlation coefficient of 0.5 or more were selected mainly from time-domain heart rate variability index data. Because the explanatory variables to be combined can be limited to time-domain heart rate variability index data, the total number of explanatory variable combinations can be reduced, thereby reducing the processing load. Furthermore, with combination number 3, explanatory variables that are highly correlated with electrocardiogram data can be consolidated, which is expected to improve the estimation rate.

Here we will explain combination number 5 in Table 2 in more detail. By performing principal component analysis on the explanatory variables shown in numbers 3 to 13 in Table 1, there were three principal components with eigenvalues of 1 or more. The explanatory variables with large principal component loadings for each principal component are as shown in Table 4.

As shown in Table 4, explanatory variables with large principal component loadings for the first principal component were the explanatory variable dif_avg shown at number 8 in Table 1, the explanatory variable avg_30 shown at number 11 in Table 1, the explanatory variable avg_60 shown at number 12 in Table 1, and the explanatory variable avg_90 shown at number 13 in Table 1. In addition, explanatory variables with large principal component loadings for the second principal component were the explanatory variable avg shown at number 3 in Table 1, the explanatory variable rmssd shown at number 7 in Table 1, and the explanatory variable dif_min shown at number 9 in Table 1. In addition, explanatory variables with large principal component loadings for the third principal component were the explanatory variable sd shown at number 4 in Table 1, the explanatory variable cv shown at number 5 in Table 1, and the explanatory variable pp shown at number 6 in Table 1.

For each principal component shown in Table 4, the number of explanatory variables generated in S7 is reduced by replacing explanatory variables with those with large principal component loadings. This makes it possible to reduce the total number of combinations of explanatory variables and reduce the processing load.

As shown in FIG. 3, the CPU 21 repeats the learning process of S13 and the evaluation index calculation process of S15 a predetermined number of times for each regression model selected in S1, and selects the optimal estimation model for estimating the heart rate variability indexes (LF, HF) (S11 to S21).

That is, the CPU 21 extracts an unprocessed regression model that has not been subjected to learning and evaluation from the four regression models selected in S1 (S11). For example, if four regression models, Gaussian process regression, linear multiple regression of degree 1, random forest regression, and SVM regression, are selected in S1, one of the regression models is extracted.

Then, the CPU 21 executes a learning process for the regression model extracted in S11 (S13). S13 is an example of a "learning step". In the learning process, 70% of the feature amounts calculated in S7 are used as learning data. The remaining 30% are used as evaluation data. The method for dividing the beat interval time series data is random so as to avoid bias in the data.

The CPU 21 calculates an evaluation index using the evaluation data for the trained model that has undergone the learning process in S13 (S15). S15 is an example of an "evaluation index calculation process" or an "evaluation index calculation step". In this embodiment, the coefficient of determination calculated using the following formula 4 is used as the evaluation index. ^R2 is the coefficient of determination, the HF bar is the average value of HF obtained from the electrocardiogram data, and the arc HF is an estimated value of HF by the model.

The evaluation index is not limited to the coefficient of determination, but may be the root mean square error (RMSE), the mean absolute error (MAE), etc.

As shown in FIG. 3, the CPU 21 determines whether the learning process of S13 and the evaluation index calculation process of S15 have been performed a predetermined number of times (S17). The predetermined number of times can be set arbitrarily, but is set to 30 times in this explanation. By repeating the learning process and the evaluation index calculation process 30 times while arbitrarily changing the combination of explanatory variables, 30 determination coefficients, which serve as evaluation indexes, are calculated for one regression model.

When the CPU 21 has performed the learning process and the evaluation index calculation process for the regression models extracted in S11 a predetermined number of times (S17: YES), it determines whether or not the process has been completed for all regression models selected in S1 (S19). For example, if the process has been completed for only some of the regression models selected in S1 (S19: NO), the CPU 21 returns to the process of S11 and extracts unprocessed regression models. The process from S13 onwards is the same as above, so the explanation will be omitted.

After performing the learning process of S13 and the evaluation index calculation process of S15 a predetermined number of times for all regression models selected in S1 (S19: YES), the CPU 21 selects an optimal model (S21). S21 is an example of an "optimal model selection process" or an "optimal model selection step."

For example, the CPU 21 calculates statistically significant differences by analysis of variance of 30 coefficients of determination obtained by 30 learning processes and evaluation index calculation processes, and compares the types of models and the selection methods of explanatory variables to verify the accuracy between models. An example of the verification results is shown in Table 5. Table 5 shows the combinations of explanatory variables that maximize the average value of the 30 coefficients of determination in each regression model of Gaussian process regression, first-order linear multiple regression, random forest regression, and SVM regression, separated by LF and HF.

As shown in Table 5, in the Gaussian process regression, when the combination of explanatory variables shown in Table 2 (combination number 3) was used, the average value of the coefficient of determination for LF was maximized, and when the combination of coefficients of determination for combination number 2 shown in Table 2 was used, the average value of the coefficient of determination for HF was maximized. In the first-order linear multiple regression, when the combination of explanatory variables shown in Table 2 (combination number 2) was used, the average value of the coefficient of determination for LF was maximized, and when the combination of coefficients of determination for combination number 3 shown in Table 2 was used, the average value of the coefficient of determination for HF was maximized.

In the random forest regression, when the combination of explanatory variables shown in Table 2 (combination number 2) was used, the average value of the coefficient of determination for LF was maximized, and when the combination of coefficients of determination shown in Table 2 (combination number 3) was used, the average value of the coefficient of determination for HF was maximized. In the SVM regression, when the combination of explanatory variables shown in Table 2 (combination number 1) was used, the average value of the coefficient of determination for LF was maximized, and when the combination of coefficients of determination shown in Table 2 (combination number 5) was used, the average value of the coefficient of determination for HF was maximized.

Therefore, it can be seen that the combination of explanatory variables with a high estimation rate differs for each regression model. In other words, by executing the learning process and the evaluation index calculation process for each of multiple regression models, the regression model that provides the best estimation can be selected.

Figure 4 is a graph showing the average and standard deviation of the coefficients of determination when HF was maximized in each regression model in Table 5. Figure 5 is a graph showing the average and standard deviation of the coefficients of determination when LF was maximized in each regression model in Table 5. Note that the horizontal axis of each graph shows G for Gaussian process regression, M for linear multiple regression of degree 1, R for random forest regression, and S for SVM regression. The vertical axis of each graph shows the average value of the coefficients of determination calculated by repeating the processes of S13 to S15 30 times.

As shown in Figure 4, the coefficient of determination for LF was largest for random forest regression, followed by first-order linear multiple regression, Gaussian process regression, and SVM regression. As shown in Figure 5, the coefficient of determination for HF was largest for random forest regression, followed by Gaussian process regression, first-order linear multiple regression, and SVM regression.

In this way, the evaluation of the regression models differs between HF and LF. Therefore, by performing the learning process and the evaluation index calculation process for multiple regression models while arbitrarily selecting a combination of explanatory variables, it is possible to select the combination of explanatory variables and the regression model that provides the best estimation of LF and HF. By selecting the optimal model, it becomes possible to examine the psychological state of the subject from the heart rate measured by the wearable device 50.

The CPU 21 selects the optimal model as a model consisting of a combination of the regression model (random forest regression in FIG. 4 and FIG. 5) with the maximum average value of the coefficient of determination and explanatory variables. The CPU 21 then associates the model of the wearable device 50 selected in S3 (Polar A370) with the selected optimal model, and stores it in the third DB 29. The CPU 21 then ends the process shown in FIG. 3.

Incidentally, when constructing optimal models for LF and HF, it is possible to use different combinations of explanatory variables and adopt the model that maximizes the coefficient of determination. However, constructing the models using the same combination of explanatory variables as much as possible can avoid the complexity and effort involved in calculating estimated values for LF and HF.

Therefore, for the 30 coefficients of determination obtained from constructing 20 types of models [four regression models (Gaussian process regression, first-order linear multiple regression, random forest regression, SVM regression) and five combinations of explanatory variables (combinations in Table 2)], one-way analysis of variance was performed for the five combinations of explanatory variables for each regression model, and combinations of explanatory variables that can be treated as being of the same quality as the combinations of explanatory variables in random forest regression in Table 5 were extracted. The results are shown in Table 6.

As shown in Table 6, the coefficient of determination based on combination number 3 in Table 2, which was used in the random forest regression for HF with the maximum coefficient of determination, and the combination of explanatory variables treated as being homogeneous, was the combination of combination number 2 in Table 2. As a result, the algorithms for the estimation models for LF and HF are both random forest regression, and the combination of explanatory variables is the combination of combination number 2 in Table 2, and it is believed that by adopting one that includes all of the time-domain heart rate variability index data (avg, SD, cv, rmssd), it is possible to avoid the complexity and effort involved in calculating the estimated values for LF and HF. Therefore, the CPU 21 associates the random forest regression with the combination of explanatory variables when the coefficient of determination was maximum (the combination shown in combination number 3), and stores this as an optimal model in the third DB 29.

In addition, in the process of S21, it is preferable to store the optimal model in the third DB 29 in association with the model of the wearable device 50 selected in S1. This is because the measurement characteristics of the heart rate differ depending on the model of the wearable device 50.

Figure 6 is a graph showing heart rate time series data measured simultaneously using the first and second wearable devices. A Polar V800 was used as the first wearable device, and a Polar A370 was used as the second wearable device. Measurements were performed under the following three conditions. The first condition is that the first wearable device is worn on the left arm. The second condition is that the wristwatch-type second wearable device is worn on the left arm. The third condition is that the wristwatch-type second wearable device is worn on the right arm. It is generally recommended that wristwatch-type heart rate monitors be worn on the arm opposite the dominant arm.

As shown in Figure 6, the large movements are similar under conditions 1 to 3. However, as shown in conditions 1 and 2, the small fluctuations during the time periods when large movements occurred for the first and second wearable devices worn on the same left arm hardly match, indicating that errors occur depending on the model. Also, as shown in conditions 2 and 3, errors occur depending on the arm on which the device is worn, even when using the same model.

For example, assume that heart rate data based on the heart rate measured by a first wearable device is input to an optimal model constructed based on the heart rate measured by a wearable device 50 of the same model as the second wearable device. In this case, the HF and LF/HF calculated from the optimal model are considered to be unreliable as evaluation indices, and applying conventional interpretations of the indices as is may lead to misinterpretation of the results.

Therefore, the optimal model is created by selecting one model of wearable device 50 (S3 in FIG. 3). This allows an optimal model to be constructed for each wearable device 50, suppressing estimation errors due to differences in the model of the wearable device 50, and making it possible to estimate the heart rate variability index with high accuracy.

<Application of the estimation model>
Based on the above estimation model selection method, an optimal model for the frequency domain index HF was derived using a random forest regression algorithm, and an HF estimate was calculated from the optimal model using heart rate time series data collected by the wearable terminal 50 in other experiments, and compared with previous results.

The validation data was time series heart rate data obtained by a wearable device (Polar A370) from 11 healthy female high school students. The 11 students learned how to manage fire when using a gas stove using VR (Virtual Reality) teaching materials before participating in a cooking class. It is known that students who learn using VR teaching materials before participating in a cooking class tend to be more careful about fire management during the class and become more nervous when cooking compared to before the class (e.g., Kawahara Yuko, Ito Hisatoshi, Shimizu Ayako, Maruyama Tomomi: Learning Effects of VR-Based Learning Materials on "Fire," Abstracts of Research Presentations at the Japan Society of Food Science and Technology Conference, 31(0), p.31, 2019).

The estimated HF values (HF_est) were separated into those before the cooking lesson (Pre) and those during heating (Heating), and the average values for the 11 participants were calculated to verify whether the estimated HF values during heating had decreased compared to before the cooking lesson. For statistical judgment, a test was conducted on the difference in the average values before and during heating, and a judgment was made with a significance probability of less than 1%. The results of the estimated HF values are shown in Figure 7. The values during heating were significantly lower than before the cooking lesson, which can be interpreted as indicating that participants were nervous when cooking. This result supports the logical development of the previous literature mentioned above.

<Usage example>
For example, the optimal model calculated as described above is used for emotion determination in a lecture or the like. For example, a wearable device 50 (e.g., Polar A370) corresponding to the optimal model is worn by a student attending a lecture. The wearable device 50 measures the heart rate of the student during the lecture and transmits the heart rate time series data to, for example, a cloud server. The cloud server accumulates the heart rate time series data of the student. The cloud server stores the heart rate in association with device information including the model of the wearable device 50 or identification information that identifies the student.

A personal computer (an example of a psychological state evaluation device, hereinafter referred to as "PC") used to evaluate the psychological state of a student stores an optimal model selected for each model of wearable terminal 50. The PC accepts an operation to select one model of wearable terminal 50 corresponding to the optimal model. The PC also accepts an operation to select a heart rate variability index (e.g., LF, HF) to be estimated. The PC downloads the heart rate time series data associated with the selected model from the cloud server. Then, the PC reads out the optimal model corresponding to the selected model and reads the downloaded heart rate time series data. The optimal model calculates and outputs an estimate of the selected heart rate variability index from the input heart rate time series data using an algorithm including a combination of explanatory variables used when the model was constructed. For example, the PC displays or draws the calculated HF and LF estimates on a display. The PC may calculate the ratio between the estimated value of LF and the estimated value of HF calculated from the optimal model and display or print it out.

Professors who gave lectures will be able to analyze students' interest in the lecture based on the output results and revise the content of the lecture.

As described above, the heart rate variability index estimation device 10 of this embodiment converts the heart rate time series data of the wearable terminal 50 into beat interval time series data, calculates multiple feature amounts from the beat interval time series data to set them as explanatory variables, and applies a regression model algorithm to the data combining the explanatory variables to learn. Then, the model with the best evaluation index calculated based on the actual value of the heart rate variability index calculated from the electrocardiogram data and the estimated value of the heart rate variability index calculated from the learned model is selected as the optimal model, so that the estimated value of the heart rate variability index calculated by the selected optimal model approximates the actual value of the heart rate variability index calculated from the electrocardiogram data. Therefore, according to the heart rate variability index estimation device 10 of this embodiment, it is possible to construct a model that can estimate heart rate variability with high accuracy from the heart rate measured by the wearable terminal 50. This makes it possible to interpret the estimated heart rate variability index calculated by inputting the heart rate measured using the heart rate measurement function of a wearable device into the optimal model in the same way as a heart rate variability index calculated from conventional electrocardiogram data, and to examine the subject's psychological state.

The present invention is not limited to the above embodiment, and various applications are possible. For example, in the above embodiment, the heart rate variability index estimation device 10 receives heart rate time series data and electrocardiogram data from the wearable terminal 50 and the electrocardiograph 70, and stores them in the first DB 27 and the second DB 28, thereby acquiring the heart rate time series data and the electrocardiogram data. In response to this, the wearable terminal 50 and the electrocardiograph 70 may each transmit the measured heart rate to the heart rate variability index estimation device 10 at any time, and the heart rate variability index estimation device 10 may generate the heart rate time series data and the electrocardiogram data.

For example, in the above embodiment, the heart rate time series data and the electrocardiogram data are acquired via the communication unit 12. Alternatively, the heart rate time series data and the electrocardiogram data may be stored in an external storage device such as a cloud server or a storage medium such as a memory card, and the heart rate variability index estimation device may read and acquire the heart rate time series data and the electrocardiogram data from the external storage device.

For example, in the above embodiment, explanatory variables 1 to 13 are given as examples, but explanatory variables other than these may be used. In addition, in the above embodiment, a model is constructed using combination numbers 1 to 5 as examples, but explanatory variables may be combined in other ways. However, by using the combination of explanatory variables with combination number 3, explanatory variables that are highly correlated with electrocardiogram data can be aggregated (collected) to create a combination, and an improvement in the estimation rate can be expected. Furthermore, by limiting the heart rate variability index data that serves as the explanatory variable to time domain beat variability index data, the total number of combinations of explanatory variables can be reduced, and the processing load can be reduced. Furthermore, by using the combination of explanatory variables with combination number 5, the number of explanatory variables to be combined is narrowed down, and the total number of combinations of explanatory variables can be reduced, and the processing load can be reduced.

The actual and estimated values of the heart rate variability index do not have to be the actual and estimated values of HF and LF. For example, the actual and estimated values of the heart rate variability index may be pNN50 (percent of difference between adjacent normal RR intervals greater than 50 ms), rMSSD (rootmean square successive difference), SDNN (standard deviation of NN intervals), SDANN (standard deviation of the normal-normal R-R period), etc. In other words, the actual and estimated values of the heart rate variability index are not limited to indices that can be used to examine psychological states, and may also be indices that can be used to predict the severity of illness.

For example, the optimal model selected by the optimal model selection unit 107 may not have the same combination of explanatory variables between the regression model that calculates HF and the regression model that calculates LF. However, by using the same combination of explanatory variables, the complexity and effort required for calculating the estimated values can be reduced.

For example, the process of S1 in FIG. 3 may be omitted and multiple regression models may be specified in advance. However, by allowing multiple regression models to be selected in S1, it becomes easier to select a regression model that suits the purpose.

For example, the process of S3 in Fig. 3 may be omitted. However, by selecting the model of the wearable device 50 in S3, even when the heart rate is measured using a plurality of wearable devices 50, a model according to the characteristics of each model of the wearable device 50 can be constructed.
For example, when selecting an optimal model, it is not necessary to calculate the average value of the evaluation index (e.g., the coefficient of determination) for each regression model. However, by calculating the average value of the evaluation index for each regression model and selecting the model with the largest average value of the evaluation index as the optimal model, it becomes possible to select a regression model having an algorithm suitable for estimating a heart rate variability index as the optimal model.

For example, in the above embodiment, the second application 26 calculates the actual value of the heart rate variability index, but the first application 25 may calculate the actual value of the heart rate variability index.

For example, in the flowchart disclosed in the above embodiment, the execution order of multiple processes in any of the multiple steps can be changed as desired or the processes can be executed in parallel, as long as no inconsistencies occur in the process content.

The processes disclosed in the above embodiments may be executed by a single CPU, multiple CPUs, hardware such as an ASIC, or a combination of these. The processes disclosed in the above embodiments may be realized in various forms, such as a recording medium having a program for executing the processes recorded thereon, or a method or system.

10 Heart rate variability index estimation device 21 CPU
27 First DB
28 2nd DB
101 First acquisition unit 102 Second acquisition unit 103 Explanatory variable calculation unit 104 Explanatory variable combination unit 105 Learning unit 106 Evaluation index calculation unit 107 Optimal model selection unit 108 Heart rate variability index actual value calculation unit

Claims (11)

a first acquisition unit that acquires, for each of a plurality of wearable devices that are attached to a part of a body of a subject and have a heart rate measurement function that measures the heart rate of the subject based on a pulse wave, heart rate time series data that indicates the heart rate of the subject measured by the wearable device in a time series;
a second acquisition unit that acquires electrocardiogram data indicating heart rate fluctuations of the subject measured by a plurality of electrocardiographs that are attached to the subject and measure electrical signals generated by the heart of the subject, for each of the plurality of electrocardiographs;
an explanatory variable generating unit that converts the heart rate time series data acquired by the first acquiring unit into beat interval time series data, calculates a plurality of feature amounts from the beat interval time series data, and sets the feature amounts as explanatory variables;
an explanatory variable combination unit that combines a plurality of explanatory variables generated by the explanatory variable generation unit;
a heart rate variability index actual value calculation unit that calculates an actual value of a heart rate variability index from the electrocardiogram data acquired by the second acquisition unit;
a learning unit that selects a plurality of regression models using the actual value of the heart rate variability index as a response variable, and applies a regression model algorithm to the combination data of the explanatory variables combined by the explanatory variable combination unit for each selected regression model to learn the regression models;
an evaluation index calculation unit that calculates an evaluation index for each of the selected regression models based on the actual value of the heart rate variability index calculated by the heart rate variability index actual value calculation unit and an estimated value of the heart rate variability index calculated from the trained model trained by the learning unit;
an optimal model selection unit that selects, as an optimal model, a model consisting of a combination of a regression model and explanatory variables that has the best evaluation index calculated by the evaluation index calculation unit;
having
A heart rate variability index estimation device comprising: 2. The heart rate variability index estimation device according to claim 1,
the explanatory variable combination unit analyzes a correlation between each of the plurality of feature amounts calculated by the explanatory variable generation unit and a value calculated from the electrocardiogram data acquired by the second acquisition unit and corresponding to the feature amount, and combines feature amounts having a high correlation among the plurality of feature amounts;
A heart rate variability index estimation device comprising: 3. The heart rate variability index estimation device according to claim 1,
The explanatory variable combination unit selects explanatory variables indicating time domain pulsation variation index data as combination targets among the explanatory variables generated by the explanatory variable generation unit;
A heart rate variability index estimation device comprising: 2. The heart rate variability index estimation device according to claim 1,
the explanatory variable combination unit performs principal component analysis on the explanatory variables generated by the explanatory variable generation unit, and replaces the explanatory variables with principal components having large eigenvalues;
A heart rate variability index estimation device comprising: The heart rate variability index estimation device according to any one of claims 1 to 3,
The heart rate variability index is a high frequency component and a low frequency component of the heart rate variability,
the optimal model selected by the optimal model selection unit has the same combination of explanatory variables between a regression model for calculating the high frequency components and a regression model for calculating the low frequency components;
A heart rate variability index estimation device comprising: The heart rate variability index estimation device according to any one of claims 1 to 5,
Selecting the optimum model for each model of the wearable device;
A heart rate variability index estimation device comprising: The heart rate variability index estimation device according to any one of claims 1 to 6,
The estimated value of the heart rate variability index and the actual value of the heart rate variability index indicate indices that can be used to examine a psychological state;
A heart rate variability index estimation device comprising: A method for estimating a heart rate variability index, comprising:
a first acquisition step of acquiring, for each of a plurality of wearable devices having a heart rate measurement function that is attached to a part of a body of a subject and measures the heart rate of the subject based on a pulse wave, heart rate time series data indicating the heart rate of the subject measured by the wearable device in a time series;
a second acquiring step of acquiring electrocardiogram data indicating heart rate fluctuations of the subject measured by a plurality of electrocardiographs each attached to the subject and measuring an electrical signal generated by the heart of the subject, the electrocardiographs;
an explanatory variable generating step of converting the heart rate time series data acquired in the first acquisition step into beat interval time series data, calculating a plurality of feature amounts from the beat interval time series data, and setting the feature amounts as explanatory variables;
an explanatory variable combination step of combining a plurality of explanatory variables generated in the explanatory variable generation step;
a heart rate variability index actual value calculation step of calculating an actual value of a heart rate variability index from the electrocardiogram data acquired in the second acquisition step;
a learning step of selecting a plurality of regression models using the actual value of the heart rate variability index as a response variable, and applying a regression model algorithm to the combination data of the explanatory variables combined in the explanatory variable combination step for each selected regression model to learn the regression models;
an evaluation index calculation step of calculating an evaluation index for each of the selected regression models based on the actual value of the heart rate variability index calculated in the heart rate variability index actual value calculation step and an estimated value of the heart rate variability index calculated from the trained model trained in the learning step;
an optimal model selection step of selecting, as an optimal model, a model consisting of a combination of the regression model and explanatory variables for which the evaluation index calculated in the evaluation index calculation step is the best;
To carry out the following:
A method for estimating a heart rate variability index, comprising: Incorporated into information processing equipment,
The information processing device includes:
a first storage unit configured to store, for each of a plurality of wearable devices that are attached to a part of a body of a subject and have a heart rate measurement function that measures the heart rate of the subject based on a pulse wave , heart rate time series data that indicates the heart rate of the subject measured by the wearable device in a time series;
a second storage unit configured to store, for each of a plurality of electrocardiographs attached to the subject and measuring an electrical signal generated by the heart of the subject, an actual value of a heart rate variability index calculated from electrocardiogram data indicating the heart rate variability of the subject measured by the electrocardiograph;
Equipped with
A computer of the information processing device
an explanatory variable generation process of converting the heart rate time series data stored in the first storage unit into beat interval time series data, calculating a plurality of feature amounts from the beat interval time series data, and setting the feature amounts as explanatory variables;
an explanatory variable combination process for combining a plurality of explanatory variables generated in the explanatory variable generation process;
A learning process of selecting a plurality of regression models using the actual value of the heart rate variability index as a response variable, and applying a regression model algorithm to the combination data of the explanatory variables combined in the explanatory variable combination process for each selected regression model to learn the regression models;
an evaluation index calculation process for calculating an evaluation index for each of the selected regression models based on the actual value of the heart rate variability index stored in the second storage unit and an estimated value of the heart rate variability index calculated from the trained model trained in the learning process;
an optimal model selection process for selecting, as an optimal model, a model consisting of a combination of a regression model and explanatory variables for which the evaluation index calculated in the evaluation index calculation process is the best;
To carry out the
A heart rate variability index estimation program comprising: 10. The heart rate variability index estimation program according to claim 9,
causing the computer to execute, for each model of the wearable device, the explanatory variable generation process, the explanatory variable combination process, the learning process, the evaluation index calculation process, and the optimal model selection process;
A heart rate variability index estimation program comprising: 11. The heart rate variability index estimation program according to claim 9,
The learning process and the evaluation index calculation process are performed a predetermined number of times for each selected regression model;
In the optimal model selection process, an average value of the evaluation index is calculated for each selected regression model, and a model consisting of a combination of the regression model and explanatory variables having the largest average value of the evaluation index is selected as an optimal model.
A heart rate variability index estimation program comprising:

Images