KR102483567B1 - Method and system for predicting HRV index of 5 minutes or 3 minutes based on artificial intelligence from HRV data of less than 3 minutes - Google Patents

Abstract

The present invention uses the heart rate variability (HRV) measured for a short measurement period, that is, less than 3 minutes, to output the same heart rate variability index as when measuring 5 minutes (or 3 minutes) of the standard recommendation. The heart rate variability measured for a short time of less than 3 minutes is input to the artificial neural network, and the heart rate variability index output from the artificial neural network is output as a heart rate variability (HRV) index of 5 minutes (or 3 minutes). A method and system for predicting a heart rate variability index of 5 minutes (or 3 minutes) from heart rate variability data of a short measurement period of less than 10 minutes.
The system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes of the present invention detects heart rate data from a subject as heart rate data of a measurement period of less than 3 minutes during a measurement period of less than 3 minutes , heart rate monitor; The heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device is applied to the HRV (heart rate variability) index prediction model of the previously trained artificial neural network using deep learning, and the HRV index output from the HRV index prediction model is 3 Characterized in that it includes; a calculation processing unit that outputs as a minute or five-minute HRV indicator.
As the input data of the HRV index prediction model of the present invention, a linear regression equation is obtained from the heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device, and the 3-minute heart rate data or the 5-minute heart rate data is obtained by using the linear regression equation , predicted 3-minute heart rate data or predicted 5-minute heart rate data, obtained by detecting and performing FFT (Fast Fourier Transform) on the predicted 3-minute heart rate data or predicted 5-minute heart rate data, or predicted 3-minute heart rate FFT data or prediction 5 Contains heart rate FFT data, per minute.

Description

Method and system for predicting HRV index of 5 minutes or 3 minutes based on artificial intelligence from HRV data of less than 3 minutes}

The present invention uses the heart rate variability (HRV) measured for a short measurement period, that is, less than 3 minutes, to output the same heart rate variability index as when measuring 5 minutes (or 3 minutes) of the standard recommendation. A short measurement that inputs heart rate variability measured for a short time of less than 3 minutes to the artificial neural network, and outputs the heart rate variability index output from the artificial neural network as a 5-minute (or 3-minute) heart rate variability (HRV) index. A method and system for predicting a 5-minute (or 3-minute) heart rate variability index from heart rate variability data of a section and a system thereof.

The heart rate variability (HRV) test is a useful, simple, and non-invasive method to evaluate the autonomic nervous system control function. HRV is a representation of successive heart beats and the time interval between beats as time series data.

The HRV index, which is a result of analyzing the HRV, includes an HRV time domain index, which is an index of a result of analyzing the HRV signal in the time domain, and an HRV frequency domain index, which is an index of a result of analyzing the HRV signal in the frequency domain. HRV time domain indicators include mean heart rate (mHR), SDNN (standard deviation of normal-to-normal interval), RMSSD (root mean square differences of successive R-R intervals), and HRV frequency domain indicators include There are VLF (very low frequency, range <= 0.04Hz), LF (power in low frequency, range 0.04~0.15Hz), HF (power in high frequency, range 0.15~0.4Hz) and LF/HF. Since the HRV indicators are described in detail in Patent Registration No. 10-0493714, a detailed description thereof will be omitted.

In order to obtain these indicators, the measurement time of time series data can be an important issue. Time-domain indicators are bound to be time-dependent. In the case of SDNN or RMSSD, in particular, the size increases as the measurement time increases. Therefore, HRV time domain indicators must be compared with those of the same measurement time. The frequency band of the VLF having the longest period among the frequency domain indicators is 0.04 Hz or less, and expressed in terms of time, it has a period of 25 seconds or more. In the sense of looking at a cycle repeated about 10 times, a measurement time of 250 seconds, which is 10 times 25 seconds, is recommended, approximately 5 minutes. In addition, a measurement time of 5 minutes is appropriate to secure frequency domain resolution capable of distinguishing VLF, LF, and HF domains in Fast Fourier Transformation (FFT). Therefore, in a long-term analysis, such as 24-hour HRV, the time and frequency domain metrics are calculated by dividing into 5-minute segments.

Despite the recommendation of 5-minute measurement in the guidelines, many research papers and clinical applications published today use different measurement times such as 2-minute, 3-minute, 5-minute, 10-minute, and 15-minute measurement data depending on the research purpose and situation. there is. The guideline's recommendation for a 5-minute measurement is based on the 1996 draft "Heart Rate" by the Task Force of the European Society of Cardiology, the North American Society of Pacing, and Electrophysiology "Heart rate variability: standards of measurement, physiological interpretation, and clinical use".

In addition, with the recent development of mobile and wearable technologies, it has become possible to easily acquire HRV signals, and most devices require shorter measurement times.

In 2015, Byung-chae Lee et al. published in the Korean J Fam Pract., “Clinical usefulness through comparison of heart rate variability (HRV) performed for 3 minutes and 5 minutes” in the guidelines of the Korean Society of Family Medicine (Korean J Fam Pract.) There was no significant difference between the recommended 5-minute heart rate variability and 3-minute heart rate variability results, and it was said that if sufficient rest was taken before the test, the result value for 3 minutes could be regarded as 5-minute measurement data. That is, it can be seen that 3-minute measurement data can be used instead of 5-minute measurement data.

In 2017, Byeong-chae Lee et al., in “The Minimum Measurement Time Affecting the Reliability of Heart Rate Variation Test” published in the journal of the Korean Society for Clinical Health Promotion (Korean J Health Promot), measured data for 5 minutes into segments at 30-second intervals. Divided, the indicators of the 5-minute data and the indicators measured in a short period of 5 minutes or less were compared. In this study, reliable minimum measurement time with no statistically significant difference from the 5-minute data for each indicator was provided. When compared with the standard recommended 5-minute heart rate variability test results, the minimum measurement time without a statistically significant difference was 180 seconds for SDNN, 270 seconds for NN50, and 180 seconds for TP, LFn, and HFn for each variable, respectively. It was suggested that the minimum measurement time for heart rate variability test should be more than 180 seconds (3 minutes). Among the time domain indicators, indicators such as average heart rate and RMSSD can obtain reliable values even after 30 seconds of measurement, so most non-medical wearable devices use RMSSD as the main indicator. However, in order to evaluate the ability to regulate the autonomic nervous system, an index in the frequency domain, which requires a measurement time of at least 180 seconds, must be secured along with the index in the time domain. When the 5-minute (or 3-minute) data was viewed as a true value, the frequency domain data obtained through a measurement time of less than 3 minutes had a significant difference compared to the true value, so it could not be used clinically.

Therefore, since the measurement time cannot be further shortened according to statistical significance, the present invention proposes a method of predicting indicators of 5-minute or 3-minute data with data of a short measurement interval.

In the present invention, deep learning, an artificial intelligence technique that has been actively researched and utilized recently, is applied for index prediction. In addition, in the present invention, the data measured for 1 minute is used as input data of the learning model, and the data measured for 3 minutes is used by learning to predict the HRV index value.

There are many methods for predicting myocardial ischemia by applying heart rate variability to an artificial neural network, such as in Korean Patent Registration No. 10-1273652. , and using the 5-minute (or 3-minute) heart rate variability (HRV) output from the artificial neural network to detect the heart rate variability (HRV) index, no prior art related to the technology could be found.

An object to be achieved by the present invention is to use heart rate variability (HRV) measured in a short measurement period, that is, less than 3 minutes, to output a heart rate variability index, such as the 5-minute or 3-minute measurement of the standard recommendation. Inputs the heart rate variability measured for a short time of less than 3 minutes to the pre-learned artificial neural network, and outputs the heart rate variability index output from the artificial neural network as a 5-minute (or 3-minute) heart rate variability (HRV) index, A method and system for predicting a 5-minute or 3-minute heart rate variability index from heart rate variability data of a short measurement period are provided.

In order to solve the above problems, the system for predicting the heart rate variability index of 5 minutes or 3 minutes from the heart rate data of the measurement period of less than 3 minutes of the present invention, during the measurement period of less than 3 minutes, the heart rate data from the subject, the measurement period of less than 3 minutes As heart rate data, detecting, a heart rate measuring device; The heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device is applied to the HRV (heart rate variability) index prediction model of the previously trained artificial neural network using deep learning, and the HRV index output from the HRV index prediction model is 3 It includes; a calculation processing unit that outputs as a minute or five-minute HRV indicator.

The calculation processing unit generates first input data of the HRV indicator prediction model and second input data of the HRV indicator prediction model using the heart rate data of the measurement period of less than 3 minutes, wherein the first input data of the HRV indicator prediction model has a preset data size Within the data size, the calculation processing unit arranges the heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device in the center, and the front and back of the heart rate data of the measurement period of less than 3 minutes arranged in the center are less than 3 minutes. The data filled with the average value of the heart rate data in the measurement period is used as the first input data of the HRV index prediction model.

The calculation processing unit obtains a linear regression equation from the heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device, and using the linear regression equation, the 3-minute heart rate data or the 5-minute heart rate data is predicted, the 3-minute heart rate data or the prediction As the 5-minute heart rate data, predicted 3-minute heart rate FFT data or predicted 5-minute heart rate FFT data obtained by performing FFT (Fast Fourier Transform) on the predicted 3-minute heart rate data or the predicted 5-minute heart rate data, as an HRV index It is used as the second input data of the predictive model.

HRV indicators are mean RR interval (mRR), mean heart rate (mHR), SDNN (standard deviation of normal-to-normal interval), RMSSD (root mean square differences of successive R-R intervals), 5-minute total intensity (5-minute total power, TP), VLF (very low frequency), LF (power in low frequency, low frequency), HF (power in high frequency, high frequency), low frequency normalization It includes one or more of normalized intensity (normalized LF, nLF), normalized intensity of high frequency band (normalized HF, nHF), and LF/HF (LF/HF ratio, band intensity ratio of low frequency band to high frequency band).

The operation processing unit normalizes the first input data of the HRV index prediction model to a (0, 1) value of a float type, and inputs the normalized data to the HRV index prediction model.

The calculation processing unit normalizes the second input data of the HRV index prediction model to a (0, 1) value of a float type, and inputs the data to the HRV index prediction model.

The driving method of the system for predicting the heart rate variability index of 5 minutes or 3 minutes from the heart rate data of the measurement period of less than 3 minutes is to obtain the heart rate data from the subject during the measurement period of less than 3 minutes from the heart rate measuring device, and the heart rate of the measurement period of less than 3 minutes. heart rate measurement step, detecting as data; The calculation processing unit receives heart rate data for a measurement period of less than 3 minutes from the heart rate measuring device, and applies the received heart rate data for a measurement period of less than 3 minutes to a pre-trained heart rate variability (HRV) indicator prediction model using deep learning, HRV index prediction step of outputting the HRV index output from the HRV index prediction model as a 3-minute or 5-minute HRV index.

The HRV indicator prediction step includes the prediction model input data generation step, wherein the prediction model input data generation step includes the operation processing unit 200 receiving the heart rate data of the less than 3-minute measurement section measured in the heart rate measurement step, and receiving the received heart rate data. removing noise from heart rate data of a measurement period of less than 3 minutes and storing the noise-removed heart rate data of a measurement period of less than 3 minutes as real heart rate data of less than 3 minutes in a memory unit; An interpolation step of interpolating the data of the noise-removed part in the heart rate data of the measurement period of less than 3 minutes from which the noise was removed in the noise-removal step; a resampling step of, by the calculation processing unit, re-sampling the heart rate data of the measurement interval of less than 3 minutes interpolated in the interpolation step at a preset sampling rate; The calculation processing unit may include a linear regression equation extraction step of obtaining a linear regression equation from the heart rate data of a measurement period of less than 3 minutes resampled in the resampling step; a predicted heart rate data generation step of generating, as predicted heart rate data, the predicted 3-minute data or the predicted 5-minute data by using the linear regression equation obtained in the linear regression equation extraction step; The arithmetic processing unit performs an FFT (Fast Fourier Transform) on the predicted heart rate data generated in the predicted heart rate data generating step, and stores the FFT predicted heart rate data in a memory unit.

The HRV index prediction step further includes a preprocessing step, wherein the preprocessing step converts the actual heart rate data of the measurement period of less than 3 minutes, from which noise is removed in the noise removal step, to a preset data size, Within the size, the actual heart rate data of the measurement period of less than 3 minutes is placed in the center, and the front and back of the actual heart rate data of the measurement period of less than 3 minutes is filled with the average value of the actual heart rate data of the measurement period of less than 3 minutes. Resizing and normalizing the heart rate data of a measurement period of less than 3 minutes, taking the model first input data and normalizing the HRV index prediction model first input data to a (0, 1) value of a float type; The operation processing unit normalizes the predicted heart rate data FFTed in the FFT step as first input data of the HRV index prediction model to a float type (0, 1) value, normalized predictive data generation step; includes.

In the HRV indicator prediction step, the operation processing unit inputs the normalized HRV indicator prediction model first input data and the normalized HRV indicator prediction model second input data to the HRV indicator prediction model of the artificial neural network to execute the HRV indicator prediction model , HRV indicator prediction model driving step is further included.

The calculation processing unit receives 3-minute heart rate data or 5-minute heart rate data measured by the heart rate measuring device as HRV raw data for training of the HRV indicator prediction model of the artificial neural network, and from the received HRV raw data, 3 In a preset time unit of less than 3 minutes, heart rate data for less than 3 minutes is extracted as heart rate data for training of less than 3 minutes, and the extracted heart rate data for learning for less than 3 minutes is converted to a preset data size, but within the data size ,The heart rate data for learning less than 3 minutes is placed in the center, and the data filled with the average value of the actual heart rate data for the measurement period less than 3 minutes before and after the arranged heart rate data for learning less than 3 minutes is used as the first data for learning, and the heart rate data for less than 3 minutes A linear regression equation is obtained using the training heartbeat data, and 3-minute heartbeat data or 5-minute heartbeat data is detected as predicted heartbeat data using the linear regression equation, and FFT (Fast Fourier Transform) is performed on the predicted heartbeat data. is used as the second data for learning, and the HRV index is detected from the HRV raw data as the correct answer data for learning, and the first data for learning, the second data for learning, and the correct answer data for learning are used to calculate the HRV index of the artificial neural network Train a predictive model.

The present invention uses the heart rate variability (HRV) measured for a short measurement period, that is, less than 3 minutes, to output the same heart rate variability index as when measuring 5 minutes (or 3 minutes) of the standard recommendation. A short measurement that inputs heart rate variability measured for a short time of less than 3 minutes to the artificial neural network, and outputs the heart rate variability index output from the artificial neural network as a 5-minute (or 3-minute) heart rate variability (HRV) index. A method and system for predicting a 5-minute (or 3-minute) heart rate variability index from heart rate variability data of a section are provided.

In order to verify the reliability of the method of the present invention, deep learning, an artificial intelligence technique, is applied, the data measured for 1 minute is used as the input data of the learning model, and the HRV index value of the data measured for 3 minutes is learned to predict As a result of verifying whether there is a statistically significant difference between the application result of the artificial neural network learned in this way (i.e., the predicted value) and the actual value, it was confirmed that it can be used clinically because there is no significant difference. did. In other words, according to the results of previous studies, a measurement time of at least 3 minutes was required to obtain a clinically reliable HRV index, but in the present invention, it is possible through a measurement time of 1 minute through deep learning learning.

In this way, even HRV data measured in a short time can be used as highly accurate data in various medical devices and household appliances using HRV data. That is, it is possible to develop medical devices and household devices that do not require a long HRV measurement time.

1 is an explanatory diagram schematically illustrating the configuration of a system for predicting indicators of 5-minute or 3-minute heart rate variability data using heart rate variability data of a short measurement period according to the present invention.
Figure 2 is an explanatory diagram for explaining the input data form and learning label of the HRV indicator prediction model of the present invention.
3 is a schematic flowchart illustrating the process of generating learning data in the artificial neural network of the present invention.
4 is an explanatory diagram for explaining an example of a method for generating 1-minute data in the step of generating learning data of FIG. 3 (S200).
FIG. 5 is a flowchart illustrating the step (S200) of generating learning data of FIG. 3. Referring to FIG.
6 is an explanatory diagram explaining the configuration of data for artificial intelligence learning of the present invention.
7 is a flowchart schematically illustrating the artificial intelligence learning process of the present invention.
FIG. 8 is a flowchart illustrating the learning-related data preprocessing step 510 of FIG. 7 .
9 is a flowchart schematically illustrating a prediction process of predicting indicators of heart rate variability data of 5 minutes or 3 minutes using heart rate variability data of a short measurement period (less than 3 minutes) in the artificial intelligence of the present invention.
FIG. 10 is an explanatory diagram for explaining generation of short measurement intervals (1 minute data), which is input data for prediction in FIG. 9 .
11 is a flowchart illustrating a preprocessing step of input data for prediction in FIG. 9 .
12 is an explanatory diagram of an example of the configuration of an artificial neural network of the LF/HF prediction model of the LF/HF prediction unit of the present invention.
13 is an explanatory diagram of an example of the artificial neural network configuration of the average heart rate (mHR) prediction model of the average heart rate (mHR) predictor of the present invention.
14 is an explanatory diagram of an example of the artificial neural network configuration of the RMSSD prediction model of the RMSSD prediction unit of the present invention.
15 is an explanatory diagram of an example of the artificial neural network configuration of the SDNN prediction model of the SDNN prediction unit of the present invention.
16 is an explanatory diagram of an example of the artificial neural network configuration of the predictive model of other variables of the present invention.
17 is an example of the transition of loss values for SDNN learning epochs.
18 shows an example of a difference between an actual value and a predicted value of a TP indicator.
19 shows an example of a correlation between an actual value of an index of data measured for 3 minutes and a predicted value by deep learning, and a correlation between an actual value and 1 minute measurement data for TP and VLF.
20 shows an example of a correlation between an actual value of an indicator of data measured for 3 minutes and a predicted value by deep learning, and a correlation between an actual value and 1 minute measurement data for LF and HF.
21 shows an example of a correlation between an actual value of an index of data measured for 3 minutes and a predicted value by deep learning, and a correlation between an actual value and 1 minute measurement data for LFn and HFn.
22 shows an example of a correlation between an actual value of an index of data measured for 3 minutes and a predicted value by deep learning, and a correlation between an actual value and 1 minute measurement data for LF/HF and mHR.
23 shows an example of the correlation between the actual value of the indicator of the data measured for 3 minutes and the predicted value by deep learning, and the correlation between the actual value and the 1-minute measurement data for RMSSD and SDNN.
24 shows the degree of agreement between actual values and predicted values in a Bland-Altman plot for TP, LF, HF, and VLF.
25 shows the degree of agreement between actual values and predicted values in a Bland-Altman plot for LFn, HFn, LF/HF, and mHR.
26 shows the degree of agreement between the actual value and the predicted value in a Bland-Altman plot for RMSSD and SDNN.

Hereinafter, a method and system for predicting indicators of heart rate variability data of 5 minutes or 3 minutes using heart rate variability data of a short measurement period according to the present invention will be described with reference to the accompanying drawings.

1 is an explanatory diagram schematically illustrating the configuration of a system for predicting indicators of 5-minute or 3-minute heart rate variability data using heart rate variability data of a short measurement period according to the present invention.

The heart rate measuring device 100 measures heart rate variability for a short measurement period (less than 3 minutes) and transmits it to the calculation processing unit 200 . The heart rate measuring device 100 is a device for detecting a heart rate from electrocardiogram (ECG) or photoplethysmogram (PPG), and a commercially available heart rate variability tester or heart rate measuring device may be used. Here, the calculation processing unit 200 may be composed of a computer, a microprocessor, or the like.

The short measurement period is a measurement period of less than 3 minutes, and may be, for example, a measurement period of 1 minute.

The calculation processing unit 200 inputs the heart rate variability data received from the heart rate measuring device 100 to the pre-learned artificial neural network, and converts the HRV index output from the artificial neural network into a 5-minute (or 3-minute) HRV index. It is stored in the memory unit 300 or output through the output unit 500. The artificial neural network uses a HRV index prediction model using deep learning. Here, the memory unit 300 may be a database.

As HRV indicators, there are HRV time domain indicators and HRV frequency domain indicators. HRV time domain indicators include mean RR interval (mRR), mean heart rate (mHR), SDNN (standard deviation of normal-to-normal interval) ), RMSSD (root mean square differences of successive R-R intervals), and HRV frequency domain indicators include 5-minute total power (TP), VLF (very low frequency, very low frequency), LF (power in low frequency, low frequency band intensity), HF (power in high frequency, high frequency band intensity), normalized intensity of low frequency band (normalized LF, nLF), normalized intensity of high frequency band (normalized HF, nHF), LF/HF (LF/HF ratio, band intensity ratio of low frequency to high frequency) and the like. Since this is described in detail in Korean Registered Patent No. 10-0493714, a detailed description thereof will be omitted.

In an experiment to find out if there is a difference between the indicators (actual values) of the data measured for 3 minutes and the predicted value by the deep learning of the present invention, data was collected through a target group with evenly distributed age and gender, It was analyzed using the standard analysis method suggested by the guidelines. At this time, HRV data measured for 3 minutes of 43,504 cases collected by a heart rate variability tester (SA3000P, Medicore, Seoul, Korea) were used. A total of 34,885 data measured for 3 minutes were used for actual learning, except for 8,619 cases, such as data with abnormal heart rate, data with average heart rate less than 60, 100, and data in which the state changed significantly during the 3-minute measurement period. . In this experiment, the index of 3-minute HRV was predicted using the 1-minute data extracted from the heart rate variability tester.

Table 1 shows the HRV index used in the present invention.

Here, mRR can be calculated through mHR prediction, and LFn and HFn in the frequency domain are also calculated directly from the predicted LF and HF values, so they are excluded from learning.

In this experiment, HRV time series data was generated through peak detection in the ECG (or PPG) raw signal measured for 3 minutes, and then 1 minute data after 1 minute and 30 seconds was used as training data.

In order to process the input data of the HRV index prediction model of the artificial neural network, the calculation processing unit 200 stores the heart rate data of a short measurement period of less than 3 minutes received from the heart rate measuring device 100 in the memory unit 300 In the input data temporary storage unit (not shown), based on the heart rate data of the short measurement period of less than 3 minutes, in the front and rear parts, that is, in the short measurement period of less than 3 minutes Fill in the part except for the heart rate data in with the average value of the heart rate data of the short measurement period of less than 3 minutes. At this time, the size (number) of the input data of the HRV indicator prediction model is determined at the time of shipment from the factory or at the beginning of use by the user at the size (number) of data of 5 minutes (or 3 minutes). The data of the input data temporary storage unit filled in this way is input to the HRV indicator prediction model of the artificial neural network as input data of the HRV indicator prediction model.

Figure 2 is an explanatory diagram for explaining the input data form and learning label of the HRV indicator prediction model of the present invention.

2 shows the case of using 1-minute heartbeat data as input data. Since the number of bits of 1-minute data varies depending on the heart rate, the size of the input data is set to 300, and 1-minute real heartbeat ) data is placed in the center, and the remaining front and back parts are filled with the average value of 1 minute data to create 300 input data. This can be automatically processed in the calculation processing unit 200.

For the frequency domain index data, 256 pieces of data obtained by FFT-converting the extracted 1-minute data were used, and the index value of the data measured for 3 minutes was designated as a label for learning.

In the present invention, in predicting the HRV index of 5-minute or 3-minute data using deep learning, since the purpose of data is to reduce the measurement time, age and gender do not need to be considered.

As the learning model of the present invention, various models including or excluding general DNN (Deep Neural Network), CNN (Convolutional Neural Network), CNN1D, CNN2D, RNN and dropout are applied to each index to obtain mean absolute error (mean_absolute_error: MAE) The model with the smallest can be selected as the learning model for the indicator.

3 is a schematic flow chart explaining the learning data generation process in the artificial neural network of the present invention, and FIG. 4 is an explanatory diagram for explaining an example of a 1-minute data generation method in the step of generating learning data (S200) of FIG. 5 is a flowchart illustrating the step (S200) of generating learning data of FIG. 3.

As shown in FIG. 3, the calculation processing unit 200, as data stored in the memory unit 300 or data measured from the heart rate measuring device 100, 3-minute HRV raw data (or 5-minute HRV raw data) is received from the memory unit 300 or the heart rate measuring device 100 (S100), and from the received 3-minute HRV raw data (or 5-minute HRV raw data), training data for artificial intelligence learning of the artificial neural network is generated (S200), and at the same time, correct answer data for learning for artificial intelligence learning of the artificial neural network is generated (S300). The thus generated data for learning and answer data for learning are stored in the memory unit 300 as data for artificial intelligence learning (S350).

As the data for learning, first data for learning and second data for learning are provided.

For the first data for learning, the calculation processing unit receives 3-minute heart rate data or 5-minute heart rate data measured by the heart rate measuring device or read from the memory unit as HRV raw data, and from the received HRV raw data, Heart rate data for less than 3 minutes is extracted as heart rate data for learning for less than 3 minutes in a predetermined time unit of less than 3 minutes. The extracted heart rate data for training of less than 3 minutes is converted into a preset data size, within the data size, the heart rate data for training of less than 3 minutes is placed in the center, and the front and back of the arranged heart rate data for training of less than 3 minutes is 3 Data filled with the average value of the actual heart rate data in the measurement interval of less than a minute is used as first data for learning.

For the second data for learning, the calculation processing unit obtains a linear regression equation using heartbeat data for training of less than 3 minutes, and detects 3-minute heartbeat data or 5-minute heartbeat data as predicted heartbeat data using the linear regression equation, , data obtained by FFT (fast Fourier transform) on the predicted heart rate data is used as second data for learning.

For learning answer data, the arithmetic processing unit detects an HRV index from the HRV raw data and sets it as answer data for learning.

In order to input the first data for learning, the second data for learning, and the correct answer data for learning to the artificial neural network, these data are normalized to the (0, 1) values of the float type and input to the artificial neural network.

Hereinafter, a process of generating data for learning will be described in detail.

The 3-minute HRV raw data (or the 5-minute HRV raw data) is data obtained by detecting a heartbeat, ie, a peak, in an electrocardiogram.

The answer data for learning is data obtained by the calculation processing unit 200 detecting an HRV indicator from 3-minute HRV raw data (or 5-minute HRV raw data).

As shown in FIG. 5, the step of generating learning data (S200) is performed as follows.

As a 1-minute data extraction step for training data, the calculation processing unit 200 extracts 1-minute data from the received 3-minute HRV raw data (post peak detection data) (or 5-minute HRV raw data) (S210) . For example, as shown in FIG. 4, the calculation processing unit 200 cuts 3-minute HRV raw data (eg, 90,000) into 30-second data (eg, 15,000), and then combines them to obtain 1 minute data can be extracted.

As a noise removal step, the calculation processing unit 200 removes artifacts from the 1-minute data extracted in the 1-minute data extraction step for training data (S220), and converts the noise-removed 1-minute data into 1-minute actual data. Heartbeat data is stored in the memory unit 300 (S230).

As an interpolation step, the operation processing unit 200 interpolates the data of the part from which the noise is removed in the noise removal step (S230). Here, data interpolation may be an average value of 3-minute HRV raw data (or 5-minute HRV raw data), or may be interpolated with a value immediately before the removed part or a value immediately after the removed part.

In the resampling step, the calculation processing unit 200 samples the 1-minute data interpolated in the interpolation step again at a preset sampling rate (S240).

As a linear regression equation extraction step, the calculation processing unit 200 obtains a linear regression equation from the 1-minute data resampled in the resampling step (S250).

In the 3-minute data generation step, the calculation processing unit 200 uses the linear regression equation obtained in the linear regression equation extraction step to concatenate the 3-minute data (or 5-minute data), that is, the predicted 3-minute data (or predicted 5-minute data) is generated (S260).

In the FFT step, the calculation processing unit 200 performs FFT (Fast Fourier Transform) on the predicted 3-minute data (or 5-minute data) in the predicted 3-minute data generating step (S270), and the FFT 3-minute data (or 5-minute data) is stored in the memory unit 300 as learning data (S280).

6 is an explanatory diagram explaining the configuration of data for artificial intelligence learning of the present invention.

Data for artificial intelligence learning consists of 1-minute actual data (first data for learning, 3-minute FFT data (second data for learning), and correct answer data for learning).

7 is a flowchart schematically illustrating the artificial intelligence learning process of the present invention, and FIG. 8 is a flowchart illustrating the learning-related data preprocessing step 510 of FIG. 7 .

As shown in FIG. 7, the artificial intelligence learning process includes an artificial intelligence learning data input step (S505), a preprocessing step (S510), an artificial intelligence learning step (S550), and a learning result model storage step (S600).

In the step of inputting artificial intelligence learning data (S505), the calculation processing unit 200 inputs data for artificial intelligence learning to the artificial neural network.

As in FIG. 8, the preprocessing step (S510) is as follows.

In the sizing and normalized 1-minute real data generation step, the calculation processing unit 200 inputs the actual heart rate data of a short data section of less than 3 minutes (for example, 1-minute real heart rate data input in the artificial intelligence learning data input step S505). data) to a preset data size (eg 300), but the actual heart rate data (eg 1 minute actual data) is placed in the center, and the front and rear of the actual heart rate data placed is converted to the average value of 1 minute actual data. It is filled (S520), and the 1-minute data changed to have a constant size is normalized to the (0, 1) value of the float type (S525).

As a normalized learning data generation step, it proceeds simultaneously with the size adjustment and normalized 1-minute actual data generation step, but the operation processing unit 200 converts the training data (ie, 3-minute FFT data) into a float type (0, 1) value Normalize to (S530).

As a normalized learning correct answer data generation step, it proceeds simultaneously with the resizing and normalized 1-minute actual data generating step and the normalized learning data generating step, but the operation processing unit 200 converts the correct answer data for learning into float type (0, 1) Normalize with the value (S540).

By inputting the scaled and normalized 1-minute actual data, normalized training data, and normalized training answer data into the artificial neural network, artificial intelligence learning is executed to create a learning result model, an HRV index prediction model, and the HRV index generated in this way The prediction model is stored in the memory unit 300 .

In the HRV index prediction process of the present invention, for the input data to the HRV index prediction model, the calculation processing unit uses the heart rate data of the measurement period of less than 3 minutes to obtain HRV index prediction model first input data and HRV index prediction model second Generate input data.

The first input data of the HRV index prediction model has a preset data size, and the calculation processing unit arranges the heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device in the center within the data size, and It is data filled with the average value of the heart rate data of the measurement period of less than 3 minutes before and after the heart rate data of the measurement period of less than 3 minutes.

For the second input data of the HRV indicator prediction model, the calculation processing unit obtains a linear regression equation from the heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device, and uses the linear regression equation to obtain 3-minute heart rate data or 5-minute heart rate data. Data obtained by detecting data as predicted 3-minute heart rate data or predicted 5-minute heart rate data, and performing FFT (fast Fourier transform) on the predicted 3-minute heart rate data or predicted 5-minute heart rate data.

The first input data of the HRV indicator prediction model and the second input data of the HRV indicator prediction model are normalized to float type (0, 1) values and applied to the artificial neural network.

Hereinafter, the HRV indicator prediction process of the present invention will be described in detail.

9 is a flowchart schematically illustrating a prediction process of predicting indicators of 5-minute or 3-minute heart rate variability data with heart rate variability data of a short measurement period (less than 3 minutes) in the artificial intelligence of the present invention, and FIG. 10 is 9 is an explanatory diagram for explaining generation of short measurement intervals (1 minute data), which is input data for prediction in FIG. 9, and FIG.

As shown in FIG. 9, in the prediction process using the HRV index prediction model, the predictive model input data generation step (S705), the preprocessing step of the predictive model input data (S765), the HRV index prediction model driving step (S780), and the prediction result output Step S800 is included.

In the predictive model input data generating step (S705), the calculation processing unit 200 generates predicted 3-minute FFT data from the heart rate data for a short measurement period of less than 3 minutes from the heart rate measuring device 100.

As shown in FIG. 10, the predictive model input data generating step (S705) is obtained through the following process.

As a step of receiving heart rate data for a measurement period of less than 3 minutes, during a short measurement period of less than 3 minutes, the heart rate measuring device 100 measures the heart rate (S710), and the calculation processing unit 200 measures the heart rate measured from the heart rate measuring device 100. Heartbeat (post peak detection data) is received as heartbeat data (for example, 1 minute data) of a measurement period of less than 3 minutes, and is temporarily stored in the memory unit 300 (S720).

As a noise removal step, the calculation processing unit 200 removes artifacts from the heart rate data (for example, 1 minute data) received in the heart rate data reception step of the measurement period of less than 3 minutes (S725), and noise The removed heart rate data (eg, 1 minute data) of the measurement period of less than 3 minutes is used as actual heart rate data of the measurement period of less than 3 minutes, that is, first input data (eg, 1 minute actual heart rate data) of the HRV indicator prediction model. 300) (S730). Here, in the heart rate data of the measurement period of less than 3 minutes from which the noise is removed, “0” is input to the noise-removed part.

As an interpolation step, the calculation processing unit 200 interpolates the data of the part from which the noise is removed in the noise removal step (S735). Here, for data interpolation, linear interpolation may be applied, or an average value of actual heart rate data for a measurement period of less than 3 minutes may be used, or interpolation may be performed with a value immediately before the removed portion or immediately after the removed portion.

Preferably, linear interpolation is applied. In general, in the case of linear interpolation, an arbitrary point (x, y) located between two coordinate points (a, f (a)) and (b, f (b)) To estimate, apply the following formula:

In the resampling step, the calculation processing unit 200 resamples (resamples) the heart rate data of the measurement interval of less than 3 minutes interpolated in the interpolation step at a preset sampling rate and stores it in the memory unit 300 (S740). . Here, the preset sampling rate may be 2Hz. In this case, if the heart rate data is sampled at 500 Hz, the interpolated heart rate data for a measurement period of less than 3 minutes is sampled at 500 Hz and resampled at 2 Hz. That is, resampling from 500 data per second to 2 data per second.

As a linear regression equation extraction step, the calculation processing unit 200 obtains a linear regression equation from heart rate data of a measurement period of less than 3 minutes resampled in the resampling step (S745).

In the 3-minute data generation step, the calculation processing unit 200 uses the linear regression equation obtained in the linear regression equation extraction step to concatenate the 3-minute data (or 5-minute data), that is, the predicted 3-minute data (or predicted 5-minute data) is generated (S750).

In the FFT step, the calculation processing unit 200 performs an FFT (fast Fourier transform) on the predicted 3-minute data (or 5-minute data) in the predicted 3-minute data generating step (S755), and the FFT 3-minute data (or 5-minute data) is stored in the memory unit 300 as predicted 3-minute FFT data, that is, second input data of the HRV indicator prediction model (S760).

As in FIG. 11, the preprocessing step (S710) is as follows.

In the step of resizing and normalizing the heart rate data of the measurement period of less than 3 minutes, the operation processing unit 200 measures less than 3 minutes generated in the prediction model input data generation step (S705) (ie, detected in the noise removal step). The actual heart rate data of the section (eg 1 minute actual data) is converted to a preset data size (eg 300), but the actual heart rate data of the measurement section less than 3 minutes (eg 1 minute actual data) is placed in the center and placed The front and back of the actual heart rate data obtained are filled with the average value of the actual heart rate data of the less than 3-minute measurement period (S770), and the actual heart rate data of the less than 3-minute measurement period changed to have a constant size (ie, HRV index prediction model first input data) is normalized by applying the Min-Max Normalization technique, and as a result, it is normalized to (0, 1) values of the float type (S775).

The Min-Max Normalization technique, which is well known, is obtained by the following formula.

Normalized value =　(Value to be normalized - Minimum value in range) / (Maximum value in range - MIN value in range)

Restore to original value =　(normalized value × (maximum value in range - minimum value in range))

Minimum value in range +　

In other words, using the minimum value of the actual heart rate data range of the measurement period of less than 3 minutes, which is changed to a constant size, and the maximum value of the actual heart rate data range of the measurement period of less than 3 minutes, which is changed to a constant size, the actual measurement period of less than 3 minutes, which is changed to a constant size Heart rate data is normalized by the above formula.

As a step of generating normalized data for prediction, it is performed simultaneously with the step of resizing and normalizing the heart rate data of the measurement period of less than 3 minutes, but the operation processing unit 200, Model second input data) is normalized by applying a Min-Max Normalization technique, and as a result, it is normalized to a float type (0, 1) value (S777).

HRV indicator prediction model execution step, scaled and normalized measurement interval actual data (HRV indicator prediction model first input data), normalized and predicted 3-minute FFT data (HRV indicator prediction model second input data) is input to the HRV index prediction model of the artificial neural network to execute the HRV index prediction model (S780).

The HRV indicator prediction model is a LF/HF prediction unit (not shown), an average heart rate (mHR) prediction unit (not shown), an RMSSD prediction unit (not shown), an SDNN prediction unit (not shown), and a TP of the calculation processing unit 200. Prediction unit (not shown), LF prediction unit (not shown), HF prediction unit (not shown), VLF prediction unit (not shown), normalized LF (LFnorm) prediction unit (not shown), normalized HF (HFnorm) Including a predictor (not shown), from the predicted 3-minute FFT data (HRV index prediction model second input data), 5-minute or 3-minute heart rate variability index LF/HF, average heart rate (mHR), RMSSD, Find SDNN, TP, LF, HF, VLF, normalized LF (LFnorm), and normalized HF (HFnorm). At this time, one or more of LF/HF, average heart rate (mHR), RMSSD, SDNN, TP, LF, HF, VLF, normalized LF (LFnorm), and normalized HF (HFnorm) is obtained according to the user's settings.

The HRV index output from the HRV index prediction model of the artificial neural network is output to the output unit 500 or stored in the memory unit 300.

12 is an explanatory diagram of an example of the configuration of an artificial neural network of the LF/HF prediction model of the LF/HF prediction unit of the present invention.

As a convolution processing step for LF/HF, the calculation processing unit 200 normalizes the normalized predictive data generation step (S777) with the LF/HF prediction unit (LF/HF prediction model) of the calculation processing unit 200 If the predicted 3-minute FFT data (HRV indicator prediction model second input data) is input as input data in the form of 16x16x1 (S1110), the LF/HF prediction unit convolutions the input data with a size 32 kernel, then It is processed as a relu (Rectified Linear Unit) activation function (S1120).

As a fully connected layer processing step with 256 kernels for LF / HF, the LF / HF prediction unit outputs data of the convolution processing step (S1120) (ie, input data processed by the relu activation function after convolution processing with 32 kernels) ) is processed into a fully connected layer (dense layer) with a 　256 kernel, and then the relu activation function is processed (S1130).

Although DepthwiseConv2D is used as a layer function in the fully connected layer processing step (S1130) with 256 kernels in FIG. 12, it should be noted that this is not intended to limit the present invention. For this purpose, it is a layer function provided by Tensorflow 2.x, an end-to-end open source platform for machine learning. Since general convolution filters are affected by all channels of the input, it is impossible to completely extract the spatial features of a specific channel. Do. That is, it is unavoidable that information of other channels is involved. However, depthwise convolution used filters performed only for each single channel.

As a fully connected layer processing step to 1 kernel for LF / HF, the LF / HF prediction unit outputs the output data of the fully connected layer processing step (S1130) to 256 kernels for LF / HF (i.e. Fully conndeted with 256 kernels After layer processing, fully connected layer processing is performed on the relu activation function-processed input data) with 1 kernel to output one LF/HF prediction value (S1140).

13 is an explanatory diagram of an example of the artificial neural network configuration of the average heart rate (mHR) prediction model of the average heart rate (mHR) predictor of the present invention.

As a fully connected layer processing step to 256 kernels for the average heart rate, the calculation processing unit 200, the average heart rate (mHR) prediction unit (average heart rate (mHR) prediction model) of the calculation processing unit 200, the noise removal step ( S725, S730), the scaled and normalized measurement section actual data (HRV index prediction model first input data) output in 300 sizes is input (S1150), and the average heart rate (mHR) The prediction unit processes this input data into a fully connected layer (dense layer) with a 　256 kernel, and then processes the relu activation function (S1160).

As a fully connected layer processing step with 1 kernel for average heart rate, the average heart rate (mHR) prediction unit outputs the output data of the fully connected layer processing step (S1150) with 256 kernels for average heart rate (i.e., Fully conndeted with 256 kernels). After layer processing, fully connected layer processing is performed on the relu activation function-processed input data) with 1 kernel to output one average heart rate prediction value (S1170).

14 is an explanatory diagram of an example of the artificial neural network configuration of the RMSSD prediction model of the RMSSD prediction unit of the present invention.

As a 4-dimensional data construction step, the calculation processing unit 200, the RMSSD prediction unit (RMSSD prediction model) of the calculation processing unit 200, output from the noise removal step (S725, S730), scaled and normalized less than 3 minutes The actual data of the measurement period (the first input data of the HRV indicator prediction model) is input as input data of size 300 (S1210), and a reshape function (function for changing the dimensions of the array) is used to facilitate convolution processing. The shape of the input data is configured in 4 dimensions by using (S1220).

As a convolutional multiplication step for RMSSD, the RMSSD prediction unit performs convolutional multiplication of the data output in the 4-dimensional data construction step (ie, 4-dimensional input data) with 32 kernels of size 3, and then processes it with a relu activation function ( S1230).

As an average pooling step, the RMSSD prediction unit converts the output data (that is, data processed by the relu activation function after convolutional processing with 3 size 32 kernels) in the convolution processing step (S1230) for RMSSD to the reshape function. After reconstructing using (S1240), an average pooling process is performed (S1250). That is, through the GlobalAveragePooling1D function provided by Tensorflow, 1-dimensional average pooling processing is performed.

As a fully connected layer processing step with 32 kernels for RMSSD, the RMSSD prediction unit converts the output data (i.e., average pooled data for one dimension) of the average pooling step (S1240, S1250) into 32 kernels. A fully connected layer is processed (S1260).

As a fully-connected layer processing step with 1 kernel for RMSSD, the RMSSD predictor outputs the output data of the fully-connected layer processing step (S1260) with 32 kernels for RMSSD (that is, fully-connected layer processed data with 32 kernels). A fully connected layer is processed with 1 kernel to output one RMSSD predicted value (S1270).

15 is an explanatory diagram of an example of the artificial neural network configuration of the SDNN prediction model of the SDNN prediction unit of the present invention.

As a convolution processing step for SDNN, the calculation processing unit 200, to the SDNN prediction unit (SDNN prediction model) of the calculation processing unit 200, the scaled and normalized 3 output from the noise removal steps (S725, S730) When the actual data (HRV indicator prediction model first input data) of the measurement period of less than a minute is input as input data in the form of 15x20x1 (S1310), the SDNN prediction unit convolutions the input data with a 3x3 size 32 kernel and uses the relu activation function It is processed as (S1320).

As a fully connected layer processing step with 256 kernels for SDNN, the SDNN prediction unit outputs data of the convolution processing step (S1320) for SDNN (ie, input data processed by the relu activation function after convolution processing with 32 kernels) After processing 　256 kernel as a fully connected layer (dense layer), the relu activation function is processed (S1330).

As a fully connected layer processing step with 1 kernel for SDNN, the SDNN prediction unit activates relu after output data of the fully connected layer processing step (S1330) with 256 kernels for SDNN (ie, fully connected layer processing with 256 kernels) Function-processed input data) is fully connected layer processed with 1 kernel to output one SDNN prediction value (S1340).

16 is an explanatory diagram of an example of the artificial neural network configuration of the predictive model of other variables of the present invention.

That is, FIG. 16 shows other variable predictors (ie, TP predictor, LF predictor, HF predictor, VLF predictor, normalized LF (LFnorm) predictor, normalized HF (HFnorm) predictor) of the present invention. , a description of artificial neural network configurations applied to other variable prediction models (i.e., TP prediction model, LF prediction model, HF prediction model, VLF prediction model, normalized LF (LFnorm) prediction model, and normalized HF (HFnorm) prediction model). Doida,

As a fully connected layer processing step to 256 kernels for the other variable prediction model, the operation processing unit 200 generates normalized prediction data with the other variable prediction model of the other variable prediction unit of the operation processing unit 200 (S777) Normalized and predicted 3-minute FFT data (HRV index prediction model second input data) is input as 256 size input data (S1510), and the other prediction unit converts this input data into a fully connected layer (with a 256 kernel) After processing with a fully conndeted layer, dense layer), a relu activation function is processed (S1520).

As a fully connected layer processing step to 1 kernel for the other variable prediction model, the other prediction unit outputs the output data of the fully connected layer processing step (S1520) to 256 kernels for the other variable prediction model (i.e., Fully conndeted with 256 kernels). After layer processing, fully connected layer processing is performed on the relu activation function-processed input data) with 1 kernel to output one predicted value (S1530).

Here, the other variable is one of TP, LF, HF, VLF, normalized LF (LFnorm), and normalized HF (HFnorm), and the other predictor is a predictor for predicting other variables. The TP predictor, the LF predictor, and the HF It is one of a prediction unit, a VLF prediction unit, a normalized LF (LFnorm) prediction unit, and a normalized HF (HFnorm) prediction unit, and the other prediction model is a model for predicting other variables, TP prediction model, LF prediction model, HF It is one of a prediction model, a VLF prediction model, a normalized LF (LFnorm) prediction model, and a normalized HF (HFnorm) prediction model.

For example, if the other variable to be predicted is TP, the other predictor is a TP predictor, and the other predictor model becomes a TP predictor model, resulting in the following.

In the fully connected layer processing step to the 256 kernel for the other variable prediction model (ie, the TP prediction model), the calculation processing unit 200, to the TP prediction unit of the calculation processing unit 200, receives the HRV index prediction model second input data , 256-sized input data, and the TP prediction unit processes this input data into a fully connected layer (dense layer) with a 256 kernel and then processes the relu activation function (S1520).

As a fully connected layer processing step to 1 kernel for other variable prediction models (ie, TP prediction model), the TP prediction unit outputs the output data of the fully connected layer processing step (S1520) to 256 kernels for other predictive models (ie, , After fully conndeted layer processing with 256 kernels, input data processed with relu activation function) is fully conndeted layer processed with 1 kernel to output one TP prediction value (S1530).

Of course, the same method will be applied when the variable is LF, HF, VLF, normalized LF (LFnorm), or normalized HF (HFnorm).

In the present invention, in an experiment to find out if there is a difference between the indicators (actual values) of the data measured for 3 minutes and the predicted value by the deep learning of the present invention, the training set is 60% of the total data 20,931 data were randomly selected for the validation set, 10,465 data (30%), and 3,489 data (10%) for the test set. In the case of SDNN, CNN2D is used as the learning model, Huber_loss is used as the loss function, Adam (adaptive moment assessment, Adam) is used as the optimizer, the learning rate is 1e-05, and the batch size is 5000, epoch is learned with 14,150 times. In addition, the transition of the loss value for the SDNN learning epoch at this time is shown in FIG. 17.

Table 2 shows an example of a learning model and hyperparameters for each frequency domain predictive index. That is, it shows the learning model and hyperparameters for each frequency domain predictive index in the experiment.

Table 3 shows an example of a learning model and hyperparameters for each predictive index in the time domain. That is, it shows the learning model and hyperparameters for each time-domain predictive index in the experiment.

In this experiment, a model with the smallest mean absolute error (MAE) was selected by applying various models and parameters to select a learning model and parameters optimized for each indicator.

Table 4 is an example of prediction results for each index. That is, it represents the minimum MAE value of each index in the above experimental results.

18 shows an example of the difference between the actual value (blue line) and the predicted value (orange line) of the TP indicator. The difference between the actual value (blue line) and predicted value (orange line) of the TP index in the above experimental results is shown.

In the statistical processing of the experimental results, the data indicators measured for 3 minutes are set as actual values (reference values) and the comparison with the predicted values of the indicators of the 3-minute data from the data measured for 1 minute through machine learning is an independent sample. A t test (independent-sample t-test) and Pearson's correlation coefficient were used. In addition, the degree of agreement between the actual value and the predicted value was verified using Bland-Altman plots based on the actual value. In all statistics, the level of significance was 5% or less as a statistically significant difference. In the independent sample t test, a significance level of 5% or higher was regarded as no significant difference between the actual value and the predicted value, and eventually the same result value was seen. All statistics were performed using SPSS19 and Excel descriptive statistics packages.

First, in the independent sample t test result, an independent sample t test (independent sample t-test ) The results are shown in Table 5.

Table 5 shows an example of independent sample t-test results for actual and predicted values. That is, it shows the independent sample t-test results for the actual value and the predicted value in the above experiment.

As a result of comparing the indicators of the data measured for 1 minute with the actual values, there was no significant difference between mHR and RMSSD, and there was a statistically significant difference from the actual data in all indicators (P<0.05). This result was confirmed to be consistent with the results of previous studies.

The predicted value by deep learning did not show a statistically significant difference between the actual value and all indicators used in this study (P > 0.05). Therefore, in the present invention, the predicted value predicted by deep learning using data measured for 1 minute can be used as an indicator of data measured for 3 minutes.

Second, in the Pearson correlation analysis results, as shown in FIGS. 19 to 23, data measured for 3 minutes for TP, VLF, LF, HF, LFn, HFn, LF / HF, mHR, RMSSD and SDNN The correlation between the indicator (actual value) of , the predicted value by deep learning, and the actual value and 1-minute measurement data was shown. That is, it was confirmed that the predicted values by deep learning in all indicators were very similar to the actual values (all indicators P-value < .0000). As shown in FIGS. 19 to 23, the correlation coefficient with the predicted value was much higher than the correlation coefficient with the index value of the data measured for 1 minute, rather than the actual value and the predicted value. TP is 0.4727 to 0.9307, LF is 0.4421 to 0.9773, HF 0.8147, 0.9929 at 0.8147, 0.6979 to 0.2433, 0.5177 for LFN, 0.9553 to 0.5167, 0.9579 to 0.9579, and LF/HF from 0.4345 to 0.9099 and the correlation increased.

In the above experimental results, it was found that the predicted value by deep learning had no statistically significant difference and a very high correlation with the actual 3-minute data through the independent sample t test result and the Pearson's correlation analysis result.

Looking at the degree of agreement between the actual value and the predicted value by deep learning in the above experiment, the degree of agreement between the actual value and the predicted value by the Bland-Altman plot is shown in FIGS. 24 to 26. 24 to 26, for TP, VLF, LF, HF, LFn, HFn, LF / HF, mHR, RMSSD and SDNN, the tolerance and limit of the Bland Altman plot for the actual value and the predicted value of agreement: The ratio of data within LOA) is shown together. is the ratio

As described above, the present invention has been described by the limited embodiments and drawings, but the present invention is not limited to the above embodiments, and those skilled in the art in the field to which the present invention belongs can make various modifications and transformation is possible Therefore, the spirit of the present invention should be grasped only by the claims described below, and all equivalent or equivalent modifications thereof will be said to belong to the scope of the spirit of the present invention.

100: heart rate measuring device
200: calculation processing unit
300: memory unit

Claims (18)

During a measurement period of less than 3 minutes, a heart rate measuring device that detects heart rate data from the subject as heart rate data of a measurement period of less than 3 minutes;
The heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device is applied to the HRV (heart rate variability) index prediction model of the previously trained artificial neural network using deep learning, and the HRV index output from the HRV index prediction model is 3 An arithmetic processing unit that outputs as an HRV index of minutes or 5 minutes;
A system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that it comprises a. According to claim 1,
The calculation processing unit generates first input data of the HRV index prediction model and second input data of the HRV index prediction model using the heart rate data of the measurement period of less than 3 minutes,
The HRV indicator prediction model first input data has a preset data size,
Within the data size, the calculation processing unit arranges the heart rate data of the less than 3-minute measurement interval received from the heart rate measuring device in the center, and the front and back of the heart rate data of the less than 3-minute measurement interval disposed in the center, the measurement interval of less than 3 minutes A system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that the data filled with the average value of the heart rate data is used as the first input data of the HRV indicator prediction model. According to claim 2,
The calculation processing unit obtains a linear regression equation from the heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device, and using the linear regression equation, the 3-minute heart rate data or the 5-minute heart rate data is predicted, the 3-minute heart rate data or the prediction As the 5-minute heart rate data, predicted 3-minute heart rate FFT data or predicted 5-minute heart rate FFT data obtained by performing FFT (Fast Fourier Transform) on the predicted 3-minute heart rate data or the predicted 5-minute heart rate data, as an HRV index A system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that the prediction model is used as the second input data. According to claim 1,
HRV indicators are mean RR interval (mRR), mean heart rate (mHR), SDNN (standard deviation of normal-to-normal interval), RMSSD (root mean square differences of successive RR intervals), 5-minute total intensity (5-minute total power, TP), VLF (very low frequency), LF (power in low frequency, low frequency), HF (power in high frequency, high frequency), low frequency normalization Characterized in that it includes one or more of normalized intensity (normalized LF, nLF), normalized intensity of high frequency band (normalized HF, nHF), LF / HF (LF / HF ratio, band intensity ratio of low frequency band and high frequency band) , A system that predicts a 5-minute or 3-minute heart rate variability index from heart rate data of a measurement period of less than 3 minutes. According to claim 2,
The calculation processing unit normalizes the first input data of the HRV index prediction model to a (0, 1) value of a float type, and inputs the data to the HRV index prediction model. A system that predicts a 5-minute or 3-minute heart rate variability index. According to claim 3,
The calculation processing unit normalizes the second input data of the HRV index prediction model to a (0, 1) value of a float type, and inputs the data to the HRV index prediction model. A system that predicts a 5- or 3-minute heart rate variability index from a heart rate measurement step of detecting heart rate data from the subject as heart rate data for a measurement period of less than 3 minutes from a heart rate measuring device during a measurement period of less than 3 minutes;
The calculation processing unit receives heart rate data for a measurement period of less than 3 minutes from the heart rate measuring device, and applies the received heart rate data for a measurement period of less than 3 minutes to a pre-trained heart rate variability (HRV) indicator prediction model using deep learning, HRV index prediction step of outputting the HRV index output from the HRV index prediction model as a 3-minute or 5-minute HRV index;
A method of driving a system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that it comprises a. The method of claim 7, wherein the HRV index prediction step,
The calculation processing unit is configured to generate first input data of the HRV index prediction model and second input data of the HRV index prediction model using the heart rate data of the measurement period of less than 3 minutes,
The HRV indicator prediction model first input data has a preset data size,
For the first input data of the HRV indicator prediction model, within the data size, the heart rate data received from the heart rate measuring device is centered, and the heart rate data of the measurement period of less than 3 minutes is centered. A method of driving a system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that the data filled with the average value of the heart rate data of the measurement period of less than 3 minutes. According to claim 7,
HRV indicators are mean RR interval (mRR), mean heart rate (mHR), SDNN (standard deviation of normal-to-normal interval), RMSSD (root mean square differences of successive RR intervals), 5-minute total intensity (5-minute total power, TP), VLF (very low frequency), LF (power in low frequency, low frequency), HF (power in high frequency, high frequency), low frequency normalization Characterized in that it includes one or more of normalized intensity (normalized LF, nLF), normalized intensity of high frequency band (normalized HF, nHF), LF / HF (LF / HF ratio, band intensity ratio of low frequency band and high frequency band) , A method of driving a system that predicts a 5-minute or 3-minute heart rate variability index from heart rate data of a measurement period of less than 3 minutes. According to claim 8,
The second input data of the HRV index prediction model is
The calculation processing unit obtains a linear regression equation from the heart rate data of the measurement period of less than 3 minutes received from the heart rate measuring device, and using the linear regression equation, the 3-minute heart rate data or the 5-minute heart rate data is predicted, the 3-minute heart rate data or the prediction As 5-minute heart rate data, it is characterized in that it is data obtained by detecting and performing FFT (Fast Fourier Transform) on predicted 3-minute heart rate data or predicted 5-minute heart rate data. A method of operating a system that predicts heart rate variability index per minute. According to claim 8,
The calculation processing unit normalizes the first input data of the HRV index prediction model to a (0, 1) value of a float type, and inputs the data to the HRV index prediction model. A method of driving a system that predicts a 5-minute or 3-minute heart rate variability index. According to claim 8,
The calculation processing unit normalizes the second input data of the HRV index prediction model to a (0, 1) value of a float type, and inputs the data to the HRV index prediction model. A method of driving a system that predicts a 5-minute or 3-minute heart rate variability index from According to claim 7,
The HRV index prediction step includes the step of generating input data for the prediction model,
The step of generating the predictive model input data is,
The calculation processing unit 200 receives the heart rate data of the measurement period of less than 3 minutes measured in the heart rate measurement step, removes noise from the received heart rate data of the measurement period of less than 3 minutes, and removes noise from the heart rate data of the measurement period of less than 3 minutes. a noise removal step of storing in a memory unit as real heartbeat data for a measurement period of less than 3 minutes;
An interpolation step of interpolating the data of the noise-removed part in the heart rate data of the measurement period of less than 3 minutes from which the noise was removed in the noise-removal step;
a resampling step of, by the calculation processing unit, re-sampling the heart rate data of the measurement interval of less than 3 minutes interpolated in the interpolation step at a preset sampling rate;
The calculation processing unit may include a linear regression equation extraction step of obtaining a linear regression equation from the heart rate data of a measurement period of less than 3 minutes resampled in the resampling step;
a predicted heart rate data generation step of generating, as predicted heart rate data, the predicted 3-minute data or the predicted 5-minute data by using the linear regression equation obtained in the linear regression equation extraction step;
an FFT step of, in the calculation processing unit, performing FFT (Fast Fourier Transform) on the predicted heart rate data generated in the predicted heart rate data generating step, and storing the FFT predicted heart rate data in a memory unit;
A method of driving a system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that it comprises a. According to claim 13,
The HRV index prediction step further includes a preprocessing step,
The preprocessing step is
The calculation processing unit converts the actual heart rate data of the measurement period of less than 3 minutes from which noise is removed in the noise removal step to a preset data size,
Within the data size, the actual heart rate data of the measurement period of less than 3 minutes is placed in the center, and the front and back of the actual heart rate data of the measurement period of less than 3 minutes is filled with the average value of the actual heart rate data of the measurement period of less than 3 minutes. HRV As the index prediction model first input data,
sizing and normalizing heart rate data of a measurement period of less than 3 minutes, normalizing first input data of the HRV indicator prediction model to a (0, 1) value of a float type;
a normalized predictive data generating step of normalizing, by the calculation processing unit, the predicted heart rate data FFTed in the FFT step to (0, 1) values of a float type as first input data of the HRV indicator prediction model;
A method of driving a system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that it comprises a. The method of claim 14, wherein the HRV index prediction step,
The operation processing unit inputs the normalized HRV indicator prediction model first input data and the normalized HRV indicator prediction model second input data to the HRV indicator prediction model of the artificial neural network to execute the HRV indicator prediction model, driving the HRV indicator prediction model A method of driving a system for predicting a heart rate variability index of 5 minutes or 3 minutes from heart rate data of a measurement period of less than 3 minutes, characterized in that it further comprises the step. The method of any one of claims 7 to 15,
The calculation processing unit receives 3-minute heart rate data or 5-minute heart rate data measured by the heart rate measuring device as HRV raw data for training of the HRV indicator prediction model of the artificial neural network, and from the received HRV raw data, 3 In a preset time unit of less than 3 minutes, extract heart rate data for less than 3 minutes as training heart rate data for less than 3 minutes,
Convert the extracted heart rate data for training of less than 3 minutes to a preset data size,
Within the data size, the heart rate data for training of less than 3 minutes is placed in the center, and data filled with the average value of the actual heart rate data of the measurement period of less than 3 minutes is used as the first data for learning. do,
A linear regression equation is obtained using heart rate data for training of less than 3 minutes, using the linear regression equation, 3-minute heart rate data or 5-minute heart rate data is detected as predicted heart rate data, and the predicted heart rate data is FFT (fast Fourier transform) is used as second data for learning,
Detect HRV index from HRV raw data and use it as correct answer data for learning,
Characterized in that the HRV indicator prediction model of the artificial neural network is trained using the first data for learning, the second data for learning, and the correct answer data for learning, a 5-minute or 3-minute heart rate variability index from the heart rate data of a measurement period of less than 3 minutes How to drive the predictive system. The method according to any one of claims 1 to 6, wherein the calculation processing unit,
The calculation processing unit receives 3-minute heart rate data or 5-minute heart rate data measured by the heart rate measuring device as HRV raw data for training of the HRV indicator prediction model of the artificial neural network, and from the received HRV raw data, 3 In a preset time unit of less than 3 minutes, extract heart rate data for less than 3 minutes as training heart rate data for less than 3 minutes,
Convert the extracted heart rate data for training of less than 3 minutes to a preset data size,
Within the data size, the heart rate data for training of less than 3 minutes is placed in the center, and data filled with the average value of the actual heart rate data of the measurement period of less than 3 minutes is used as the first data for learning. do,
A linear regression equation is obtained using heart rate data for training of less than 3 minutes, using the linear regression equation, 3-minute heart rate data or 5-minute heart rate data is detected as predicted heart rate data, and the predicted heart rate data is FFT (fast Fourier transform) is used as second data for learning,
Detect HRV index from HRV raw data and use it as correct answer data for learning,
Characterized in that the HRV indicator prediction model of the artificial neural network is trained using the first data for learning, the second data for learning, and the correct answer data for learning, a 5-minute or 3-minute heart rate variability index from the heart rate data of a measurement period of less than 3 minutes predictive system. A computer in which a program for implementing a method of driving a system for predicting a 5-minute or 3-minute heart rate variability indicator according to any one of claims 7 to 15 from heart rate data of a measurement period of less than 3 minutes by computer is recorded thereon. readable recording medium.

Images