KR102493879B1 - Method for analyzing parameter of photoplethysmography using deep neural network filters - Google Patents

Abstract

A receiving step of receiving PPG data measured from a PPG measurement sensor; a differential data generation step of generating differential data obtained by differentiating the PPG data received in the receiving step; a signal processing step of generating input data by pre-processing the PPG data and differential data generated corresponding thereto for each predetermined analysis section; Through a predefined analysis algorithm, while learning a plurality of PPG data and learning data in which a label for a target singularity in differential data corresponding to each PPG data is stored, using an optimized analysis algorithm, the signal is processed in the signal processing step. A detection step of detecting a target singularity within the input data of; and a determination step of deriving a parameter from the target singularity detected in the detection step. Including, wherein the analysis algorithm is a DNN (Deep Neural Network) filter composed of at least two or more hidden layers between an input layer and an output layer Optical using a DNN filter, characterized in that A volume pulse wave (PPG) parameter analysis method is provided.

Description

Photoplethysmogram (PPG) parameter analysis method using DNN filter {METHOD FOR ANALYZING PARAMETER OF PHOTOPLETHYSMOGRAPHY USING DEEP NEURAL NETWORK FILTERS}

The present invention relates to a method for analyzing photoplethysmogram parameters using a DNN filter.

Photoplethysmography (PPG) is a signal that measures the change in blood volume in blood vessels according to heartbeat using light absorption, reflection, and scattering, and is one of the most frequently measured biosignals along with electrocardiogram (ECG). Parameters such as heart rate (HR), heart rate variability (HRV), and oxygen saturation (SpO ₂ ) can be measured by monitoring the waveform variation period of the PPG signal sensed in the photoplethysmogram. At this time, the measured parameters are applied to various diagnoses such as arrhythmia, blood pressure change, heart disease, and autonomic nervous system change.

That is, the photoplethysmogram can provide a diagnostic parameter for abnormal blood flow in blood vessels. When measuring biosignals through the photoplethysmography sensor, noise generated by heartbeat and motion noise generated by the subject's movement etc., it was difficult to accurately specify the locations of target singularities for deriving parameters (eg, HR), and it was difficult to implement an automated algorithm due to large differences in individual characteristics of subjects.

Accordingly, prior arts 1 to 3 have proposed various analysis methods for improving the accuracy of analysis of the photoplethysmogram parameter.

Prior Art 1 (Non-Patent Document 001 among the following prior art documents) has proposed an algorithm for extracting target singular points such as the maximum value and inflection point of the PPG signal by first and second differentiation of the PPG signal, but the following problems ensued. When performing the second derivative, it was greatly affected by motion noise and was mistaken for an erroneous heart rate. The second derivative is mainly used because it is easy to create an automated algorithm. Problems with the accuracy of the HR, such as not detecting the heartbeat when the difference is large and the cardiac output temporarily decreases due to arrhythmia, have been continuously raised. In addition, it was difficult to distinguish the change in the waveform caused by the heartbeat and the change in the waveform caused by the reflected wave in the peripheral blood vessels.

Prior Art 2 (Patent Documents 001 to 002 among the following prior art documents) has proposed a technique for measuring parameters through frequency analysis of the entire PPG waveform measured for a certain period of time, but a lot of calculation time is required for frequency analysis and fast It was difficult to provide measurement results. In addition, when applied to HR measurement among parameters, it is difficult to quickly detect changes in heart rate, and in the case of arrhythmia patients, it is difficult to accurately measure heart rate.

In prior art 3 (Non-Patent Document 002 among the prior art documents below), an optical pulse wave analysis method using an image recognition artificial intelligence technology such as CNN (CONVEOLUTION NEURAL NETWORK) was proposed, but the following problems were accompanied. A very large amount of learning data is required, expensive resources such as high-performance computers for machine learning are required, real-time analysis is difficult because a lot of input data is used, and the accuracy of heart rate measured by analyzing PPG is not objectively verified. There have been reports of cases where errors have been observed.

Republic of Korea Patent Registration No. 10-1409800 Republic of Korea Patent Publication No. 10-2012-0021098

K. Takazawa, N. Tanaka, M. Fujita, O. Matsuoka, T. Saiki, M. Aikawa and C. Ibukiyama, Assessment of vasoactive agents and vascular aging by the second derivative of photoplethysmogram waveform, Hypertension, 32 (1998), no. 2, 365-370. A. Reiss, I. Indlekofer, P. Schmidt and K. Van Laerhoven, Deep PPG: large-scale heart rate estimation with convolutional neural networks, Sensors, 19 (2019), no. 14, 3079.

The present invention is to solve the above-mentioned problems, and the reliability of the measurement parameter can be improved by automating the detection of target singularity of the PPG waveform and excluding noise that could not be distinguished in the conventionally proposed photoplethysmogram analysis method from the heart rate determination. An object of the present invention is to provide a method for analyzing PPG parameters using a DNN filter.

In addition, the present invention is intended to solve the above-mentioned problems, and PPG parameters using a DNN filter capable of real-time analysis by using a smaller number of input data and reducing the amount of calculation compared to the conventionally proposed photoplethysmogram analysis method. Its purpose is to provide an analysis method.

To achieve this object, a PPG parameter analysis method using a DNN filter according to an embodiment of the present invention includes a receiving step of receiving PPG data measured from a PPG measuring sensor; a differential data generation step of generating differential data obtained by differentiating the PPG data received in the receiving step; a signal processing step of generating input data by pre-processing the PPG data and differential data generated corresponding thereto for each predetermined analysis section; Through a predefined analysis algorithm, while learning a plurality of PPG data and learning data in which a label for a target singularity in differential data corresponding to each PPG data is stored, using an optimized analysis algorithm, the signal is processed in the signal processing step. A detection step of detecting a target singularity within the input data of; and a determination step of deriving a parameter from the target singularity detected in the detection step. , and the analysis algorithm may be a Deep Neural Network (DNN) filter composed of at least two hidden layers between an input layer and an output layer.

Here, the signal processing step is a focusing step of reducing the data size included in the PPG data corresponding to the analysis section and the differential data corresponding thereto by selecting peripheral data other than the target zone at a low frequency. ; and a normalization step of normalizing the maximum and minimum values of the PPG data signal-processed in the focusing step and the differential data corresponding thereto to 1 to -1 and converting them into a plurality of input data; Including, the analysis interval may include the size and movement interval of the window for determining the interval of the input data input to the DNN filter.

At this time, the target singularity is the maximum value point (Systolic, S) by heartbeat among the PPG data, the minimum value point immediately before heartbeat (Onset, O), the maximum value point (W) among the differential data, and heartbeat measurement among the PPG data. It may include at least one of an error start point (ES) and an error end point (EE) of the error region, which is an unavailable region.

In the detecting step, the target singularity is detected from a plurality of DNN filters for detecting each target singularity from the input data, but when one input data is analyzed and another input data is analyzed , the window for determining the other input data is moved and a partial region of the one input data overlaps, so that the target singularity of the overlapped region can be detected redundantly.

In addition, the determining step may include a deriving step of deriving a parameter through the target singularity detected in the detecting step; a scoring step of calculating a recognition score obtained by scoring the recognition level of the DNN filter for the parameters derived in the deriving step; and a determination step of determining, as a final parameter, a parameter whose recognition degree calculated in the scoring step is 80 to 100 points. Including, the recognition degree can be calculated from the following equation.

[mathematical expression]

Recognition Score = max(R _S [n])_0.2s+max(R _O [n])_0.2s+max(R _W [n])

+{sum(40-R _ES [n]-R _EE [n])_0.2s}

(Where, R _S [n], R _O [n], R _W [n], R _ES [n], and R _EE [n] are S, O, W, ES, and EE repeatedly detected by the DNN filter, respectively. The number of times, max(R[n])_0.2s is the maximum value of R[n] within 0.2 seconds from when the maximum value of R _W [n] is detected, sum(R[n])_0.2s is: R Sum of R[n] within 0.2 seconds from when the maximum value of _W [n] is detected)

Further, the method for analyzing pulse wave parameters using a DNN filter proposed by the present invention includes a storage step of storing the learning data; and a learning step of optimizing the DNN filter to determine whether a target singularity exists in arbitrary PPG data and differential data corresponding thereto while learning the learning data through the DNN filter. may further include.

As described above, according to the present invention, there are the following effects.

First, it is possible to improve the reliability of the decision result by excluding noise that cannot be distinguished by the differential method and the digital filter among the previously proposed PPG signal analysis methods from the parameter decision.

Second, it automates the target singularity detection of waveforms that could not be automated due to individual differences, and uses fewer input data compared to PPG signal analysis methods using frequency analysis or existing artificial intelligence technology. . That is, parameters analyzed in real time can be provided from PPG measurement data.

Third, by updating learning data and performing machine learning (deep learning), accuracy can be improved more easily than when a person directly corrects a program.

Fourth, it is possible to utilize the determination result in various ways for different purposes by providing an indicator (eg, recognition level) capable of indicating the reliability of the determination result.

1 is a block diagram schematically showing an apparatus for analyzing PPG parameters using a DNN filter according to an embodiment of the present invention.
2 is a flowchart schematically illustrating a method for analyzing PPG parameters using a DNN filter according to an embodiment of the present invention.
3 schematically illustrates an algorithm and a signal processing process for detecting a target singularity using a DNN filter.
4 schematically illustrates a data processing process in the PPG parameter analysis method using a DNN filter proposed by the present invention.
5 is a table illustrating training data and test data used in deep learning and evaluation in the present invention.
6 is an illustration to explain target singularities of PPG data and differential data to be detected in the present invention.
Figure 7 shows the results obtained through the photoplethysmogram parameter analysis method using the DNN filter proposed by the present invention. (a) PPG data and differential data corresponding thereto (b) target recognized by the DNN filter singularities (S, O, W, ES, EE) and (c) heart rate (HR) determined in real time.
FIG. 8 is a table showing the number of HRs measured from the ECG and the number of S, O, and W detected by the DNN filter for the data illustrated in FIG. 5 .
Figure 9 shows the correlation between HR derived through the DNN filter proposed by the present invention and HR measured from ECG. (a), (b) and (c) are from ECG when recognition is not considered (d), (e) and (f) are comparisons between HR and S, O, W measured from ECG when the recognition level is 100, ( g), (h) and (i) are comparisons of HR and S, O, and W measured from ECG when the recognition level is 80 points or higher, respectively.

A preferred embodiment of the present invention will be described in more detail with reference to the accompanying drawings, but for the sake of brevity, technical parts that have already been well-known will be omitted or compressed.

For reference, the photoplethysmogram parameter analysis method and device using the DNN filter proposed in the present invention are designed to measure HR among the above-mentioned parameters, but can be used to measure other parameters according to configuration.

<Brief description of photoplethysmogram (PPG) parameter analysis device using DNN filter>

1 schematically illustrates a photoplethysmogram (PPG) parameter analyzer (hereinafter, referred to as a 'PPG analyzer') using a DNN filter according to an embodiment of the present invention.

Referring to FIG. 1, the PPG analysis device 100 according to an embodiment of the present invention includes a receiver 110, a differential data generator 120, a signal processor 130, a storage unit 140, and a learning unit 150. ), a detection unit 160, a determination unit 170, and a control unit 180.

The receiving unit 110 receives PPG data measured from the PPG measurement sensor. At this time, the receiving unit 110 can receive PPG data measured in real time from the subject to be measured, and of course can receive PPG data for learning a DNN filter (not shown) to be described later.

The differential data generation unit 120 generates differential data obtained by differentiating the PPG data received from the receiver 110 . At this time, the differential data generation unit 120 may perform the first differentiation or the second differentiation according to the parameter to be analyzed through the PPG analysis device 100. The PPG analysis device 100 according to an embodiment of the present invention ) wants to obtain HR through PPG data analysis, so it may be desirable to perform only the first derivative.

The signal processing unit 130 is a step of preprocessing the PPG data received by the receiving unit 110 and differential data generated corresponding thereto for each predetermined analysis section to generate input data. At this time, the signal processing unit 130 focuses on reducing the size of PPG data corresponding to the analysis section and differential data generated corresponding thereto so as to reduce the amount of neurons and calculations of the DNN filter including the detection unit 160 to be described later ( Signal processing for conversion into input data may be performed through normalization to unify the y-axis range of the focused data to -1 to 1.

At this time, the analysis section includes the size and movement interval of the window for determining the section of the input data.

The storage unit 140 stores a plurality of PPG data and training data labeled with a target singularity in differential data corresponding to each PPG data. At this time, it goes without saying that the detection result of the detection unit 160, that is, the output data of the DNN filter, which will be described later, can be accumulated and additionally stored in the storage unit 140.

At this time, the target singularity is the maximum value point by heartbeat among the PPG data (Systolic, S), the minimum value point just before the heartbeat (Onset, O), the maximum value point among the differential data (W), and the PPG data where the heartbeat cannot be measured. It means the start point (Error Start; ES) and the end point (Error End; EE) of the error region.

The learning unit 150 optimizes the analysis algorithm so that a target singularity is detected in arbitrary PPG data and differential data corresponding thereto while learning learning data using a predefined analysis algorithm. That is, the learning unit 150 applies the learning data stored in the storage unit 140 to the analysis algorithm, and performs learning, verification, and testing based on this, even when given arbitrary PPG data and differential data corresponding thereto. High judgment accuracy can be maintained so that the target singularity can be detected.

For reference, optimizing the predefined analysis algorithm by the learning unit 150 means to improve the probability of an accurate judgment result, that is, the accuracy of judgment, by adjusting the weights of elements included in the predefined analysis algorithm through learning. it means.

In this case, the analysis algorithm means a Deep Neural Network (DNN) filter composed of at least two or more hidden layers between an input layer and an output layer.

The detection unit 160 detects a target singularity in any input data signal-processed by the signal processing unit 130 using the analysis algorithm optimized through the learning unit 150.

At this time, the detection unit 160 may be composed of a plurality of DNN filters for detecting each target singularity from the input data, more specifically, five DNN filters for detecting S, O, W, ES, and EE, respectively. can be configured.

Here, when the DNN filter analyzes any one input data and derives the analysis result, the window for determining the other input data moves and the other input data is continuously input. In the data, a partial area of any one of the previously analyzed input data overlaps, so that a target singularity of the overlapped area may be detected redundantly.

For reference, an analysis result through the detection unit 160, that is, detection accuracy of a target singularity may be improved as the number of redundant detections in the overlapping area increases.

The determination unit 170 derives a parameter using the target singularity detected by the detection unit 160 . At this time, a recognition score obtained by scoring the recognition level of the DNN filter for the derived parameter may be calculated, and a parameter corresponding to a recognition score of 80 to 100 may be determined and provided as a final parameter.

Here, the degree of recognition may be determined from the following equation.

[mathematical expression]

Recognition Score = max(R _S [n])_0.2s+max(R _O [n])_0.2s+max(R _W [n])

+{sum(40-R _ES [n]-R _EE [n])_0.2s}

(Where, R _S [n], R _O [n], R _W [n], R _ES [n], and R _EE [n] are S, O, W, ES, and EE repeatedly detected by the DNN filter, respectively. The number of times, max(R[n])_0.2s is the maximum value of R[n] within 0.2 seconds from when the maximum value of R _W [n] is detected, sum(R[n])_0.2s is: R Sum of R[n] within 0.2 seconds from when the maximum value of _W [n] is detected)

The control unit 180 is a component that controls each unit described above and can provide additional data input from the outside to each unit that requires the corresponding data. For example, when learning data is input, it may be provided to the storage unit 140 .

Hereinafter, a method for analyzing PPG parameters using a DNN filter by the PPG analyzer 100 will be described. In the above-mentioned PPG analysis device 100, insufficient description of its configuration describes a PPG parameter analysis method using a DNN filter and will be described in more detail.

<Description of PPG parameter analysis method using DNN filter>

Figure 2 is a flowchart schematically showing a photoplethysmogram (PPG) parameter analysis method (hereinafter referred to as 'PPG analysis method') using a DNN filter according to an embodiment of the present invention, and Figure 3 is a DNN filter using Figure 4 schematically shows the data processing process in the photoplethysmogram (PPG) parameter analysis method using the DNN filter proposed by the present invention.

Hereinafter, the PPG analysis method proposed by the present invention will be described in detail with reference to FIGS. 2 to 4.

1. Receiving step <S100>

The receiving step (S100) is a step of receiving PPG data measured from a PPG measuring sensor. For reference, in the receiving step (S100), it is of course possible to receive learning data to be learned in a learning step (not shown) to be described later.

2. Differential data generation step <S200>

The differential data generation step (S200) is a step of generating differential data obtained by differentiating the PPG data received in the receiving step (S100). At this time, suitable differential data may be generated by performing primary differential or secondary differential according to biosignal parameters to be measured. Since HR is measured as a bio-signal parameter in an embodiment of the present invention, it is preferable to generate differential data obtained by performing a primary differential of PPG data in the differential data generation step (S200).

For reference, referring to FIG. 3, in the 1 ^st STEP, in the waveform of the received PPG data, local high or low peaks (HPs, LPs) could be detected within the data length of 0.2 s, and the PPG data was differentiated Locally high peaks could be detected in one differential data. At this time, in order to automate the parameter measurement, it should be possible to determine whether the detected peaks are generated from pulsation or reflection, so in the present invention, the accuracy of the judgment can be improved without expert judgment intervention by evaluating the PPG data, the first differential data, and their target singular points. We would like to present a PPG analysis method that can be improved.

3 . Signal processing step <S300>

The signal processing step (S300) is a step of generating input data by pre-processing the PPG data and differential data corresponding thereto for each preset analysis section. At this time, it goes without saying that the analysis section may be set before PPG data is received or set by the user after PPG data is received.

Here, the signal processing step (S300) may include a focusing step (S310) and a normalizing step (S320).

The focusing step (S310) is a step of reducing the data size included in the PPG data corresponding to the analysis section and the differential data corresponding thereto by selecting peripheral data other than the target zone at a low frequency.

The normalization step (S320) is a step of normalizing the maximum and minimum values of the PPG data signal-processed in the focusing step (S320) and the corresponding differential data to 1 to -1 and converting them into a plurality of input data.

For reference, the analysis section may include the size and movement interval of a window for determining a section of input data input to a DNN filter to be described later.

3 to 4, in the 2 ^nd STEP of FIG. 3, the input data to be input to the DNN filter through the signal processing step (S300) can be reconstructed from the signal processed in 1 ^st STEP . The 74 PPG data and the corresponding differential data go through a signal processing step (S300), and the 20 consecutive PPG data and the corresponding differential data can be configured as input data of a DNN filter to be described later. If P[n] (where n is 0 to 73), the target area where target singularities appear may be an area where n is 27 to 46. Of course, this may vary depending on the case.

Thereafter, in the signal processing step (S300), all data from the target area, that is, P[n+27] to P[n+46], is included in the input data, but in data where n is 0 to 26 and 47 to 73, 12 It is possible to reduce the amount of data that the DNN filter has to calculate by adjusting the frequency so that only the data of the DNN filter is selectively included in the input data.

That is, the input data includes 32 PPG data, 32 differential data corresponding thereto, and 20 target singular points (S, W, O) located in the target area, for a total of 124 input data (X[0] to X[123 ]) could be generated.

For reference, the target singularity is the point of maximum value by heartbeat among PPG data (Systolic, S), the point of minimum value right before heartbeat (Onset, O), the point of maximum value among differential data (W), and the area where heartbeat cannot be measured among PPG data. It may include at least one of the start point (Error Start; ES) and the end point (Error End; EE) of.

4 . Detection step <S400>

The detection step (S400) is a signal processing step using an optimized analysis algorithm while learning a plurality of PPG data through a predefined analysis algorithm and learning data in which a label for a target singularity in differential data corresponding to each PPG data is stored. This is a step of detecting a target singularity in any input data signal-processed in (S300).

Here, the analysis algorithm means a Deep Neural Network (DNN) filter composed of at least two or more hidden layers between an input layer and an output layer. Since this is one embodiment, it can be varied according to the user's purpose, of course.

For reference, the input layer of the DNN filter proposed by the present invention may consist of an input layer composed of 124 neurons, two hidden layers composed of 248 neurons, and an output layer composed of 21 neurons.

In this case, in the detection step (S400), target singular points may be detected from five DNN filters for each of the aforementioned five target singular points. Three of the five DNN filters determined S, O, and W due to heartbeat, and two filters determined ES and EE, which should be excluded from decision making.

In the detection step (S400), the target singularity is detected from a plurality of DNN filters for detecting each target singularity from the input data, but when one input data is analyzed and then another input data is analyzed, A window for determining another input data is moved and a partial area of one input data overlaps, so that the target singularity of the overlapped area is detected redundantly.

In this case, the DNN filter re-evaluates most of the continuous PPG data. Looking at the windows illustrated in FIGS. 3 and 4, the time length of the window containing 74 data was 1.184 s and the data was 0.016 Processed every s and input to the DNN filter, it was possible to determine the locations of S, O, W, ES, and EE within the 0.32 s long target region. At this time, as the window moves to analyze the continuous PGG data, 20 parts of the analysis section in which the analysis is performed in the DNN filter may overlap. Here, the reliability of target singularity detection through the DNN filter can be improved as the number of overlapping regions increases. That is, the PGG analysis method proposed by the present invention can improve the accuracy of the decision result although the number of constituent neurons is smaller than that of the existing artificial intelligence-related PGG analysis model.

And, the output from the output layer can be derived as 1 and 0. For example, in the case of a DNN filter for detecting S, when the target singularity, that is, S is detected in the nth PPG data block, R _S [n] is incremented by 1, otherwise it was not incremented. If the DNN filter does not find the target singularity, the 21st value of the output layer (Y _S [20], Y _O [20], Y _W [20], Y _ES [20], Y _EE [20]) is 1 , and all other values may be output as 0. That is, if there is no target singularity that the DNN filter is looking for, the output indicating the location (Y _S [0] to Y _S [19], Y _O [0] to Y _O [19], Y _W [0] to Y _W [ 19], Y _ES [0] to Y _ES [19], Y _EE [0] to Y _EE [19]).

For reference, prior to the detection step (S400), while learning the learning data through a storage step (not shown) for storing learning data and a DNN filter, it is determined whether a target singularity exists in arbitrary PPG data and differential data corresponding thereto. A learning step (not shown) of optimizing the DNN filter may be preceded.

More specifically, when learning data, which is PPG data for learning, is received through the receiving step (S100), it is labeled by an expert through the aforementioned differential data generation step (S200) and signal processing step (S300). The learning step is performed based on the learning data for which the answer is predetermined, and the analysis algorithm, that is, the DNN filter can be optimized. At this time, weights are added to the neurons included in each layer through a learning step, and adjustments are made repeatedly. Through this, an optimal algorithm that provides highly accurate judgment results can be configured. For reference, the recognition accuracy of S, O, W, ES, and EE increased as the deep learning of the DNN filter, that is, the learning progressed. ), 99.29% (O), 99.96 % (W), 99.45 % (ES), and 98.81 % (EE), and when 100 epochs were performed, the accuracy was 98.81 % (S), 97.03 % (O), 98.59 % (W) ), 92.49 % (ES), and 92.17 % (EE), it was confirmed that the accuracy increased to 6.96 % as learning progressed.

In the PPG analysis method proposed by the present invention, learning data for learning the DNN filter is illustrated in FIG. For reference, the subject data illustrated in FIG. 5 is obtained from Physionet (physionet.org), which distributes data collected by medical institutions around the world for clinical research or for other reasons as open data.

At this time, all data was recorded for 8 minutes at 125 Hz, and the sampling rate (125 Hz) to provide data to the DNN filter was reduced to 62.5 Hz. This is to reduce the amount of training data while increasing the deep learning efficiency of the DNN filter.

For reference, in PPG data and differential data, LP, HP, S, O, and W are defined as illustrated in FIG. 6. At this time, through the learning step, the DNN filter learns the training data labeled with the positions of S, O, W, ES, and EE in the PPG data and the corresponding differential data by the expert, and outputs the target singularity as output data. it will be optimized

5. Judgment step <S500>

The determination step (S500) is a step of deriving bio-signal parameters from the target singularities detected in the detection step (S400).

Here, the determination step (S500) may include a derivation step (S510), a scoring step (S520) and a determination step (S530).

The derivation step (S510) is a step of deriving a parameter (eg, HR) through the target singularity detected in the detection step (S400). At this time, HR was calculated using the ns gap difference and the position of the consecutive maxima of R _S [n], R _O [n], and R _W [n]. At this time, the final HR was determined through a scoring step (S520) and a decision step (S530) to be described later in order to minimize a decision error due to motion noise.

The scoring step (S520) is a step of calculating a recognition score obtained by scoring the recognition level of the DNN filter for the parameters derived in the derivation step (S510). In this case, the degree of recognition may be calculated from the following equation.

[mathematical expression]

Recognition Score = max(R _S [n])_0.2s+max(R _O [n])_0.2s+max(R _W [n])

+{sum(40-R _ES [n]-R _EE [n])_0.2s}

(Where, R _S [n], R _O [n]), R _W [n], R _ES [n], and R _EE [n] are S, O, W, ES, and S repeatedly detected by the DNN filter, respectively. The number of EEs, max(R[n])_0.2s, is the maximum value of R[n] within 0.2 seconds from when the maximum value of R _W [n] is detected, sum(R[n])_0.2s is: Sum of R[n] within 0.2 seconds from when the maximum value of R _W [n] was detected)

At this time, in the scoring step (S520), the degree to which the DNN filter recognizes the parameters derived by each target singularity is scored by summing the number of detections of the target singularity by several DNN filters or deducting the number of times the target singularity is recognized as an error. In other words, the degree of recognition of HR by the five DNN filters is scored. Based on the value of the maximum number of recognitions of W, the maximum recognition degree of S and O within 0.2 s is summed, and the sum of the recognition times of ES and EE is subtracted. way to do it In the above-mentioned equation, when S, O, and W all acquire 20 points and ES and EE acquire 0 points, 100 points of recognition can be obtained.

The determining step (S530) is a step of determining, as a final parameter, a parameter whose recognition degree calculated in the scoring step (S520) is 80 to 100 points. For reference, as the recognition degree approaches 100 points, the allowable error size decreases. Therefore, when the recognition degree is 100 points, there is no error at all, and only certain target singularities are distinguished and used for parameter (eg, HR) calculation.

Referring to FIG. 7, the present invention detected S, O, and W from the PPG data and the corresponding differential data through the above-described series of processes using PPG data obtained through PPG measurement, and S, O, W We were able to acquire 0.7 s delayed HR from the heartbeat using the interval between the previous and current positions of .

Hereinafter, evaluation of the accuracy of parameters (eg, HR) determined from the PPG analysis method proposed by the present invention will be described.

First, FIG. 8 is a table showing the number of HRs measured using the ECG QRS detection method and the number of S, O, and W detected by the DNN filter for the data illustrated in FIG. 5 . At this time, in the case of the subject data included in the training data, the degree of agreement between the HR measured from the ECG QRS and the HR judgment of S, O, and W detected by the DNN filter was 96%, and the subject's data included in the test data In the case of , it was 97.5%.

In addition, Figure 9 compares the HR derived through the DNN filter proposed by the present invention and the HR measured using the ECG QRS detection method, and (a), (b) and (c) do not consider recognition. (d), (e), and (f) show the correlation between HR measured from ECG and S, O, W, respectively, and (d), (e) and (f) are HR measured from ECG and S, O, W respectively (g), (h) and (i) show the correlation between HR measured from ECG and S, O, and W, respectively, when the recognition level is 80 points or higher.

At this time, (a), (b) and (c) of FIG. 9 show the correlation between HR derived by the five DNN filters and HR measured from the ECG, and the Pearson correlation coefficient is a part detected by the DNN filter Due to errors in HR, they had low values of 0.36 (S), 0.33 (O) and 0.41 (W), respectively. At this time, the difference between the HR measured from the ECG and the HR derived by the DNN filter was 2.93±17.23 bpm(S), 3.20±18.38 bpm(O), and 2.22±15.30 bpm(W). That is, when the decision result of the DNN filter was accepted 100% without considering recognition, the number of abnormal decisions could increase.

And, (d), (e) and (f) of FIG. 9 show the correlation between HR derived when the recognition level of the five DNN filters is 100 and the HR measured from the ECG, as shown in FIGS. 5 and 8 Of the 10,036 measured in the training data and test data, 2456 (24.47%) scored 100 points, and Pearson's correlation coefficients showed high values of 0.914 (S), 0.871 (O), and 0.922 (W), respectively. At this time, the difference between the HR measured from the ECG and the HR derived by the DNN filter was -0.02±2.80 bpm(S), 0.55±3.58 bpm(O), and 0.06±2.60 bpm(W).

In addition, (g), (h) and (i) of FIG. 9 show the correlation between HR derived from HR and HR measured from ECG when the recognition degree of the five DNN filters is 80 to 100 points, FIG. 5 and FIG. 8643 (86.12%) out of 10,036 measured in the training data and test data presented in Figure 8 obtained a score of 80 or higher, and Pearson's correlation coefficients showed high values of 0.912 (S), 0.850 (O), and 0.919 (W), respectively. . At this time, the difference between the HR measured from the ECG and the HR derived by the DNN filter was 0.19 ± 4.07 bpm (S), 0.43 ± 5.62 bpm (O), and 0.12 ± 3.89 bpm (W).

In other words, when the judgment result with recognition level of 80 points or more was accepted, the correlation coefficient with the HR result measured using ECG increased to 0.9. Accordingly, the PPG analysis method proposed by the present invention, which determines the final HR from the judgment data when the recognition degree is 80 to 100 points, improves the reliability of the finally determined HR by removing a large number of error areas, that is, abnormal data. could make it At this time, since the permissible range of the degree of recognition means parameters, that is, the allowable error size and the accuracy of the target singularity in the HR calculation process, the user can adjust it according to the purpose.

For reference, 1356 (13.51%) had a recognition score of 1 to 79, and EE and ES scores were high due to noise, etc., and some HRs had a recognition score of 0 due to arrhythmia.

As described above, the detailed description of the present invention has been made by the embodiments with reference to the accompanying drawings, but since the above-described embodiments have only been described as preferred examples of the present invention, the present invention is limited to the above embodiments. It should not be understood as being, and the scope of the present invention should be understood as the following claims and equivalent concepts thereof.

100: PPG analysis device
110: receiver
120: differential data generation unit
130: signal processing unit
140: storage unit
150: learning unit
160: detection unit
170: judgment unit
180: control unit

Claims (6)

As an analysis method for deriving PPG parameters from PPG data measured from a photoplethysmogram (PPG) measurement sensor using an analysis algorithm to which a deep neural network (DNN) filter is applied,
a reception step in which a reception unit receives the PPG data;
a differential data generation step in which a differential data generator generates differential data obtained by differentially dividing the PPG data received in the receiving step;
The PPG signal processing unit includes all the data of the target zone in which the target singularity appears in the PPG data and the differential data generated corresponding thereto, but selectively includes the PPG data outside the target zone for each predetermined analysis section. a signal processing step of generating input data having a reduced data size by adjusting a data frequency included in the data and differential data generated corresponding thereto;
Signal processing in the signal processing step by using an optimized analysis algorithm while the detection unit learns learning data in which labels for target singular points in a plurality of PPG data and differential data corresponding to each PPG data are stored through a predefined analysis algorithm random input data A detection step of detecting my target singularity; and
a determination step in which a determination unit derives a parameter from the target singularity detected in the detection step; including,
The analysis algorithm is a Deep Neural Network (DNN) filter composed of at least two hidden layers between an input layer and an output layer,
The target singularity may measure the maximum value point (Systolic, S) by heartbeat among the PPG data, the minimum value point immediately before heartbeat (Onset, O), the maximum value point (W) among the differential data, and the heartbeat among the PPG data. Characterized in that it includes at least one of the error region's start point (Error Start; ES) and the error region's end point (Error End; EE))
Photoplethysmogram (PPG) parameter analysis method using DNN filter.
According to claim 1,
The signal processing step,
a focusing step of selecting peripheral data other than the target area at a low frequency and reducing the data size included in the PPG data corresponding to the analysis section and the differential data corresponding thereto; and
a normalization step of normalizing the maximum and minimum values of the PPG data signal-processed in the focusing step and the differential data corresponding thereto to 1 to -1 and converting them into input data; including,
Characterized in that the analysis section includes a size and movement interval of a window for determining a section of input data input to the DNN filter.
Photoplethysmogram (PPG) parameter analysis method using DNN filter.
According to claim 2,
The detecting step is characterized in that each of the target singular points is detected from a DNN filter equipped with a number corresponding to the number of target singular points to be detected in the detecting step so as to detect any one target singular point among the target singular points.
Photoplethysmogram (PPG) parameter analysis method using DNN filter.
According to claim 3,
In the detecting step, as the window moves so that most areas of continuous PPG data are re-evaluated using a single DNN filter for detecting any one target singularity, the input analyzed by the DNN filter Characterized in that some areas of the analysis section of the data overlap, and the target singularity of the overlapped area is redundantly detected
Photoplethysmogram (PPG) parameter analysis method using DNN filter.
According to claim 4,
The judgment step is
a derivation step of deriving a parameter through the target singularity detected in the detection step;
The five DNN filters for individually detecting the S, O, W, ES, and EE increase or decrease the number of times the target singularity is repeatedly detected, and the heart between the previous position and the current position of any one of the S, O, and W a scoring step of scoring Recognition Scores of the DNN filters for the parameters derived in the derivation step in beats; and
a determination step of determining a parameter whose recognition degree calculated in the scoring step is between 80 and 100 points as a final parameter; including,
The recognition degree is expressed by the following equation
[mathematical expression]
Recognition Score = max(R _S [n])_0.2s+max(R _O [n])_0.2s+max(R _W [n])
+{sum(40-R _ES [n]-R _EE [n])_0.2s}
―Where, R _S [n], R _O [n], R _W [n], R _ES [n], and R _EE [n] are S, O, W, ES, and EE repeatedly detected by the DNN filter, respectively. The number of times, max(R[n])_0.2s is the maximum value of R[n] within 0.2 seconds from when the maximum value of R _W [n] is detected, sum(R[n])_0.2s is: R Sum of R[n] within 0.2 seconds from when the maximum value of _W [n] is detected—
characterized in that it is calculated from
Photoplethysmogram (PPG) parameter analysis method using DNN filter.
According to claim 5,
a storage step in which the control unit stores the learning data in a storage unit; and
a learning step of optimizing the DNN filter so that a learning unit learns the learning data through the DNN filter and determines whether or not a target singularity exists in arbitrary PPG data and differential data corresponding thereto; characterized in that it further comprises
Photoplethysmogram (PPG) parameter analysis method using DNN filter.

Images