Pulse wave velocity detection method based on terminal volume pulse wave signal

Abstract

The present invention relates to a pulse wave velocity detection method based on terminal volume pulse wave signal, including: collecting original PPG signal, basic physiological information, health information and pulse wave velocity of human body; signal processing: denoising the original PPG signal to obtain denoised PPG waveform; extracting features of the denoised PPG waveform to obtain PPG features; introducing basic physiological information and health information, merging PPG features, screening out features and logical features characterizing PWV, and obtaining feature subsets; using machine learning algorithms to establish a PWV estimation model, inputting the feature subset of the object to be measured into the PWV estimation model, and obtaining an estimated PWV value. The present invention only needs to detect the pulse wave signal of a single part; the present invention can quickly detect the PPG signal, the detection method is convenient and fast, and has good repeatability, avoiding the problem of professionals manually searching for pulse fluctuation points on the body surface when using pressure sensors.

Description

Technical Field

[0001] The invention relates to the technical field of biomedical engineering, and in particular to a pulse wave conduction velocity detection method based on a terminal volume pulse wave signal. Background Art

[0002] Pulse Wave Velocity (PWV) is an important indicator for assessing human arterial stiffness, and is of great significance for early detection of functional arterial lesions. Changes in the arterial wall and its stiffness are the basis and key to the occurrence of early cardiovascular disease. With the advancement of chronic disease prevention and treatment, the rapid detection method of non-invasive pulse wave velocity is an important means of preventing chronic diseases, and is of great significance in the "upstream prevention and treatment" of cardiovascular and cerebrovascular diseases.

[0003] Clinically, a multi-point synchronous non-invasive pulse wave velocity detection method is often used, such as the Complior Analyser, the "gold standard" device, which requires professional medical technicians to first find the pulse beat points at multiple main arteries on the subject's body surface as detection points, and then use piezoelectric sensors to press the detection points at the same time to obtain a synchronized multi-channel pressure pulse wave, and finally obtain the pulse wave velocity based on the distance L between the detection points and the transmission time difference △T of the pressure pulse wave: PWV = L / △T. However, this method is large in size and expensive in price, and requires high professionalism from the operator, is complex to operate, involves the privacy of the subject, needs to be performed in a hospital, is not suitable for daily home or community testing, and has difficulties in monitoring non-disease populations.

[0004] In addition, many inventions or studies improve the applicability of PWV detection by changing the detection method of the pulse wave signal. For example, a pulse wave conduction velocity detection method based on dual-channel PPG signals detects PPG signals at the same time on the fingers and toes to obtain the pulse wave transmission time PTT; a pulse wave conduction velocity detection method based on ECG signals and PPG signals extracts the time difference between the ECG signal R wave and a specific point of the PPG signal (such as the starting point, the main wave peak point) as the pulse wave transmission time PTT. Both methods calculate the pulse wave conduction velocity according to the formula PWV = L / PTT, where L is the distance between the detection points of the synchronous signal. Although these methods reduce the professional requirements for the operator, they still require synchronous signal detection at two detection points. Summary of the invention

[0005] In order to solve the defects that traditional pulse wave conduction velocity needs to detect pulse waves of two or more parts and requires signal synchronization, the purpose of the invention is to provide a pulse wave conduction velocity detection method based on terminal volume pulse wave signal, which only needs to detect the pulse wave signal of a single part, has a fast and convenient detection method, and is suitable for home detection.

[0006] To achieve the above object, the present invention adopts the following technical solution: a pulse wave velocity detection method based on terminal volume pulse wave signal, the method comprising the following steps in sequence:

[0007] (1) Data collection: Collecting the original PPG signal, basic physiological information, health information and pulse wave conduction velocity of the human body;

[0008] (2) Signal processing: De-noising the original PPG signal to obtain a de-noised PPG waveform;

[0009] (3) PPG feature extraction: extract features from the denoised PPG waveform to obtain PPG features;

[0010] (4) Feature subset construction: introduce basic physiological information and health information, merge PPG features, screen out features and logical features that characterize PWV, and obtain feature subsets;

[0011] (5) PWV estimation: A PWV estimation model is established using a machine learning algorithm. The feature subset of the subject to be tested is input into the PWV estimation model to obtain the estimated PWV value.

[0012] The step (1) specifically refers to: using a jacket device to collect original PPG signals from human body ends, the human body ends include fingers, toes and earlobes, while recording basic physiological information and health information, and using a pulse wave velocity detector to detect the pulse wave velocity; the jacket device is equipped with a photoelectric capacitance pulse wave sensor, the basic physiological information includes age, gender, height, weight and body mass index BMI, and the health information includes whether the person smokes and whether the person has a history of chronic diseases.

[0013] The step (2) specifically refers to: using a denoising method based on discrete wavelet transform to remove baseline drift and high-frequency noise from the original PPG signal to obtain a denoised PPG waveform; the denoising method based on discrete wavelet transform adopts a wavelet threshold denoising method, and the wavelet threshold denoising method includes wavelet multi-layer decomposition, wavelet threshold processing and wavelet coefficient reconstruction:

[0014] (2a) Wavelet multi-layer decomposition: Perform multi-layer decomposition on the original PPG signal to obtain wavelet decomposition coefficients CD1, CD2, CD3, ..., CDn, CAn, where n is the number of decomposition layers;

[0015] (2b) Wavelet threshold processing: Set the threshold λ to perform threshold processing on each layer of wavelet decomposition coefficients in step (2a) to obtain the processed wavelet coefficients;

[0016] The threshold λ uses the VisuShrink threshold, and the formula is as follows:

[0017]

[0018] Where N is the length of each layer of wavelet coefficients, log(N) is the logarithm of N with base e, δ is the estimated noise standard deviation, median|| is the median of the absolute values ​​of the wavelet coefficients of the corresponding layer arranged in ascending order, and 0.6745 is the adjustment coefficient of the standard deviation of Gaussian white noise;

[0019] Threshold processing uses a soft threshold function processing method, the formula is as follows:

[0020]

[0021] Among them, sgn(n) is the sign function, f(x) is the processed wavelet coefficient;

[0022] (2c) Wavelet coefficient reconstruction: Reconstruct f(x) to obtain the reconstructed signal as the denoised PPG waveform.

[0023] The step (3) specifically refers to: performing a first-order difference operation on the denoised PPG waveform to obtain a first-order difference PPG waveform, and performing a second-order difference operation on the denoised PPG waveform to obtain a second-order difference PPG waveform:

[0024] (3a) Performing reference point identification: Based on the denoised PPG waveform and the first-order differential PPG waveform, a differential threshold algorithm is used to identify the starting point A on the PPG waveform, and the PPG waveform is divided into multiple single-cycle PPG waveforms according to the starting point A and the next cycle starting point A1 adjacent thereto. Similarly, the second-order differential PPG waveform is also divided into multiple single-cycle second-order differential PPG waveforms; the differential threshold algorithm is used again to identify the reference points on the single-cycle PPG waveform, which are the starting point A, the slope point K, the main wave B, the descending gorge C, the heavy beat wave D and the next cycle starting point A1, and the reference points on the single-cycle second-order differential PPG waveform are identified, which are wave point a, wave point b, wave point c, wave point d and wave point e;

[0025] (3b) According to the reference points identified in step (3a), features are extracted from multiple single-cycle PPG waveforms and single-cycle second-order differential PPG waveforms, and the value of each feature is the average value of the features extracted from multiple single-cycle PPG waveforms; the feature extraction methods on the single-cycle PPG waveform include pulse wave contour analysis method and pulse area method. The pulse wave contour analysis method uses 6 reference points, namely, starting point A, slope point K, main wave B, descending gorge C, dicrotic wave D and the starting point A1 of the next cycle to extract features. The extracted features include pulse period Period, pulse rate HR, acceleration time SlopeTime, maximum slope SlopeMax, peak time CT, peak ratio CTR, normalized peak ratio NCT, systolic time ED, diastolic time DiaTime, ejection fraction EF, peak-to-peak time PPT, sclerosis index SI, reflectivity index RI, left heart load index SPTI, myocardial perfusion index DPTI, and subendocardial myocardial vitality rate SEVR. The formula is as follows:

[0026] Period=T

[0027] HR=60 / T

[0028] SlopeTime=T1

[0029] CT=T2

[0030] CTR=T2 / T

[0031] NCT=T2 / H

[0032] ED=T3

[0033] DiaTime=T4

[0034] EF=T3 / T

[0035] PPT=T5

[0036] SI=H / T5

[0037] RI=h4 / h2

[0038] SPTI=S1

[0039] DPTI=S2

[0040] SEVR=S2 / S1

[0041] Wherein, S1 is the integral of the amplitude of the cardiac ejection period over time, S2 is the integral of the amplitude of the cardiac contraction period over time, H is the individual height, T is the cycle time of a complete contraction and relaxation activity of the heart, that is, a pulse cycle, T1 is the time from the starting point A to the slope point K, T2 is the main wave rise time from the starting point A to the main wave B, T3 is the time from the starting point A to the descending gorge C, T4 is the time from the descending gorge C to the starting point A1 of the next cycle, T5 is the time from the main wave B to the dicrotic wave D, h2 is the amplitude difference from the main wave B to the starting point A, h4 is the amplitude difference from the dicrotic wave D to the starting point A, SlopeMax is the amplitude of the first-order differential PPG waveform at the slope point K;

[0042] The pulse area method uses two reference points and the overall pulse wave area change to extract features. The two reference points include the starting point A and the next cycle starting point A1. The extracted features include waveform characteristic parameters K1, cardiac output SV, cardiac output CO, peripheral resistance TPR, and arterial compliance AC. The formula is as follows:

[0043] K1=(Q m -Q min ) / (Q max -Q min )

[0044]

[0045]

[0046] TPR=Q m / CO

[0047] AC=SV / (Q max -Q min )

[0048] Among them, Q m is the average pulse amplitude of a single-cycle PPG waveform, T is the pulse period, Q min is the minimum pulse amplitude of a single-cycle PPG waveform, Q max is the maximum pulse amplitude of a single-cycle PPG waveform, It is the integral of the single-cycle PPG waveform amplitude from time 0 to T;

[0049] The feature extraction on the single-cycle second-order differential PPG waveform uses five wave reference points, namely the amplitude ratios of wave points a, b, c, d and e. The extraction formulas of these features are: AMR1=b / a, AMR2=c / a, AMR3=d / a, AMR4=e / a, AMR5=(bc) / a, AMR6=(bcd) / a and AMR7=(bcde) / a, where a, b, c, d and e represent the amplitudes of the corresponding reference points respectively, and AMR1 to AMR7 are all amplitude ratios.

[0050] The step (4) specifically refers to: taking all the features extracted in step (3) as PPG features, and combining the PPG features with basic physiological information and health information as a feature set to be screened; using correlation analysis and Lasso regression to screen out features characterizing PWV, and using a tree-based decision mechanism, i.e., random forest, to screen out logical features, and the features characterizing PWV and the logical features are used together as a feature subset for PWV estimation;

[0051] Correlation analysis was used to determine whether the feature was significantly correlated with PWV. If the P value of the analysis result was less than 0.05, that is, P<0.05, it indicated that the feature was significantly correlated with PWV and the feature was retained. On the contrary, if the P value of the analysis result was greater than or equal to 0.05, that is, P≥0.05, it indicated that the feature was not significantly correlated with PWV and the feature was removed.

[0052] Use Lasso regression to compress the coefficient of the feature. If the regression coefficient of the feature is not 0, it means that the feature is important to the result and the feature is retained. If the regression coefficient of the feature is 0, it means that the feature is not important to the result and the feature is removed.

[0053] Use the feature importance ranking in the random forest to select the top-ranked features. Among all the base decision trees in the random forest, select the base decision trees that use these top-ranked features, remove the base decision trees that do not use the top-ranked features, and use the decision logic of the selected base decision tree as a logical feature.

[0054] The step (5) specifically refers to: using a machine learning algorithm to establish a PWV estimation model, taking a feature subset as input, taking the actually detected PWV, i.e., the pulse wave velocity detected by a pulse wave velocity detector, as a learning target, training different PWV estimation models for feature subsets corresponding to different target objects; collecting the original PPG signal, basic physiological information, and health information of the object to be tested, processing them to obtain a feature subset, inputting the feature subset into the PWV estimation model, and the PWV estimation model automatically outputs an estimated PWV value, i.e., the pulse wave velocity.

[0055] It can be seen from the above technical solutions that the beneficial effects of the present invention are as follows: first, the present invention only needs to detect the pulse wave signal of a single part, thus avoiding the problem that the traditional pulse wave conduction velocity needs to detect the pulse waves of two or more parts and requires signal synchronization; second, the present invention uses the PPG signal to replace the traditional pressure pulse wave signal, and only needs to use a jacket-type device to clamp the end of the human body to quickly detect the PPG signal. The detection method is convenient and fast with good repeatability, thus avoiding the problem that professionals need to manually search for pulse fluctuation points on the body surface when using a pressure sensor; third, the present invention is easy to popularize and suitable for home detection. The PPG feature subset is obtained based on the pulse wave contour analysis method, and the PWV prediction model is obtained using the machine learning method. The detection method is easily applied to wearable devices to realize real-time pulse wave conduction velocity monitoring and achieve the purpose of "early detection and early prevention". BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 This is the flow chart of the wavelet denoising method;

[0058] Figure 3 Schematic diagram of multi-layer wavelet decomposition of the original PPG signal;

[0059] Figure 4 It is a typical single-cycle PPG waveform and its reference points;

[0060] Figure 5 It is the vein map area method and its feature extraction method;

[0061] Figure 6 It is a typical single-cycle second-order differential PPG waveform and its reference point. DETAILED DESCRIPTION

[0062] like Figure 1 As shown, a pulse wave velocity detection method based on a terminal volume pulse wave signal comprises the following steps in sequence:

[0063] (1) Data collection: Collecting the original PPG signal, basic physiological information, health information and pulse wave conduction velocity of the human body;

[0064] (2) Signal processing: De-noising the original PPG signal to obtain a de-noised PPG waveform;

[0065] (3) PPG feature extraction: extract features from the denoised PPG waveform to obtain PPG features;

[0066] (4) Feature subset construction: introduce basic physiological information and health information, merge PPG features, screen out features and logical features that characterize PWV, and obtain feature subsets;

[0067] (5) PWV estimation: A PWV estimation model is established using a machine learning algorithm. The feature subset of the subject to be tested is input into the PWV estimation model to obtain the estimated PWV value.

[0068] The step (1) specifically refers to: using a jacket device to collect the original PPG signal of the human body end, the human body end includes fingers, toes and earlobes, and recording basic physiological information and health information at the same time, and using a pulse wave velocity detector to detect the pulse wave velocity; the jacket device is equipped with a photoelectric capacitance pulse wave sensor, the basic physiological information includes age, gender, height, weight and body mass index BMI, and the health information includes whether smoking and whether there is a history of chronic diseases. The pulse wave velocity detector uses a Complior Analyser device commonly used in clinical practice.

[0069] like Figure 2 As shown, the step (2) specifically refers to: using a denoising method based on discrete wavelet transform to remove baseline drift and high-frequency noise from the original PPG signal to obtain a denoised PPG waveform; the denoising method based on discrete wavelet transform adopts a wavelet threshold denoising method, and the wavelet threshold denoising method includes wavelet multi-layer decomposition, wavelet threshold processing and wavelet coefficient reconstruction:

[0070] (2a) Wavelet multi-layer decomposition: The original PPG signal is decomposed into multiple layers to obtain wavelet decomposition coefficients CD1, CD2, CD3, ..., CDn, CAn, where n is the number of decomposition layers, such as Figure 3 As shown;

[0071] (2b) Wavelet threshold processing: Set the threshold λ to perform threshold processing on each layer of wavelet decomposition coefficients in step (2a) to obtain the processed wavelet coefficients;

[0072] The threshold λ uses the VisuShrink threshold, and the formula is as follows:

[0073]

[0074] Where N is the length of each layer of wavelet coefficients, log(N) is the logarithm of N with base e, δ is the estimated noise standard deviation, median|coeff| is the median of the absolute values ​​of the wavelet coefficients of the corresponding layer arranged in ascending order, and 0.6745 is the adjustment coefficient of the standard deviation of Gaussian white noise;

[0075] Threshold processing uses a soft threshold function processing method, the formula is as follows:

[0076]

[0077] Among them, sgn(n) is the sign function, f(x) is the processed wavelet coefficient;

[0078] (2c) Wavelet coefficient reconstruction: Reconstruct f(x) to obtain the reconstructed signal as the denoised PPG waveform.

[0079] The step (3) specifically refers to: performing a first-order difference operation on the denoised PPG waveform to obtain a first-order difference PPG waveform, and performing a second-order difference operation on the denoised PPG waveform to obtain a second-order difference PPG waveform:

[0080] (3a) Perform benchmark identification, such as Figure 4 As shown: Based on the denoised PPG waveform and the first-order differential PPG waveform, the differential threshold algorithm is used to identify the starting point A on the PPG waveform, and the PPG waveform is divided into multiple single-cycle PPG waveforms according to the starting point A and the next cycle starting point A1 adjacent thereto. Similarly, the second-order differential PPG waveform is also divided into multiple single-cycle second-order differential PPG waveforms; the differential threshold algorithm is used again to identify the reference points on the single-cycle PPG waveform, which are the starting point A, the slope point K, the main wave B, the descending gorge C, the heavy beat wave D and the next cycle starting point A1, and the reference points on the single-cycle second-order differential PPG waveform are identified, which are wave point a, wave point b, wave point c, wave point d and wave point e;

[0081] (3b) According to the reference points identified in step (3a), features are extracted from multiple single-cycle PPG waveforms and single-cycle second-order differential PPG waveforms, and the value of each feature is the average value of the features extracted from multiple single-cycle PPG waveforms; the feature extraction methods on the single-cycle PPG waveform include pulse wave contour analysis method and pulse area method. The pulse wave contour analysis method uses 6 reference points, namely, starting point A, slope point K, main wave B, descending gorge C, dicrotic wave D and the starting point A1 of the next cycle to extract features. The extracted features include pulse period Period, pulse rate HR, acceleration time SlopeTime, maximum slope SlopeMax, peak time CT, peak ratio CTR, normalized peak ratio NCT, systolic time ED, diastolic time DiaTime, ejection fraction EF, peak-to-peak time PPT, sclerosis index SI, reflectivity index RI, left heart load index SPTI, myocardial perfusion index DPTI, and subendocardial myocardial vitality rate SEVR. The formula is as follows:

[0082] Period=T

[0083] HR=60 / T

[0084] SlopeTime=T1

[0085] CT=T2

[0086] CTR=T2 / T

[0087] NCT=T2 / H

[0088] ED=T3

[0089] DiaTime=T4

[0090] EF=T3 / T

[0091] PPT=T5

[0092] SI=H / T5

[0093] RI=h4 / h2

[0094] SPTI=S1

[0095] DPTI=S2

[0096] SEVR=S2 / S1

[0097] Wherein, S1 is the integral of the amplitude of the cardiac ejection period over time, S2 is the integral of the amplitude of the cardiac contraction period over time, H is the individual height, T is the cycle time of a complete contraction and relaxation activity of the heart, that is, a pulse cycle, T1 is the time from the starting point A to the slope point K, T2 is the main wave rise time from the starting point A to the main wave B, T3 is the time from the starting point A to the descending gorge C, T4 is the time from the descending gorge C to the starting point A1 of the next cycle, T5 is the time from the main wave B to the dicrotic wave D, h2 is the amplitude difference from the main wave B to the starting point A, h4 is the amplitude difference from the dicrotic wave D to the starting point A, SlopeMax is the amplitude of the first-order differential PPG waveform at the slope point K;

[0098] The common feature extraction on PPG waveform is shown in Table 1:

[0099] Table 1

[0100]

[0101]

[0102] like Figure 5 As shown, the pulse area method uses two reference points and the overall pulse wave area change to extract features. The two reference points include the starting point A and the next cycle starting point A1. The extracted features include waveform characteristic parameters K1, cardiac output SV, cardiac output CO, peripheral resistance TPR, and arterial compliance AC. The formula is as follows:

[0103] K1=(Q m -Qmin ) / (Q max -Q min )

[0104]

[0105]

[0106] TPR=Q m / CO

[0107] AC=SV / (Q max -Q min )

[0108] Among them, Q m is the average pulse amplitude of a single-cycle PPG waveform, T is the pulse period, Q min is the minimum pulse amplitude of a single-cycle PPG waveform, Q max is the maximum pulse amplitude of a single-cycle PPG waveform, It is the integral of the single-cycle PPG waveform amplitude from time 0 to T;

[0109] like Figure 6 As shown, the feature extraction on the single-cycle second-order differential PPG waveform uses five wave reference points, namely the amplitude ratios of wave point a, wave point b, wave point c, wave point d and wave point e. The extraction formulas of these features are: AMR1=b / a, AMR2=c / a, AMR3=d / a, AMR4=e / a, AMR5=(bc) / a, AMR6=(bcd) / a and AMR7=(bcde) / a, where a, b, c, d and e represent the amplitudes of the corresponding reference points respectively, and AMR1 to AMR7 are all amplitude ratios.

[0110] The step (4) specifically refers to: taking all the features extracted in step (3) as PPG features, and combining the PPG features with basic physiological information and health information as a feature set to be screened; using correlation analysis and Lasso regression to screen out features characterizing PWV, and using a tree-based decision mechanism, i.e., random forest, to screen out logical features, and the features characterizing PWV and the logical features are used together as a feature subset for PWV estimation;

[0111] Correlation analysis was used to determine whether the feature was significantly correlated with PWV. If the P value of the analysis result was less than 0.05, that is, P<0.05, it indicated that the feature was significantly correlated with PWV and the feature was retained. On the contrary, if the P value of the analysis result was greater than or equal to 0.05, that is, P≥0.05, it indicated that the feature was not significantly correlated with PWV and the feature was removed.

[0112] Use Lasso regression to compress the coefficient of the feature. If the regression coefficient of the feature is not 0, it means that the feature is important to the result and the feature is retained. If the regression coefficient of the feature is 0, it means that the feature is not important to the result and the feature is removed.

[0113] Use the feature importance ranking in the random forest to select the top-ranked features. Among all the base decision trees in the random forest, select the base decision trees that use these top-ranked features, remove the base decision trees that do not use the top-ranked features, and use the decision logic of the selected base decision tree as a logical feature.

[0114] The step (5) specifically refers to: using a machine learning algorithm to establish a PWV estimation model, taking a feature subset as input, taking the actually detected PWV, i.e., the pulse wave velocity detected by a pulse wave velocity detector, as a learning target, training different PWV estimation models for feature subsets corresponding to different target objects; collecting the original PPG signal, basic physiological information, and health information of the object to be tested, processing them to obtain a feature subset, inputting the feature subset into the PWV estimation model, and the PWV estimation model automatically outputs an estimated PWV value, i.e., the pulse wave velocity.

[0115] In summary, the present invention uses a single human end and a clip device to detect PPG signals. The human end is not limited to fingers, but can also be parts containing rich blood flow information such as toes and earlobes; the clip device is not limited to a finger clip device, but can also be a finger ring type, ear clip type, etc. This combined PPG signal detection method, under the existing sensor technology conditions, not only ensures the stability requirement of the PPG signal, but also avoids the synchronization requirement; the PPG signal processing method in the present invention uses a denoising method based on discrete wavelet transform, which mainly includes the determination of wavelet basis and decomposition scale for different terminal PPG signals, the selection of wavelet threshold denoising method, etc.; the PPG feature extraction method in the present invention uses a pulse wave contour analysis method to first identify reference points with clear characteristics on the waveform, and then extract various PPG features based on the reference points. The PPG feature has a clear physiological meaning, not just a number in a statistical sense; the construction method of the PPG feature subset in the present invention adopts three methods: correlation analysis, Lasso regression and random forest. For a single PPG feature, if it shows similar values ​​in three different feature selection methods, it is included in the feature subset; for the combined feature logic, if the result of the logic shows the value of PWV estimation, it is included in the PPG feature subset. The final PPG feature subset includes both traditional PPG features and logical features of grassroots decision-making.

Claims

1. A method for detecting pulse wave velocity based on terminal volume pulse wave signal, characterized in that: The method comprises the following steps in order: (1) Data collection: Collecting the original PPG signal, basic physiological information, health information and pulse wave conduction velocity of the human body; (2) Signal processing: De-noising the original PPG signal to obtain a de-noised PPG waveform; (3) PPG feature extraction: extract features from the denoised PPG waveform to obtain PPG features; (4) Feature subset construction: introduce basic physiological information and health information, merge PPG features, screen out features and logical features that characterize PWV, and obtain feature subsets; (5) PWV estimation: Use machine learning algorithms to establish a PWV estimation model, input the feature subset of the subject to be tested into the PWV estimation model, and obtain the estimated PWV value; The step (3) specifically refers to: performing a first-order difference operation on the denoised PPG waveform to obtain a first-order difference PPG waveform, and performing a second-order difference operation on the denoised PPG waveform to obtain a second-order difference PPG waveform: (3a) Performing reference point identification: Based on the denoised PPG waveform and the first-order differential PPG waveform, a differential threshold algorithm is used to identify the starting point A on the PPG waveform, and the PPG waveform is divided into multiple single-cycle PPG waveforms according to the starting point A and the next cycle starting point A1 adjacent thereto. Similarly, the second-order differential PPG waveform is also divided into multiple single-cycle second-order differential PPG waveforms; the differential threshold algorithm is used again to identify the reference points on the single-cycle PPG waveform, which are the starting point A, the slope point K, the main wave B, the descending gorge C, the heavy beat wave D and the next cycle starting point A1, and the reference points on the single-cycle second-order differential PPG waveform are identified, which are wave point a, wave point b, wave point c, wave point d and wave point e; (3b) According to the reference points identified in step (3a), features are extracted from multiple single-cycle PPG waveforms and single-cycle second-order differential PPG waveforms, and the value of each feature is the average value of the features extracted from multiple single-cycle PPG waveforms; the feature extraction methods on the single-cycle PPG waveform include pulse wave contour analysis method and pulse area method. The pulse wave contour analysis method uses 6 reference points, namely, starting point A, slope point K, main wave B, descending gorge C, dicrotic wave D and the starting point A1 of the next cycle to extract features. The extracted features include pulse period Period, pulse rate HR, acceleration time SlopeTime, maximum slope SlopeMax, peak time CT, peak ratio CTR, normalized peak ratio NCT, systolic time ED, diastolic time DiaTime, ejection fraction EF, peak-to-peak time PPT, sclerosis index SI, reflectivity index RI, left heart load index SPTI, myocardial perfusion index DPTI, and subendocardial myocardial vitality rate SEVR. The formula is as follows: Period=T HR=60 / T SlopeTime=T1 CT=T2 CTR=T2 / T NCT=T2 / H ED=T3 DiaTime=T4 EF=T3 / T PPT=T5 SI=H / T5 RI=h4 / h2 SPTI=S1 DPTI=S2 SEVR=S2 / S1 Wherein, S1 is the integral of the amplitude of the cardiac ejection period over time, S2 is the integral of the amplitude of the cardiac contraction period over time, H is the individual height, T is the cycle time of a complete contraction and relaxation activity of the heart, that is, a pulse cycle, T1 is the time from the starting point A to the slope point K, T2 is the main wave rise time from the starting point A to the main wave B, T3 is the time from the starting point A to the descending gorge C, T4 is the time from the descending gorge C to the starting point A1 of the next cycle, T5 is the time from the main wave B to the dicrotic wave D, h2 is the amplitude difference from the main wave B to the starting point A, h4 is the amplitude difference from the dicrotic wave D to the starting point A, SlopeMax is the amplitude of the first-order differential PPG waveform at the slope point K; The pulse area method uses two reference points and the overall pulse wave area change to extract features. The two reference points include the starting point A and the next cycle starting point A1. The extracted features include waveform characteristic parameters K1, cardiac output SV, cardiac output CO, peripheral resistance TPR, and arterial compliance AC. The formula is as follows: K1=(Q m -Q min ) / (Q max -Q min ) TPR=Q m / CO AC=SV / (Q max -Q min ) Among them, Q m is the average pulse amplitude of a single-cycle PPG waveform, T is the pulse period, Q min is the minimum pulse amplitude of a single-cycle PPG waveform, Q max is the maximum pulse amplitude of a single-cycle PPG waveform, dt is the integral of the single-cycle PPG waveform amplitude over time from 0 to T; The feature extraction on the single-cycle second-order differential PPG waveform uses five wave reference points, namely the amplitude ratios of wave points a, b, c, d and e. The extraction formulas of these features are: AMR1=b / a, AMR2=c / a, AMR3=d / a, AMR4=e / a, AMR5=(bc) / a, AMR6=(bcd) / a and AMR7=(bcde) / a, where a, b, c, d and e represent the amplitudes of the corresponding reference points respectively, and AMR1 to AMR7 are all amplitude ratios.

2. The pulse wave velocity detection method based on the terminal volume pulse wave signal according to claim 1, characterized in that: The step (1) specifically refers to: using a jacket device to collect original PPG signals from human body ends, the human body ends include fingers, toes and earlobes, while recording basic physiological information and health information, and using a pulse wave velocity detector to detect the pulse wave velocity; the jacket device is equipped with a photoelectric capacitance pulse wave sensor, the basic physiological information includes age, gender, height, weight and body mass index BMI, and the health information includes whether the person smokes and whether the person has a history of chronic diseases.

3. The pulse wave velocity detection method based on the terminal volume pulse wave signal according to claim 1, characterized in that: The step (2) specifically refers to: using a denoising method based on discrete wavelet transform to remove baseline drift and high-frequency noise from the original PPG signal to obtain a denoised PPG waveform; the denoising method based on discrete wavelet transform adopts a wavelet threshold denoising method, and the wavelet threshold denoising method includes wavelet multi-layer decomposition, wavelet threshold processing and wavelet coefficient reconstruction: (2a) Wavelet multi-layer decomposition: Perform multi-layer decomposition on the original PPG signal to obtain wavelet decomposition coefficients CD1, CD2, CD3, ..., CDn, CAn, where n is the number of decomposition layers; (2b) Wavelet threshold processing: Set the threshold λ to perform threshold processing on each layer of wavelet decomposition coefficients in step (2a) to obtain the processed wavelet coefficients; The threshold λ uses the VisuShrink threshold, and the formula is as follows: Where N is the length of each layer of wavelet coefficients, ln(N) is the logarithm of N with base e, δ is the estimated noise standard deviation, median|coeff| is the median of the absolute values ​​of the wavelet coefficients of the corresponding layer arranged in ascending order, and 0.6745 is the adjustment coefficient of the Gaussian white noise standard deviation; Threshold processing uses a soft threshold function processing method, the formula is as follows: Among them, sgn(n) is the sign function, f(x) is the processed wavelet coefficient; (2c) Wavelet coefficient reconstruction: Reconstruct f(x) to obtain the reconstructed signal as the denoised PPG waveform.

4. The pulse wave velocity detection method based on the terminal volume pulse wave signal according to claim 1, characterized in that: The step (4) specifically refers to: taking all the features extracted in step (3) as PPG features, and combining the PPG features with basic physiological information and health information as a feature set to be screened; using correlation analysis and Lasso regression to screen out features characterizing PWV, and using a tree-based decision mechanism, i.e., random forest, to screen out logical features, and the features characterizing PWV and the logical features are used together as a feature subset for PWV estimation; Correlation analysis was used to determine whether the feature was significantly correlated with PWV. If the P value of the analysis result was less than 0.05, that is, P<0.05, it indicated that the feature was significantly correlated with PWV and the feature was retained. On the contrary, if the P value of the analysis result was greater than or equal to 0.05, that is, P≥0.05, it indicated that the feature was not significantly correlated with PWV and the feature was removed. Use Lasso regression to compress the coefficient of the feature. If the regression coefficient of the feature is not 0, it means that the feature is important to the result and the feature is retained. If the regression coefficient of the feature is 0, it means that the feature is not important to the result and the feature is removed. Use the feature importance ranking in the random forest to select the top-ranked features. Among all the base decision trees in the random forest, select the base decision trees that use these top-ranked features, remove the base decision trees that do not use the top-ranked features, and use the decision logic of the selected base decision tree as a logical feature.

5. The pulse wave velocity detection method based on the terminal volume pulse wave signal according to claim 1, characterized in that: The step (5) specifically refers to: using a machine learning algorithm to establish a PWV estimation model, taking a feature subset as input, taking the actually detected PWV, i.e., the pulse wave velocity detected by a pulse wave velocity detector, as a learning target, training different PWV estimation models for feature subsets corresponding to different target objects; collecting the original PPG signal, basic physiological information, and health information of the object to be tested, processing them to obtain a feature subset, inputting the feature subset into the PWV estimation model, and the PWV estimation model automatically outputs an estimated PWV value, i.e., the pulse wave velocity.

Images