Optical blood pressure estimation based on PPG and FFT neural network [translation]

be based on ppg and fft Optical blood pressure estimation of neural network

Abstract

We introduced and validated light plethysmography using only fingertips (PPG) Beat by beat optical blood pressure of signal (BP) Estimation paradigm . The scheme determines the subject pair PPG The specific contribution of the signal , And most of its effects are removed by appropriate normalization . Extract key features such as amplitude and phase of cardiac components through fast Fourier transform , Training artificial neural networks , Then the neural network is used according to PPG It is estimated that BP value . from MIMIC II Database 69 Patients and 23 Volunteers verified . All the estimated results have a good correlation with the reference value . This method is fast 、 Good robustness , except BP It is estimated that , It can also be used for pulse wave analysis .

brief introduction

All over the world , Suboptimal blood pressure (BP) Is the leading risk factor for death , But at present, only one third of people with hypertension have blood pressure in [1]. One of the main reasons is the lack of proper monitoring and feedback on treatment . Hypertension patients with physical activity 、 Increased emotion and stress , Blood pressure changes rapidly , Therefore, patients with hypertension should continue to monitor blood pressure , Timely medical intervention , Avoid serious cardiovascular complications [2,3]. Compared with traditional auscultation of blood pressure , Continuous blood pressure is also considered a better predictor of cardiovascular events [4]. Unfortunately , Until today, , The most widely used non-invasive sphygmomanometer , Especially household sphygmomanometers , Still based on cuff , It interrupts normal blood flow , And it's too cumbersome , Do not wear it for a long time . Besides , This method only measures systolic blood pressure ( systolic pressure ,SBP) And diastolic pressure ( diastolic pressure ,DBP), Additional information [5] Co., LTD. .

Volume clamping method is a new non-invasive technique for continuous measurement of blood pressure . The principle is to keep the blood volume unchanged throughout the heart cycle by rapidly inflating and deflating the cuff on the finger . Then use cuff pressure to estimate blood pressure . This method is accurate and measures the pressure waveform , But the instrument is very expensive and difficult to wear . Finger cots also restrict blood flow , Wearing it for a long time will make you feel numb .

The flattening sphygmomanometer can be used continuously 、 Measure the whole circumference without cuff BP wave form . Through the appropriate transfer function , You can get the central blood pressure , Extract more diagnostic information [8,9]. But this method still needs to compress the radial artery , And the equipment is uncomfortable to wear . Besides , The flattening of arteries must be carefully controlled , To keep the vertical direction [10] Wall tension and constant deflection , This is very difficult in practice , Constant calibration is required . This method also has a potential deviation from the use of generalized transfer functions (GTF) of .

In recent years ,photoplethysmogram (PPG) measurement BP Showing great prospects . Studies have been reported BP And pulse transmission time (PTT) There is a negative correlation [11,12]. However , This method requires additional ECG measurements (ECG) Or BCG (BCG), And rely on a very complex arterial wave propagation model . Even if calibrated , Due to different physiological parameters 、 Dynamic characteristics and hydrostatic changes of human muscles ,PTT It is not always a reliable marker of pulsatile blood pressure [13,14]. therefore , Its use is limited to tracking BP Changes and qualitative analysis , For example, judgment BP Whether it is at a level worthy of attention .

But only use PPG Optics of signals BP Estimates have many practical advantages . First ,PPG And carotid pressure [15] The waveforms are similar . secondly ,PPG The device is cheap , Measurement is independent of the operator . This method has been studied in previous work [15 17]. for example ,Kurylyak Used by others PPG Waveform time information extraction 21 A feature to calculate SBP and DBP[16]. However , Their work is not reserved PPG Details of waveform , The choice of parameters is mainly empirical .S. C. Millasseau Et al. Reported a method based on fast Fourier transform (FFT) And generalized transfer function (GTF)[15] Of PPG normalization BP Waveform calculation method . They found that ,GTF It can be applied to almost all subjects , Even after taking nitroglycerin . But it and GTF There are also shortcomings , The main purpose is to compare the shape of the waveform , Instead of fitting BP The amplitude .John Allen Using linear and neural network system recognition techniques, et al BP and PPG wave form [17] Connect . They use Linear Autoregression of exogenous variables (ARX) Model extraction waveform features , And confirmed that PPG and BP The close connection between waveforms . But the initial calibration time of this method is long , And easily affected by age and other factors .Xiaorong Ding Et al. Proposed a hybrid method , take PPG Intensity information and PTT Combine the information to get BP, Compared with only PTT[18] There is an improvement . However , They just combined PPG Peak to peak amplitude information of waveform , Measurement still needs to be used ECG.

Our goal is to provide PPG Theoretical explanation of waveform , And design a method from PPG Accurate signal estimation BP The algorithm of , In order to develop a wearable 、 Successive 、 Equipment without cuff , People with hypertension can use this device to better manage their BP. The organizational structure of this paper is as follows : The first 2 Section gives PPG and BP The theoretical relationship between , The first 3 Section describes the data preprocessing steps and neural network structure , The first 4 Section gives the experimental results and discusses the future research directions , The first 5 Section gives the conclusion of this study .

Theoretical background

photoplethysmograph theory

Light plethysmography measures changes in blood volume by irradiating the skin and calculating changes in light absorption . Except for critically ill patients , The measuring part is usually on the finger . modern PPG The concept of measurement originated from Aoyagi[19] and Yoshiya[20], They put time variability PPG The signal is associated with changes in arterial blood volume , And assume that the venous blood does not pulsate .

Pulsating part of blood volume ( It is usually called the communication part ) It is mainly caused by the pulsation of blood pressure in the cardiac cycle . It is usually very small , And superimposed on a large non pulsating part ( DC part ). The normalization method is usually to divide the AC part by the DC part , Linearly scale the signal , Get a range of PPG wave form .

Arterial wall mechanics

In order to derive PPG and BP The relationship between signals , It is important to understand the elastic properties of blood vessels and how blood volume changes with blood pressure . The structure of the blood vessel wall is anisotropic , It has highly nonlinear elastic properties , So simple Hookean The model does not apply here [21].Langewouters Et al. Deduced the blood pressure and arterial radius of human fingers [22] The relationship between changes . However , The expression is complex , stay BP Estimate Application [23] Limited success in . In this paper , We will use a simpler formula to establish the relationship between the two

Arterial wall stress is generally caused by collagen 、 Elastin and smooth muscle [21] produce . The combined dynamic Young's modulus of elasticity of the two (E) Determines the stiffness of blood vessels . Many in vitro studies have been conducted to estimate E And cross wall pressure § The relationship between .Hughes Et al. Gave a widely used empirical formula [24

In the formula (1) in ,$E_0$ and $\gamma$ It depends on the location of the measurement and the specific animal . Canine aorta $\gamma$ The value is 0.017 0.004 mmhg 1[24].Bergel Et al Peterson etc. [25,26] Use different methods to calculate $E$, $E$ It can be approximately simplified to

here ,$\sigma$ It's Poisson's ratio , It is usually regarded as a constant . In some previous studies , hypothesis [25],$\sigma$ by 0.5, And will $E$ As incremental modulus . Because we only care about circumferential stress , So choose $E$ Indicates tensile strain modulus ,$\sigma$ Set to 0. Be careful ,$\sigma$ The choice of does not affect the general conclusion drawn in the next section .$H$ Is the thickness of blood vessel wall ,$r$ Is the average radius of blood vessels .

PPG Interpretation of signals

To simplify our model , We assume that PPG Signal and fingertip blood volume ($V$) Linear correlation , And the radius of artery in finger vascular bed is uniform $r$, Therefore, in the measurement area $V$ It can be written.

here $C$ It is a constant related to arterial vascular density ,$V_0$ It is related to microvascular and venous blood volume . hypothesis $V_0$ No pulsation , Combined (1)-(3), The following relation can be deduced ( See appendix )

here $b$ Is a large constant introduced by integral , And assume that $E_0$、$\gamma$ and $h$ irrelevant .$V$ It can be roughly approximated to the first order

Then normalize PPG The signal is expressed as

$V$ The time-varying component of is usually only a few percentage of the stationary component , It means $PP{G}_{norm}$ A very small . To help visualize the pulse waveform , The linear scaling function is usually used to normalize PPG The signal is amplified and shifted . The zoom function varies from manufacturer to manufacturer . In this study , We use $k_s$ As a scale factor ,$V_{off}$ As an offset factor . because $b$ It's big , To simplify the calculation , We use an improved scale factor $k$.$k$ For patients PPG Proper normalization of the signal is very important , This will be in the 3 Section describes .

Features and defects

Before normalization ,PPG The signal is affected by the elasticity of blood vessels E、 Blood volume V And so on , There are great differences between individuals . It's hard to find a universal rule that applies to everyone . After normalization , Of different subjects PPG The signal is comparable , Suppose the components of blood vessels are similar . Normalization also improves the effect of non ideal contact between the sensor and the finger , Such as eccentric placement , Because the DC component and AC component change almost in proportion .γ There may still be small changes between different populations , But the impact is relatively small .

However , Due to the assumptions used in our derivation , This method has limitations . An important assumption is , The blood composition ratio remains constant throughout the measurement , Especially those who have high absorption capacity for near-infrared wavelengths , Such as hemoglobin , This is true for most people, including people with hypertension . But in some cases , The patient has hemodynamic fluctuations , Such as severe edema or blood clots , This approach can fail , It needs to be used with caution . secondly , The vascular structure of finger artery is uneven , This reflects the distribution of vessel diameter . We only made the second-order approximation , When more accurate results are needed , This may cause considerable error . Besides , Different blood vessel composition will lead to different background absorption ratio , Lead to b The change of , So if only amplitude information is used ,BP The fitting error will be relatively large . therefore , We also need to combine BP Waveform shape information indirectly affected by horizontal [14,21], In order to improve the BP Fitting accuracy . Third , We did not consider light absorption from non vascular soft tissue , Pulsations from veins and PPG Slight pressure exerted by the sensor . Last but not least , Our theory is based on equality .(1)-(2), The properties of fragrant fruit and its main branches are well described , But the blood vessels applied to the fingertips may not be accurate . But fortunately , Because neural network is used instead of direct forward regression , These adverse effects can be effectively overcome in practice with the extensive training of the network .

Materials and methods

data source

In this study, free distributed mimiii (Multiparameter Intelligent Monitoring In Intensive Care) Data in the database .MIMIC II Collected more than 15000 name ICU Patient's multiparameter record , Including physiological signals and vital signs [27,28]. We checked 85 Complete electrocardiogram of patients 、 Blood pressure and PPG reading , At the same time, the sampling frequency is 125Hz. For these patients , Use the invasive method to measure continuously in the wrist BP, Use pulse oximeter to measure finger PPG. The length of the data varies from patient to patient , It depends on what they are ICU Time of stay . We excluded 16 The data quality of patients is poor (N = 7)、 Double wave pulse (N = 7) or BP-PPG The correlation is unstable (N = 2). Final ,69 Patients have more than 17 Ten thousand heartbeats are used to capture PPG And BP The relationship between

The data processing steps are as follows :

• Yes PPG and BP Data for calibration , To compensate for its phase lag .

• Select quality data according to established standards .

• Yes k Factors are properly normalized , Remove k Variance of factors

• Analyze the waveform data in frequency domain , Extract key features

Please note that ,MIMIC II Of PPG Record from a commercial pulse oximeter , After amplification and high filtering . The data of other resources need to be adjusted accordingly , To match the results of this article .

Data adjustment

mimicii Of BP and PPG The data are not completely consistent . To consider BP Signals and PPG Phase lag between signals , We calculate a cross-correlation function $g(\Delta t)$ To accurately align these two waveforms .

among g(Δt) Represents the time average of the cross-correlation function ,Δt by BP and PPG Phase lag between . The most accurate Δt Corresponding to the largest g(Δt).Δt Generally greater than 0, Less than the duration of a single pulse , Pictured 2 Shown . Our analysis shows that , Average Δt by 0.28s, The standard deviation is 0.04s.Δt May be used by drugs 、 Temperature change 、 Changes in posture and other effects , Recalibrate regularly .

chart 2 Shown .G (Δt) Yes 2 Patient's time . The phase lag varies from person to person .

Initial data selection criteria

BP and PPG The fidelity of data is the key to find the correct model . Abnormal or corrupted data may lead to undesirable results , For future analysis . because mimicii The data in the database comes from real critically ill patients , The sensor may fall off , Good contact may be lost occasionally . We need to find a way to filter the raw data , And use less artifacts and good internally connected data .

People have proposed various methods to evaluate BP The quality of the data . Such as Sun etc. [29] Adopt signal anomaly index (SAI) To assess the BP quality . We improved their methods , To get a correct BP and PPG Between models . First of all, SBP and DBP The signal uses Savitzky Golay Linear smoothing filtering , The data with two standard deviations from the smoothed value is considered too noisy and discarded . In this preliminary study , Systolic blood pressure is higher than 180mmHg Too little data , Cannot produce a correct fit . So they are not used . The initial selection rules are as follows ( surface 1).

surface 1.BP Data exclusion criteria

SBP: systolic pressure ;DBP: diastolic pressure ;ΔSBP = |SBP -SBP( smooth )|;ΔDBP = |DBP -DBP( smooth )|;T: Pulse width time ;std_sbp = ΔSBP Standard deviation ;std_dbp = ΔDBP Standard deviation of

PPG Signals tend to be more noisy . In order to evaluate their quality ,Silva Et al. Used multi-channel adaptive filters (MCAF) Method gives PPG Signal quality index of signal (SQI). The algorithm was originally designed to predict missing data , And depends on the measurement history .Li and Clifford[31] Using dynamic time warping technology will PPG The pulse is compared with the known excellent pulse , Get their own quality indicators . This study only focuses on PPG The signal , No other input . Because we have BP data , These data are usually very clean , We used something like Chen et al[9] A simple and effective way . They reported a correlation coefficient between radial pressure and central aortic pressure of 0.9. The sudden loss of correlation indicates that there is an error in intraocular pressure measurement . Again , In our case , Assume good PPG The signal should correspond to BP Highly correlated . For each corresponding PPG and BP Impulse linear regression . Coefficient of determination (R2) The threshold of is determined by experience as 0.8.

We also screened the double pulse data , This means that each cardiac cycle has two peaks . This may suggest aortic stenosis [32] And other diseases related to aortic insufficiency . In order to make our model simple , We exclude such data , But it will be realized in the future work .

chart 3 Shown .(a) clinical PPG Abnormal waveform (b) BP And anomalies PPG The correlation between waveforms is clinical PPG Good waveform (d) BP And good PPG Correlation between waveforms .

Standardization

Previous studies have found that , Hypertension patients and control group [33] Diameter of - There is no significant difference in the pressure curve , This means that Eq.(7) in ,γ Keep relatively constant . however , Due to the scaling factor of each patient k Different , So it's hard to get PPG And BP General correlation between , Pictured 4 Shown .

chart 4 Shown .PPG And radial BP The relationship between .

Results found , Measured by wrist BP Value and the number measured by the fingertip BP value [6] Is closely related to the . In this study , We assume that γ by 0.017, And make use of $pp{g}_{scale}$ And radius BP The relationship between all patients k Normalize . Through the least square error method , We found that k from 0.50 To 3.26 Change between , The average value is 1.86, The standard deviation is 0.5 ($N_{patient}$ = 71). When P The value is very high or close $P_{min}$ when ,$pp{g}_{scale}$ And $e^{-\gamma P_{min}}-e^{-\gamma P}$ The relationship between them will deviate from linearity , This may be caused by the reduction of the accuracy of the equation .(1)-(2), Or radial BP And figures BP Inherent differences .

Be careful , We use a commercial pulse oximeter PPG data , It is estimated that BP It is considered to be used beyond the instructions . In future research , Auto zoom should be disabled , To avoid k Errors caused by calibration . In our research , The evaluation without automatic scaling was also carried out , It turns out that in the second 4 Section shows .

PPG Feature extraction

On a large scale [34] in ,PPG The systolic amplitude of the signal has a good correlation with the local vascular expansion rate . In our research , Fast Fourier transform is used to extract amplitude and phase features from waveforms . This method does not need to identify the exact position of some feature points on the waveform , It can analyze multiple cardiac cycles at the same time , More stable than time-domain feature extraction . For beat by beat BP It is estimated that , A frame is defined as containing the previous period 10% And the next cycle 5% Complete heart cycle , Then use 0 To 150 Data points are filled .

This feature extraction method has certain limitations . When BP Change very fast , Features will be affected by adjacent beats , This will bring great error to the beat to beat estimation .

The architecture of artificial neural network

Artificial neural network is a kind of model inspired by biological neural network , It can be used to estimate highly nonlinear complex functions . in application , A multilayer feedforward back propagation neural network with a hidden layer is sufficient to fit a continuous function . The input of our artificial neural network is the amplitude and phase of Fourier components , Pictured 5 Shown . use Levenberg-Marquardt The algorithm trains the neural network .70% Data for training ,15% Used to verify ,15% Used for testing .ANN The structure is as follows

chart 5 Shown . be used for SBP and DBP Estimated neural network structure .

To test the effectiveness of our method , We first compare the output of artificial neural network with that of invasive measurement BP Compare , This is considered the gold standard . We also apply this algorithm to high-resolution pulse oximeter , And compare the output results with the commercial cuff sphygmomanometer , To determine whether the two methods are interchangeable .

experimental result

BP fitting (MIMIC II)

In order to verify the proposed method , We first studied PPG and BP Correlation between spectral components of . Pictured 6(a)-6(b) Shown ,PPG and BP The amplitude components of are highly correlated at all frequencies . Their phase correlation is better at low frequency , Poor at high frequencies , Pictured 6-6(d) Shown .

chart 6 Shown .BP Spectral components and PPG The relevance of (a)-(b) Amplitude correlation - (d) Related to .

It can be seen from the above analysis that , Only PPG Signals can be obtained SBP and DBP. However , because BP The waveform changes greatly , It is difficult to reconstruct accurately only by linear regression BP wave form .

Neural networks are often used to estimate complex nonlinear functions . Previous research found that ,8-11Hz The cut-off frequency of is BP Ideal frequency of waveform reconstruction [9,15]. In our case , A contain 35 The hidden layer of a neuron , The cut-off frequency is 10.8Hz The fitting network can be fitted stably and accurately at the same time SBP and DBP, Pictured 7 Shown . The increase in the number of neurons in the hidden layer brings a slight improvement (~0.2mmHg), But it is also easy to over fit the data , And introduce instability .

chart 7 Shown . The performance of neural networks ( Root mean square error ) And the number of hidden neurons and the cut-off frequency .

BP The accuracy of estimation requires BP and PPG There is a stable correlation between . If the heart load or medical intervention significantly changes the correlation during the measurement , The fitting accuracy will be reduced . In our research , The general neural network is not trained by patients with large fluctuation of correlation function .

To identify these outliers , We first train a single neural network for each patient ($N_{patient}$ = 71). Most of the installation errors are in the advanced medical device Association (AAMI) Within the scope of American National Standards , The maximum acceptable error is 5 8 mmHg, Pictured 8 Shown .71 Of the patients, there were 2 Causal factors BP And PPG Lack of stable correlation and excluded in general ANN Outside of training .

chart 8 Shown . Individual fitting error distribution .

Last , We use it 69 Patient data ( Include 175477 Waveform ) Trained the general neural network . Single beat to single beat fitting error SBP by 0.06 7.08mmHg, DBP by 0.01 4.66mmHg, It is also in line with the British Hypertension Society (BHS) agreement , The rating is A/A[35], As shown in the table 2 Shown . The fitting results are shown in the figure 9 Shown .

surface 2. Accuracy of blood pressure estimation . The grade is indicated in the standard 5 Mm Hg 、10 Mm Hg and 15 Cumulative percentage of readings in the range of mm Hg .

chart 9 Shown . General artificial neural network fits all data points and BP The data distribution .

If we don't care about shooting by shooting BP value , But average BP Seconds , The fitting error can be further reduced . for example , If adjacent 5 Average the waveforms , be SBP The fitting error can be reduced to 0.06 + 5.57mmHg, DBP The fitting error can be reduced to 0.01 + 3.69mmHg, Pictured 10 Shown .

chart 10 Shown . The average error is reduced

Considering that these patients are in intensive care , And possible medical intervention , This result is very promising .

But this method still has limitations , To be improved . for example , Need a ratio FFT More accurate models to capture PPG The basic waveform of . Data acquisition and preprocessing methods can be further optimized . The neural network structure needs to be fine tuned , First, classify the patients , Reduce the number of outliers . These works will be carried out in future research .

Test volunteers

To verify whether our scheme is reliable , We conducted additional tests on healthy volunteers . A total of 23 Two subjects , Age 26 ~ 34 year . Among them, men 18 example , women 5 example . First use high-resolution pulse waveform output PPG Evaluation module (AFE4490, Texas Instruments) obtain PPG The signal , The detector is located on the palm of the index finger . sampling frequency 500Hz. Use unscaled and unfiltered outputs to maintain k Constant , As formula (7) Shown . Subjects were allowed to rest for at least 5 minute , Then use Omron HEM-7051 Digital automatic upper arm blood pressure monitor measures blood pressure . All measurements are carried out under static conditions . Visual inspection PPG Quality of waveform , Then use another upper arm automatic sphygmomanometer (Microlife BP3GX1-5X) verification BP Measurement results . For each subject , Average 8 Seconds PPG Pulse to estimate systolic and diastolic blood pressure . Be careful , This comparison is to prove the equivalence of the two methods , Therefore, it cannot indicate the accuracy of any kind of equipment . Our results show that the difference between systolic and diastolic blood pressure is 1.67 2.46mmHg and 1.29 1.71mmHg. Of two measurements t There is no significant difference in the test (p<0.01), High correlation . We did it BlandAltman analysis , See if these two methods can be interchanged . The consistency limit is shown as 4.1mmHg, Pictured 11 Shown , This is acceptable in a clinical setting .

chart 11 Shown . be based on ppg Of BP Estimated with cuff sphygmomanometer Bland-Altman analysis .

Summarize and discuss

PPG Technology is wearable 、 Noninvasive and versatile , It can not only provide tissue oxygenation and heart rate information , It can also provide blood pressure information , It has the potential of early screening of vascular diseases [36 39]. However , Due to cardiovascular 、 Effects of respiratory and autonomic systems ,PPG Signal analysis is also complex . A comprehensive understanding of its morphology and basic physiological mechanism is still lacking .

We studied BP and PPG Internal connection between , And derive a simple mathematical formula to describe the relationship between them , It's better than using PTT The method is simpler , Suppose there's less . be based on FFT Feature extraction uses several key parameters to retain the amplitude and shape information of the entire waveform , Make analysis easier . We also use the power of neural networks to improve BP Fitting accuracy . After proper data processing 、 Normalization and training , This method has good performance , There is no need for complex and repeated calibration of each individual . Through observation BP and PPG Spectral composition of , It can also provide more information about patients' cardiovascular health

In this work , We used 69 Of patients 17 More than 10000 waveforms to train general neural networks , This is much larger than the population used in previous studies [15 18]. We have proved that our method is robust and accurate , In a considerable number of subjects . But this is by no means a thorough training and test . In the real world BP The distribution may be different from the training samples we use . for example , stay MIMIC II There are more than 15000 Famous patients , And we only randomly selected dozens of patients . Ideally , We should include the blood pressure data of healthy and unhealthy subjects into a broader reading range , But analyzing all this data will require more sophisticated computational tools .

Data quality is very important for the initial training and subsequent use of artificial neural networks in this method . about MIMIC II Data sets , According to the data in the same database BP Choose good information PPG Waveform segment , This enables us to easily and reliably obtain clean BP and PPG The signal , For neural network training and verification . However , For volunteer experiments , We must rely on PPG The signal itself determines and discards anything that is considered bad or unavailable PPG fragment . There? , A basic method , Compare the positive and negative standard deviations of the signal and the average signal using .

In the future work , We are committed to developing a more robust 、 More general data selection scheme , To evaluate PPG The quality of the signal , And filter out bad waveforms from good waveforms in real time . Besides , You also need to measure the numbers at the same time BP And radial BP, To establish a more accurate BP and PPG The relationship between . It's going on BP I guess before , We should also try to classify different diseases . The original... Will be used PPG data , Instead of the data processed by commercial pulse oximeters , To identify differences PPG- bp Reactive patients . for example ,PPG The morphological characteristics and low-frequency drift of waveform can be used to mark some diseases , Such as atrial fibrillation 、 Arteriosclerosis and hypovolemia . Do that , New technologies such as pattern recognition and recurrent neural networks may be helpful .

PG The quality of the signal , And filter out bad waveforms from good waveforms in real time . Besides , You also need to measure the numbers at the same time BP And radial BP, To establish a more accurate BP and PPG The relationship between . It's going on BP I guess before , We should also try to classify different diseases . The original... Will be used PPG data , Instead of the data processed by commercial pulse oximeters , To identify differences PPG- bp Reactive patients . for example ,PPG The morphological characteristics and low-frequency drift of waveform can be used to mark some diseases , Such as atrial fibrillation 、 Arteriosclerosis and hypovolemia . Do that , New technologies such as pattern recognition and recurrent neural networks may be helpful .