Estimate blood pressure according to PPG using spectral spectrum time depth neural network [turn]

according to PPG Estimate blood pressure using spectrum - Time depth neural network

Blood Pressure Estimation from Photoplethysmogram Using a Spectro-Temporal Deep Neural Network

Abstract

Blood pressure (BP) It is a direct indicator of hypertension , High blood pressure is a dangerous 、 Potentially fatal diseases . therefore , It is important to monitor blood pressure regularly , But many people don't like cuff equipment , Their limitation is that they can only be used at rest . In our research , Use only light volume diagrams (PPG) To estimate BP Is a potential solution . We analyzed MIMIC III High quality in the database PPG And arteries BP wave form , After being pretreated, more than 700 Hour signal , Belong to 510 Two subjects . then , We will PPG And its first and second derivatives as input , Input to a new spectrum with residual connections - Time depth neural network . We showed in an experiment with one subject , The network is able to PPG and BP Modeling dependencies between , The average absolute errors of systolic and diastolic blood pressure are 9.43 and 6.88. Besides , We have shown that , Personalization of the model is important , And fundamentally improved the results , It is difficult to derive a good general prediction model . We have disclosed the key parts of the research , In particular, the list of subjects used and our neural network code , To provide a reliable baseline , And simplify future research in explicit MIMIC III Potential comparisons on subsets .

brief introduction

Blood pressure (BP) Measurement is the most important , The most commonly used medical office test . Blood pressure is a direct indicator of hypertension , It is an important risk factor for a variety of cardiovascular diseases (CVDs) yes 2015 The most common cause of death in , According to the World Health Organization ( Who )[2] The data of , It has caused nearly 1500 Ten thousand deaths . therefore , Regular monitoring of blood pressure is important for the general population , But it is especially important for people who already have high blood pressure or related diseases , Because these people are particularly vulnerable to blood pressure [3] The effect of rising .

Even though it's important , But people have a general BP Monitoring is clearly disgusting . This can be attributed to the nature of measuring equipment to a large extent . Cuff based devices are still the gold standard , Usually recommended by doctors . These devices provide the highest measurement accuracy , However , They also have some disadvantages . People who use cuff type equipment must follow relatively strict measurement protocols , To ensure that the measured value is correct [4]. The measurement process can be tedious , It requires special time and energy . Physical activity ( Like sports ) It is usually not allowed to use cuff to measure blood pressure at the same time . Besides , Measuring the event itself will cause the subject's stress or anxiety , This in turn will affect the measurement BP value . This is often called white coat syndrome [5].

Because of the above factors , The focus of work in this field is to develop robust and unobtrusive BP Estimation system , It can provide users with near real-time periodicity BP to update . The catalyst for the development of this system is the emergence of more and more wearable devices , Like bracelets , It can collect countless physiological signals in a rather inconspicuous way . One of the signals that reflect the state of the cardiovascular system is the light volume diagram (PPG), We will describe in detail in the next section .

PPG background

Light plethysmography is a fairly simple and inexpensive technique , It is widely used in medicine to measure heart rate and blood oxygen saturation . It is based on the illumination of the skin and measures the changes in its light absorption . It usually consists of a light-emitting diode (LED) To illuminate the skin and a photodetector ( photodiode ) To measure through or from the skin [8] The amount of light reflected .

The change of tissue color and the absorption of light are controlled by the blood circulation in the body , Blood circulation is driven by the contraction and relaxation of the beating heart . When the heart contracts , The heart muscles contract , Blood flows around the body , So blood pressure will rise . The propagation of this blood pulse wave causes blood pressure to rise . Again , When the heart is full of blood during diastole , Blood pressure will also drop . This produces a periodic signal , It has a clear and definite systolic peak ( The first passage of blood when it is pushed out of the heart during systole ) And subtle peak diastolic ( Blood returns to the heart from the periphery during diastole ). Each periodically repeated cycle contains a systolic peak and a diastolic peak . The former is usually easy to detect , It can be used to determine the heart rate (HR), The latter is sometimes obvious , But more often it is difficult to detect correctly . Examples of ideal waveforms and common anomalies are shown in Figure 1 Shown .

chart 1 Shown . Ideal light volume diagram (PPG) The first and second derivatives next to the periodic waveform and its distorted waveform . Ideal examples are a large systolic peak and a low diastolic peak , And the anomaly has too many or too few peaks . All data comes from mimiciii database [11].

If you use PPG To measure the blood flow velocity , That is usually called pulse wave velocity (PWV)[9], that PPG and BP The relationship between them has been well confirmed . When blood vessels harden or contract more severely , Blood flows faster , The pressure is also greater . Again , When blood vessels are more relaxed or elastic , Blood flow is slower , The pressure is also less . The time required for the pulse from the heart to reach the peripheral point of the body is called pulse transmission time (PTT)[10]. When PTT Short time , It means high blood pressure , and PTT Longer means lower blood pressure . However , In order to use this method , Two sensors are required , It's usually an electrocardiogram on the heart (ECG) And points around the body PPG

Subtle changes in blood volume during cardiac activity , Reflected in PPG The shape of the signal , It also seems to be related to BP of . However , This relationship is currently modeled only by machine learning . The derivation of such a model is recent 1.3 The topic of intensive research discussed in section , It's also our job

BP Estimate of the PTT Method

1922 year ,Bramwell Et al. First studied PWV And vascular stiffness , And formalize it . Two copper tubes are connected to the artery , Mercury passes through the device in the form of waves . Mercury was chosen instead of actual blood because the arterial segment was very short , Then used for measurement PWV The time interval is also very short .

Geddes And other people in 1981 year [10] Preliminary analysis of arterial blood pressure (ABP) And PTT The specific connection between . They assessed 10 Pulse arrival time and diastolic blood pressure in anesthetized dogs (DBP) The relationship between . Use chemical methods to control blood pressure , Measure ECG signal R The peak in the carotid and femoral arteries corresponds to the time of systole .PTT and DBP The change of showed a good near linear correlation

Chan Et al. 2001 Created a sleeveless for telemedicine BP Simple system of estimation system [12]. An undetermined number of subjects participated in the laboratory experiment , Their ECG and PPG With 1500 Hz Sampling rate measurement , Their ground truth BP Calculate with a device with handcuffs . measurement PTT, Each subject 20 A course of treatment is used for calibration . After this calibration ,PTT and BP There is a linear relationship between . They used the average error between actual blood pressure and predicted blood pressure ( In millimeters of mercury (mmHg) Is the metric , Get systolic blood pressure (SBP) Of ME by 7.5, diastolic pressure (DBP) Of ME by 4.1. Use ME Evaluation may cause problems , Because in the overall average , Equally large positive and negative differences will offset each other , Shows a low overall ME, Even a single ME It's big .

PTT A recent example of this method is Kachuee Et al. 2017 year [13] Published , The scope of this method is similar to our own work .BP It's through mimicii About... In the database 1000 Estimated by subjects . Adopt pretreated PPG and ECG To obtain a ptt Relevant features , And ABP As the basic true value . use 10 Multiple cross validation (CV) Experimental evaluation of several regression models . Use Adaptive Boosting (AdaBoost) The best results have been achieved , especially SBP and DBP The mean absolute error of (MAE) by 11.17 and 5.35. They also use personal calibration in their work , This improves the results . Compared with the work mentioned above , The evaluation indicators of the report are more appropriate , And better insight into the results

The recent rise of deep learning , It has proved to be very effective in many fields , It is also reflected in this research field .Su Et al. 2018 year [14] Published a paper , It emphasizes the existing PPG BP The accuracy of the estimation model decreases in a long time . This problem means frequent calibration . Adopt a long-term and short-term memory (LSTM) Deep loop neural network (RNN) Yes BP Time dependent modeling . With PPG and ECG As input ,PTT And other features used to predict blood pressure . Compared with the existing method , their BP The prediction accuracy has been improved for a long time , At the same time, the real-time prediction accuracy equivalent to that of related work is maintained .

In the past 10 In the year ,PTT The method has been thoroughly tested [15,16], And put forward some slight changes , However , The basic principles have not changed . The disadvantage of this method is that it requires two sensors to measure PTT, This makes it more inconvenient than the single device method . This also means that two signals must be monitored 、 Pretreatment and analysis . Last , Because the same pulse wave must be tracked , Two ( Usually wireless ) Accurate synchronization and peak detection between sensors are crucial , It takes a lot of effort before processing begins .

single ppg BP Estimation method

The shortcomings mentioned in the previous section , And the emergence of more and more wearable devices , Gave birth to the use of single PPG New research on signals . The focus of this study is PPG Morphological analysis , With the help of the usual description of each cycle PPG Characteristics of waveform shape .

Teng Et al. 2003 year [17] Carried out using only PPG The signal goes on BP One of the earliest attempts to estimate . They analyzed ABP and PPG The relationship between certain features in the waveform . Data collected from 15 Young healthy subjects , Use professional equipment , In a highly controlled environment , Ensure constant temperature 、 No exercise and quiet . Feature selection through correlation analysis , And use linear regression algorithm to model the relationship . Systolic and diastolic ME Values, respectively 0.21 7.32 mmHg and 0.02 4.39 mmHg. For the reasons mentioned above , Use ME instead of MAE Once again, it becomes a potential problem . Besides , The sample size is very small , Conditions are highly controlled , So it is doubtful how these results can be translated into a more practical environment

Kurylyak Et al. 2013 year [18] Published an article on the use of neural networks from PPG In the middle of BP The paper with the highest estimated citation rate . They used MIMIC II A small amount of data in the database , First, we calculate 21 Features , These features describe a single PPG The shape of the period . Study these characteristics and BP The correlation between , Then these features are input into the artificial neural network (ANN) Training prediction model . They report systolic blood pressure MAE by 3.80 3.46, diastolic pressure MAE by 2.21 2.09. The result is very good. , But only used MIMIC A small unpublished subset of the database

Xing Et al. 2016 year [19] Published an important recent study . They used from MIMIC II Database 69 Patients and others 23 Volunteers . First, all signals are strictly preprocessed , To obtain high-quality waveforms . Different from previous work using time characteristics , They use fast Fourier transform (FFT) Extract amplitude and phase features from the waveform in the frequency domain . Input these features into the artificial neural network again , Good results were obtained , accord with BP Estimate the main criteria of the device . A special contribution is to standardize in patients PPG Wave proposal , It is said to eliminate the need for calibration .

We know , No direct use of the original PPG Signals as deep learning BP Estimated input , However ,Gotlibovych Et al. 2018 year [20] Studied the use of primitive PPG Data for atrial fibrillation detection , And achieved reasonable success , This implies the potential of raw signal input

In addition to the related work summarized , There are also a large number of other high-quality papers published in this field , But what they follow is to use PPG Signal to calculate periodic characteristics , Then use these characteristics to train regression ML Model , So they didn't discuss it alone .

We carry out relevant work in the context of research

Establish a good PTT Methods dominate this field , This is to be expected , Because description PTT and BP The underlying mechanism of correlation has been formalized and well tested . On wearable devices 、 The era of mobile health and telemedicine , pure PPG Methods do have many advantages , It also makes sense , But people are right about PPG There is less understanding and Research on signals , This urges us to focus only on PPG The signal .

In related work, it is common to use unavailable privately collected data sets or unspecified MIMIC Database subset . One side , For privately collected data , It's understandable , Because medical data is sensitive , Researchers who collect these data may not be able to obtain permission to share . On the other hand , When using publicly available data sets ( Such as MIMIC) when , You should specify a selected subset or process to get it . In view of the repeatability crisis prevailing in science [21], Using unpublished or publicly unavailable data may be problematic . This problem has also been encountered several times by the author of this article , Because the specific data used and the experimental code are almost never obtained from the author

In related work , It is also common not to compare with regression factors that always predict the mean value of training data . This is very important , Because there may be a basic reality of a selected subset of data BP The range of values is very small ( for example ,BP Subjects with relatively constant values ), under these circumstances , The error will be very low , But this does not necessarily mean that the model has learned much more than the predicted average .

This prompted our research from complete to publicly available MIMIC database ( edition III) Start , Reduce it to a subset of high-quality waveforms with clearly described filtering , This leads us to the final subset for training and evaluation . Because the database is huge ( Tens of thousands of themes , Each topic may have several recorded sessions ), It is also very suitable for use with deep learning , The latter can use the original signal as input , Derive your own characteristics . It's valuable , Because related work depends on a fairly standard set of features , These features require perfect waveforms , Heavily dependent on a single clear diastolic peak , And this is usually difficult to detect , Even missing . Besides ,[22] Some related work also shows , come from PPG First order of waveform (PPG) And second order (PPG) Additional features of are more difficult to detect consistently , Pictured 1 Shown .

We deal with the issues discussed in the following ways :

• Use precisely specified large MIMIC III Database subset , It contains available id And the corresponding code used to get it , as well as

• take PPG And its derivatives are directly input into a new spectrum - Time residual neural network , Successfully simulated PPG And BP The relationship between . As far as we know , The neural network architecture we proposed is by far the most complex in this field , Because it considers PPG The time and frequency information contained in the waveform and its derivatives . The details of the architecture will be described in later sections , The code of the model is available .

Materials and methods

As mentioned earlier , Our work is based on MIMIC III database [11,23], Can be found in https://physionet.org/physiobank/database/mimic3wdb/ Get on . The database contains 3 More than ten thousand 16 Various types of data recorded by patients aged or older during hospitalization . The median age of adult patients is 65.8 year , Men occupy 55.9%, Female occupation 44.1%. Each patient may have multiple records , From a few seconds ( It is usually an abnormal situation ) To a few hours . We are very interested in patients who record bedside waveforms , Especially those who have PPG and ABP Signal patients . Once the data is cleaned and preprocessed , First calculate the characteristics , Then train a classic ML Model ( Random forests ). Last , Take the raw data as the input to the depth neural network (DNN), To predict SBP and DBP.

Get and clean up raw data

Because the complete database is very large , There are some challenges in obtaining data initially and then selecting patients with high-quality waveforms suitable for machine learning . The advantage of such a large amount of data is , We can be very strict in the cleaning process , Because a lot of data will be left after each cleaning step .

We initially used customization DataMiner.sh bash Script download data . In order to download the required MATLAB Formatted data , Used WFDB software package [23], To be specific wfdb2mat function . A dedicated server is responsible for downloading data and saving it to local storage . We first noticed that some downloaded files were empty , Or include a very small number of examples . To save space , The size threshold determined by experience is 17 Kilobyte files are Cleaner.sh bash Script delete , Not even considering further cleaning steps . The script also removes the script that does not contain PPG and ABP Waveform file ( Specify in the file as PLETH and ABP). This step reduces us to about 1 Ten thousand patients .

after , We have carried out a more detailed cleaning procedure , Processing waveform quality . First , take PPG and ABP The minimum length required is set to 10 minute . Delete all short recordings , Because we want a waveform long enough , At least some SBP and DBP The change of . And then PPG The signal is normalized to zero mean unit variance , use 4 Butterworth bandpass filter filtering , The cut-off frequencies are 0.5 Hz and 8 Hz. lower than 0.5 Hertz can be considered as baseline wandering , And higher than 8 Hertz is high-frequency noise . then , utilize Hampel The filter filters the signal again , To remove outliers . This requires a containing 7 A follow-up PPG Sliding window of the sample , And calculate the median value of the window . Then it estimates the standard deviation of each sample with respect to the median window . If the sample differs from the median of the window by more than three standard deviations , Then use the median [24] Replace . choice Hampel Filter instead of median , Because it proved to be slightly better than signal processing [25] Some related work , Also because its effect has proved to be satisfactory visual inspection . It should be noted that , Although sometimes it can be in ABP Some high-frequency noise is observed in the waveform , However, these noises are not filtered or preprocessed in any way , Because it may slightly affect the signal value , And the signal value is SBP and DBP Of ground truth.SBP It's a sign of systole ABP cycle ( Or segment ) Peak value ( Or the average value of the detection peak ), and DBP It appears in diastole ABP Valley between cycles .

Then these signals are divided into periods , Each cycle corresponds to a heartbeat . utilize Elgendi wait forsomeone [26] The proposed peak valley detector is used for segmentation . The author proves that this algorithm is better than the traditional one for PPG The peak detector of the signal is faster 、 More effective , most important of all , More precise . In the initial empirical test , It works well on our dataset , But because of ground truth Peak value is unknown , It is difficult to make an accurate assessment . This can be manually segmented by experts through visual inspection PPG cycle , But such a job requires a lot

Split into cycles , Then the waveform is empirically evaluated . We plotted a large number of random records PPG and ABP Short random fragment of , This enables us to identify common post filtering problems in waveform morphology . These can be divided into two main categories

• Horizontal line : stay PPG and ABP A long time between normal cycles , Sometimes horizontal lines appear , Pictured 2 Shown . When three or more consecutive signal samples do not change their values , A flat line is detected . This horizontal line can be observed in several separate segments of the signal , We assume that they are caused by periodic sensor anomalies or sensor separation . These areas are useless , So it is cut from the waveform

  

chart 2. arterial pressure (ABP) The signal can be seen that the horizontal line is abnormal .

• Flat peak : Similarly ,ABP The waveform usually has flat peaks , The top part is missing , Pictured 3 Shown . take PPG Split into multiple cycles , Similarly detect peaks by checking whether three or more consecutive samples have the same value in a given period . The reason is still unknown , But it's probably a sensor problem .ABP The peak of is crucial , Because its value is SBP, This is what machine learning requires ground truth.

  

  chart 3. stay ABP Abnormal flat peak can be observed in the signal .

Same as before , Because of the amount of data , We applied a fairly strict standard , If exceeded 5% The period of has a flat peak , Or more than 10% The recording duration of consists of flat lines , Let's delete this record from the data set .

Last , Remove the remaining flat line or peak period from the waveform , Just remove the... Between the start and end points of the remaining horizontal lines PPG and ABP part , perhaps , If it is a cycle , Then cut off the whole cycle . chart 4 Summarize the whole process .

chart 4. Pipeline diagram of this system

Once the cleaning is completed ,510 Patients were retained , Each patient has at least one high-quality waveform record , Continuous at least 10 minute . Our goal is to create one that includes naturally occurring PPG Data set of complete changes in waveform ( Pictured 1 Waveform shown ), At the same time, avoid anomalies that are likely to be attributed to sensor problems ( This is very common in our data set ). On average, , After cleaning, the subjects had 5 Time 1.3 An hour's record , In all, it's close to 700 Hours of training data . Our final data SBP and DBP The distribution is shown in the figure 5 Shown .

chart 5. systolic pressure (SBP) And diastolic pressure (DBP) Distribution in our final data .

MIMIC The author of the database reported PPG and ABP There may be a slight delay between . Consider this , because BP In common cases ( Arterial bleeding is not included 、 Powerful drugs, etc ) There will be no substantial change , We take tradition ML In specific PPG Near the cycle ABP 2 The average of the peak and valley values detected in the short segment , So as to obtain basically real SBP and DBP. For those used in deep learning PPG fragment , because PPG and ABP Have very similar waveforms , So use the same peak detection algorithm for detection , Time consuming 5 s. Because the spectral layer in the neural network has enough periodicity to the signal , In order to extract meaningful information of frequency content , We choose a longer period for in-depth learning . Traditional machine learning is based on cycles , Therefore, the segment taken is shorter ( Ideally , Take only one PPG The quantity corresponding to the period , But in some marginal cases , It could happen ABP The systolic peak of was not captured , So I chose a longer length ).

Classic machine learning

As relevant work has shown , Traditional regression models can deal well with a set of widely used features , We conducted such an experiment , In order to compare with deep learning . We calculated a set of features commonly used in the related work discussed earlier .Yousef Others also hinted PPG and PPG[22] The potential usefulness of , So we also try to extract features . All common features describe each cycle PPG The shape of the waveform . therefore , The data is initially divided into multiple cycles , Points required for feature calculation ( for example , Contraction peak ) It's in the original PPG Period and its first derivative . Although the second derivative has also been proposed , But its waveform did not show the expected number of peaks , Therefore, it is impossible to use these features . surface 1 All the features used are given . With our previous work [27] comparison , We extend our feature set semantically by adding frequency-domain features . Most frequency domain features are represented by Welch ’ s Method [28] Got PSD To calculate the . meanwhile , We have removed some functions that have proved worthless in the past .

Domain	features
Time domain	$T_c$ Cycle duration
	$T_s$ From the beginning of the cycle to the peak of systole
	$T_d$ The time from the peak of systole to the end of cycle
	$T_{steepest}$.PPG '( Steepest point ) The time from the beginning of the cycle to the first peak
	$T_{dianortch}$. From the beginning of the cycle to the second peak PPG’ Time for
	$T_{SysToDiaNotch}$. Time from peak systole to fovea
	$T_{diaToEnd}$. The time from the heavy pulse to the end of the cycle
	ratio . Ratio of systolic and diastolic amplitudes

surface 1. According to the PPG Characteristics of calculation .

Domain	features
frequency domain	The three largest ranges . from PSD Three peaks of the largest magnitude are considered . These tell us the main frequencies in the cycle . In amplitude and frequency (Hz) For the characteristic .
	energy . Calculate as fast Fourier transform (FFT) Sum of squares of component sizes . Then normalize the energy by dividing by the period length .$ energy =\frac{1}{N}\sum_{n=0}^{N-1}
	entropy . Calculated as normalized FFT Information entropy of component size .$entropy=-su{m}_{n=0}^{N-1}x(n)log(x(n))$
	Statistical distribution . A normalized histogram , Essentially, FFT The size of is distributed to 10 In containers of equal size , Range from 0 Hz To 62.5 Hz.
	Skewness and kurtosis . These describe the shape of the period . More precisely , Skewness tells us symmetry , And kurtosis tells us flatness .

In an ideal situation , expect PPG The cycle has two peaks ( Systole and diastole ),PPG There are also two peaks . However , Our data is about PPG Characteristics of diastolic notch ,PPG The two peaks of are difficult to calculate , Because the diastolic notch is very subtle , Not even , Pictured 1 Shown . In these cases , Between peak systole and end of cycle PPG The point closest to zero (PPG The most gentle point of the signal ) It is used to calculate and process relaxation notch Characteristics of .

These features are eventually input into the use 100 It's a tree scikit-learn Default random forest (RF) In the implementation , And run one to keep one subject (LOSO) experiment , We will describe this experiment in detail in the following chapters .

Deep learning

To provide a complete representation of input data to neural networks , We used the original PPG、PPG and PPG As the time domain input of deep learning . Besides , The frequency domain information is obtained by using the spectrum diagram calculated from the time input segment . In order to use sonogram wisely , You need signal fragments that contain at least some periodic behavior . therefore , We divide the signal into 5 Seconds , Instead of what was used before, it usually lasts less than 1 A single cycle of seconds . Training network output value SBP and DBP, So as to solve the return task .

Neural network structure and hyperparameters

We propose spectral time ResNet Schematic diagram 6 Shown .ResNet The idea of was originally created by He wait forsomeone [29] Proposed , For training very deep networks with vanishing gradient problems , In this case , Backward error propagation is weakened , The first layer does not fully update their weights . therefore , A shortcut is proposed between larger layer blocks ( The remaining ) Connect .Su et al.[14] It also shows that , The remaining connections are PTT Of BP It is estimated that this particular area is also helpful .

For three inputs (PPG, PPG and PPG) Each of them , The time representation consists of each containing three convolutions (CNN) Residual block extraction of . Every CNN There is a batch processing layer behind each layer , To reduce the offset of internal covariates , And one. ReLU Activation layer , Compared with other activation , It can speed up the training process . Each remaining block eventually has an average pooling layer , For dimensionality reduction . The network also extracts and inputs specific spectral time information . Spectral layer extracts spectral information , For each of the three inputs 5s Input segment for spectrum calculation . The output of the spectral layer is used to gate the cycle unit (GRU), To get time changes .

then , Every pile 5 The output of the remaining blocks is connected and passed to another GRU layer . Again , spectrum - The output of the time block is also connected in series . The two data streams are connected at the last moment , And in two dense ( Complete connection ) Use in layer . The final output of the network is ReLu Provided by layer , It outputs the prediction of a given window SBP and DBP value . Use after each layer of spectrum GRU unit , In order to provide additional information about the time change of the spectrum for the subsequent dense layer . Experiments show that , This is the best network structure among the subjects .

The learning rate of the training model is $10^{-4}$, The decay rate is $10^{-4}$. The batch size is set to 256, Maximum training epoch Set to 20, When three epoch Stop in advance when you don't see the improvement . Network parameters include the number of residual blocks 、 Each piece of CNN The layer number 、CNN The size of the filter 、 The learning rate and batch size are determined by experiments , Additional optimizations can be made .

chart 6. Schematic diagram of our neural network architecture .

By propagating the gradient back to all layers , Fully supervise the training of the network . By using RMSprop The optimizer minimizes MAE Loss function to optimize parameters . After the initial training process , Use the best epoch As the starting point of personalization . In the process of personalization , We choose 20% Data of a given object to train our network additionally . Other data of the subjects were used for the final test .20% The data of has been proved to be enough to significantly improve the results , As shown in the table 2 Shown .

surface 2. Through the classics ML And the average absolute error achieved by deep learning (MAEs) With and without personalization .

Experimental apparatus

In order to train and evaluate a good general model , We did LOSO experiment . Such experiments are the most robust in terms of generalization performance , Because it is completely independent of people . stay n In the next iteration (n = Number of subjects ), In every iteration , Yes n 1 The data of subjects are used for training , The data of the missing subjects are used for testing . Each iteration calculates MAE, And give the final value in the form of average MAE.

Compared with the separation of typical training and testing , One advantage of this experimental setup is , The latter may be plagued by overly optimistic or overly pessimistic results , These results may be obtained by very specific separation , And in the LOSO Setting up , The result does not depend on the choice of separation .

In traditional methods , We set up separately SBP and DBP Model , The training neural network predicts both at the same time , There are two outputs .

The disadvantage of this experiment is that the calculation is very complicated , Especially in training ResNet when , Because the whole network must be trained n Time . Considering the amount and complexity of experimental data ,GPU Clustering is almost necessary . In our experiment , We used the include 4 individual nVidia 1080Ti 8GB The unit's GPU Server and usage nVidia Quadro P6000 24gb Combination of workstations of the unit . On the entire dataset described LOSO The experiment took a week

result

Compared with related work ,LOSO The initial independent results of the experiment are not good . therefore , Adopt the neural network structure and personalization described at the end of the hyper parameter section ( calibration ) Method , The results have been greatly improved , As shown in the table 2 Shown . In every time LOSO In iteration , The total error is expressed in single mae Calculate the average value of . Compare all results with a dummy regression factor , This regression factor always outputs SBP and DBP As a prediction .

Compared with the virtual regression factor , All models achieve lower systolic blood pressure error , Confirmed that PPG There are some problems in signals and their derivatives BP Information about , And the potential relationship can be modeled to some extent . However , Dumb regression shows low DBP error , because DBP Have lower variance and generally more stable . Personalization significantly improves results , indicate BP and PPG The relationship between them is subject dependent .

Based on the traditional manual features ML Method and the deep learning method based on the original signal as input , The latter is found to be more effective . The possible reasons are , It has stronger learning ability , Handmade features cannot be obtained from the original input DNN So much information , Or common features cannot be successfully used on data sets with large changes in imperfect waveforms .

Last , We compared using only PPG As our ResNet Input and use PPG And its first and second derivatives . It can be seen that , Derivation makes total MAE Reduced by about 1 mmHg, It is suggested that the derivative waveform contains additional useful information .

Discussion and conclusion

Our work is related to the 1.2 Section and section 1.3 See table for the comparison of related work discussed in section 3. in general , We found that due to different evaluation indicators and different and inadequately specified data sets , It is difficult to compare related work in this field . Besides , Some authors did not explicitly report whether the training data included data from any test subjects ( We represent this with unknown personalization in the table ). The smallest error is achieved on a small subset of public or private collections with unknown personalization , And the work of processing large-scale data (Kachuee et al.[13] And our work ) There are bigger mistakes , This implies the difficulty of creating a robust generic model on large datasets . The reason for this may be large MIMIC III The data set contains a large number of different subjects and potentially different PPG and ABP Measuring equipment

In data usage 、 The method and error are compared with the existing related work .

From a few related papers , And Chan wait forsomeone [12] and Teng wait forsomeone [17] Comparison is not feasible , Because they use ME As a measure .Su wait forsomeone [14] A modern LSTM Deep learning architecture , But with only PPG Signal to signal , Used PTT Method .Kurylyak wait forsomeone [18] Lower error is achieved , However, a considerable amount of unspecified data is used , That is to use 15000 pat , This is equivalent to 3 A little more than an hour ( Suppose the average cycle length is 0.8 s). They also did not specify how many subjects the data belonged to , Just the total number of shots , This means that the change and independence of the data are unknown . because Xing wait forsomeone [19] The minimum error of the report , We try to reproduce their work . We can't do this completely , Use a similar network architecture and input , Our mistakes are even bigger . Although with the use of PTT Compared with the method, only one is needed PPG The signal , But our DNN The results obtained are similar to Kachuee wait forsomeone [13] The results achieved are quite , This shows that our network architecture is very suitable for this problem . As far as we know , up to now , No work has been done using the original PPG And its two derivatives as inputs to the neural network , To successfully estimate BP. This is important , Because it shows , On a big data set , Use only PPG The signal goes on BP The assessment is feasible . Besides , The proposed neural network architecture allows the fusion of three inputs , And the frequency and time information of each input .

Except for the novel DNN Outside the architecture , Another contribution is to describe in detail the process of filtering data to obtain our subset , It also provides the final theme for training and evaluation . The author is also in https://github.com/gasper321/bp-estimation-mimic3 I have provided DataMiner.sh、Cleaner.sh and MATLAB Preprocessing code , And the definition of our neural network model Python Code , This allows researchers to further construct our work , And enable them to compare with each other . Be careful , This code is experimental , To study , Not robust , Nor suitable for production .

The main limitation is personalized needs , This means that the model needs some real data to adapt to each user . In practice , This means that users must use high-quality cuff equipment to measure their PPG and SBP/DBP, Then give these data to the model for personalization . Although this requires some effort , But this seems to be a reasonable cost , Can monitor BP, Without the user taking any action afterwards . Personalized computing needs in GPU Last is a few minutes , This is not a problem on the server , But at present, it may not be feasible in the monitoring equipment itself . As an additional limitation , Noisy data may be problematic , Especially for derivative waveform , Because their shape becomes very deformed

By the medical device Promotion Association 、 The validation standard for blood pressure measuring devices proposed by the European Society for hypertension and the international organization for standardization [ reference ] It is required to estimate the tolerable error (10 mmHg) The probability is at least 85%. The result without personalization cannot meet this requirement , But with personalization, the result is close ( The average error is tolerable , Although this is less than 85% Individual mistakes of are correct ). Through some modest improvements , Using a single PPG The sensor is personalized BP It is estimated that , Therefore, it can be accurately used for household use of providing information

be based on ppg Of BP It is estimated to provide greater comfort , And potentially improve users' compliance with this important measurement in the home environment . although MIMIC III The database does not provide information about getting PPG Information of signal equipment , But they may be fingertip oxygen saturation monitors . They are certainly suitable for one measurement , Sometimes this device is used for continuous monitoring of patients , Therefore, blood pressure can be monitored at the same time . Fingertip equipment is generally not suitable for outpatient use , Therefore, different devices must be used for unobtrusive continuous blood pressure monitoring . The bracelet is equipped with PPG The most common device for sensors , At present PPG The signal is poor , Although other sensors such as cameras [30] We are also exploring for this purpose . All in all , We believe that the results of this paper show , Using a single PPG The sensor carries out BP The estimate may apply to one BP measurement , And they are continuous BP Limited applicability of monitoring .

Harmony monitor . They are certainly suitable for one measurement , Sometimes this device is used for continuous monitoring of patients , Therefore, blood pressure can be monitored at the same time . Fingertip equipment is generally not suitable for outpatient use , Therefore, different devices must be used for unobtrusive continuous blood pressure monitoring . The bracelet is equipped with PPG The most common device for sensors , At present PPG The signal is poor , Although other sensors such as cameras [30] We are also exploring for this purpose . All in all , We believe that the results of this paper show , Using a single PPG The sensor carries out BP The estimate may apply to one BP measurement , And they are continuous BP Limited applicability of monitoring .