CN115836846A - Non-invasive blood pressure estimation method based on self-supervision transfer learning - Google Patents
Non-invasive blood pressure estimation method based on self-supervision transfer learning Download PDFInfo
- Publication number
- CN115836846A CN115836846A CN202211609290.0A CN202211609290A CN115836846A CN 115836846 A CN115836846 A CN 115836846A CN 202211609290 A CN202211609290 A CN 202211609290A CN 115836846 A CN115836846 A CN 115836846A
- Authority
- CN
- China
- Prior art keywords
- blood pressure
- signal
- signals
- invasive
- self
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)
Abstract
The invention provides a non-invasive blood pressure estimation method based on self-supervision transfer learning, which comprises the following steps: collecting photoelectric plethysmography signals and invasive continuous blood pressure waveform signals of a human body, constructing a noninvasive blood pressure estimation model based on self-supervision transfer learning, training the noninvasive blood pressure estimation model according to the photoelectric plethysmography signals and the invasive continuous blood pressure waveform signals of the monitored human body, collecting the photoelectric plethysmography signals of a user, inputting the photoelectric plethysmography signals into the noninvasive blood pressure estimation model, and performing noninvasive blood pressure estimation. The noninvasive blood pressure estimation method based on the self-supervision transfer learning provided by the invention can realize noninvasive blood pressure estimation through only photoplethysmography without a large amount of supervision data, and the prediction error meets the medical standard.
Description
Technical Field
The invention relates to the technical field of artificial intelligence, pulse waves and blood pressure waveforms, in particular to a non-invasive blood pressure estimation method based on self-supervision transfer learning.
Background
A large study in 2017 estimated that about 1040 million deaths worldwide were attributed to hypertension, while 31.1% of adults (about 13.9 million) had hypertension, an important risk factor for cardiovascular disease. Development of a low-cost continuous Blood Pressure (BP) monitoring technique can reduce the burden of care, effectively improve the prognosis effect, and thus reduce the mortality. The gold standard for continuous blood pressure monitoring is obtained by invasive arterial cannulation, but this method can result in pain and risk of infection. Therefore, the continuous non-invasive blood pressure monitoring technology is more suitable for health management outside Intensive Care Unit (ICU). However, oscillometric based non-invasive blood pressure monitoring techniques can dilate and occlude the vessel during the test, which is detrimental to long-term monitoring. Therefore, novel sensing technologies, such as PPG, have the advantages of low cost, low power consumption, high sensitivity, etc., and have great potential in recent methods and research for non-invasive continuous monitoring of blood pressure. According to the Pulse Wave Velocity (PWV) theory, a combination of a Ballcardiogram (BCG) or an Electrocardiogram (ECG) signal can also be selected.
Over the past decades, most studies have elaborated on the morphological features of PPG and constructed machine learning models to link PPG and BP values. These features cover statistical measures in the time domain, spectral transforms, morphological changes, etc. However, feature engineering often faces N-P challenges, and too many features can lead to input redundancy and even dimensional disaster problems. Conversely, fewer features may ignore coupling details that are non-linear in vascular dynamics. For these reasons, some techniques and research have begun to attempt to design depth models, with the advantage that the best features can be automatically learned to map from the PPG signal to the data-driven nature of the blood pressure. Some existing methods for estimating BP include convolutional neural networks, deep convolutional auto-encoders, autoregressive models, recursive neural networks, and U-Net and V-Net models from medical image segmentation. According to the existing research results, the blood pressure estimation method based on morphology hardly meets the worldwide accepted American Association of Medical Instruments (AAMI) standard (SDE should not exceed 8 mmHg).
Notably, most existing studies or methods use a randomly scattered paradigm to partition data, limited by the size of the data set, which can lead to optimistic results when a trained model encounters data from the same test individual during the test evaluation. Therefore, it is reasonable to segment the data set by individual subjects, and this paradigm of division by subjects helps to improve the generalization and robustness of the model.
On the other hand, it is very difficult, expensive and time consuming to collect expert annotation tags for subject data on a large scale. The overall performance of supervised learning is highly dependent on the size of the paired PPG-BP dataset, resulting in suboptimal performance for smaller marker datasets. Therefore, there is an increasing interest in unsupervised and self-supervised characterization learning to overcome learning problems in small data scenarios and to reduce the required cost of recording paired data. With the popularity of smart wearable devices, PPG signals are the most readily available physiological data. In self-supervised learning, the supervision comes from the input data itself. The most critical step is how to reasonably set the learning objective. However, most existing self-surveillance methods are derived in the context of natural images and video, and no study still involves unique mapping characteristics between PPG signals and corresponding blood pressure. Therefore, it is necessary to design a non-invasive blood pressure estimation method based on the self-supervised transfer learning.
Disclosure of Invention
The invention aims to provide a non-invasive blood pressure estimation method based on self-supervision transfer learning, which can realize non-invasive blood pressure estimation through only photoplethysmography without a large amount of supervision data, and the prediction error meets the medical standard.
In order to achieve the purpose, the invention provides the following scheme:
a non-invasive blood pressure estimation method based on self-supervision transfer learning comprises the following steps:
step 1: collecting photoelectric volume pulse wave signals and invasive continuous blood pressure waveform signals of a human body;
step 2: constructing a non-invasive blood pressure estimation model based on self-supervision transfer learning;
and step 3: training a noninvasive blood pressure estimation model according to the photoplethysmography signals and invasive continuous blood pressure waveform signals of the monitored human body;
and 4, step 4: collecting the photoplethysmogram signals of a user, inputting the photoplethysmogram signals into a noninvasive blood pressure estimation model, and performing noninvasive blood pressure estimation.
Optionally, in step 1, collecting photoplethysmographic signals and invasive continuous blood pressure waveform signals of a human body specifically comprises:
collecting a photoplethysmography signal, namely a PPG signal, and an invasive continuous blood pressure waveform signal, namely an invasive BP signal, of a human body, wherein the invasive BP signal is used as a reference signal.
Optionally, in step 2, a non-invasive blood pressure estimation model based on the self-supervised transfer learning is constructed, specifically:
the method comprises the steps of constructing a non-invasive blood pressure estimation model based on self-supervision transfer learning, wherein the non-invasive blood pressure estimation model comprises a data preprocessing module, a self-supervision learning module, a mode adaptation module and a blood pressure estimation module.
Optionally, a data preprocessing module is constructed, specifically:
the method comprises the steps of acquiring collected PPG signals and invasive BP signals, resampling the PPG signals and the invasive BP signals, removing noise after sampling is finished, removing outliers of the PPG signals after the noise is removed, normalizing the PPG signals after the PPG signals are removed, performing phase alignment processing on the PPG signals and the matched invasive BP signals after the processing is finished, ensuring that the phases of the PPG signals and the invasive BP signals are consistent, segmenting the PPG signals and the invasive BP signals after the phase alignment is finished, generating a BP mode adaptive data set and a BP estimation data set according to the segmented PPG signals and the invasive BP signals, and taking the PPG signals without the matched invasive BP signals as an auto-supervision data set.
Optionally, a self-supervision learning module is constructed, specifically:
constructing a self-supervision learning module and a corresponding target optimization function, acquiring a self-supervision data set, carrying out signal transformation on the self-supervision data set, inputting a transformed signal into the self-supervision learning module for signal reconstruction or signal transformation recognition, and comparing the reconstructed or transformed signal with a signal before transformation to realize pre-training of the self-supervision learning module;
wherein a PPG signal set in an auto-supervised data set is acquired
Where s represents the number of samples, the PPG signal set is signal-transformed by a signal transformation function G (·), i.e.:
finally obtaining the transformed signal
The specific signal transformation comprises Gaussian noise addition, power frequency noise addition, motion interference addition, baseline drift addition, respiratory tract sinus arrhythmia frequency modulation, random mask masking, hard clipping, amplitude inversion, time scale interchange and time scale scaling, wherein each signal transformation mode transforms PPG signals in each input self-supervision data set according to preset probability;
for PPG signal reconstruction, the signal recovery function is learned by the Transformer based encoder-decoder architecture for recovering the transformed signal to the original PPG signal, as:
wherein the content of the first and second substances,
for the reconstructed PPG signal sample set, MSE is defined as training loss @>
For PPG signal transform identification, the pseudo labels of signal variations are also classified by the Transformer encoder architecture:
in the formula, gamma i Is the loss factor of the ith conversion task, the corresponding loss
Comprises the following steps:
in the formula, p i e.P is the label automatically generated from the ith converted PPG, κ i Is the predicted probability of the model for that class.
Optionally, a mode adaptation module and a blood pressure estimation module are constructed, specifically:
constructing a mode adaptation module and a corresponding objective optimization function, wherein the mode adaptation module comprises a mode feature extractor G a And a mode discriminator D a In the pattern feature extractor G a And a mode discriminator D a A gradient inversion layer is added in between, and the gradient inversion layer performs constant transformation in the forward propagation processFrom D a To G a In the reverse propagation process of
Pattern feature extractor G a A mode discriminator D for the remaining transform encoder after removing the transform decoder from the supervised learning module a Is composed of PatchGAN, wherein, a pattern discriminator D a A pattern feature extractor G for distinguishing a feature representation between a target abnormal blood pressure pattern and a normal blood pressure pattern with a minimized pattern loss a And a mode discriminator D a Performing antagonistic learning through a game theory principle of a min-max optimization task:
wherein the loss L is resisted a Comprises the following steps:
in the formula, y to p target (Y) represents a target label, loss of pattern adaptive training L a (j) Comprises the following steps:
in the formula, y j Is the mode label x of the jth PPG signal j ∈R t×d ；
And constructing a blood pressure estimation module and a corresponding target optimization function, wherein the blood pressure estimation module consists of a Transformer encoder after two times of progressive pre-training. Specifically, after the pre-training task of the self-supervision learning module and the mode adaptation module is completed, the pre-trained encoder G is processed a Feature extractor G migrating to estimated BP values d In the above, accurate estimation of BP value is realized by transfer learning, and MSE loss is used as regression training loss L d ：
Optionally, in step 3, training the noninvasive blood pressure estimation model according to the monitored photoplethysmography signals and the invasive continuous blood pressure waveform signals of the human body, specifically:
the data after data preprocessing is acquired and divided according to the proportion of 7.5.
Optionally, in step 4, collecting a photoplethysmogram signal of the user, inputting the photoplethysmogram signal into a noninvasive blood pressure estimation model, and performing noninvasive blood pressure estimation specifically as follows:
collecting a PPG signal of a user, and sending a signal waveform as an input into a non-invasive blood pressure estimation model to obtain a blood pressure value corresponding to the user so as to realize non-invasive blood pressure estimation.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects: the invention provides a noninvasive blood pressure estimation method based on self-supervision transfer learning, which comprises the steps of collecting photoelectric volume pulse wave signals and invasive continuous blood pressure waveform signals of a human body, constructing a noninvasive blood pressure estimation model based on self-supervision transfer learning, training the noninvasive blood pressure estimation model according to the photoelectric volume pulse wave signals and the invasive continuous blood pressure waveform signals of the monitored human body, collecting the photoelectric volume pulse wave signals of a user, inputting the photoelectric volume pulse wave signals and the invasive continuous blood pressure waveform signals into the noninvasive blood pressure estimation model, and performing noninvasive blood pressure estimation; the input of the method only consists of photoplethysmography, so that the acquisition circuit only needs to acquire PPG signals, compared with the traditional method based on the propagation speed of the pulse wave, the method omits the step of acquiring electrocardiosignals, does not need excessive derivation calculation and characteristic engineering, is conveniently integrated into devices such as a bracelet and the like, does not need blood pressure measurement devices such as a cuff and the like, breaks away from the constraint of the cuff, and enables the devices to be more portable; the method can realize continuous blood pressure estimation and long-term blood pressure monitoring, can be used for measuring the blood pressure in daily life, and can not bring the influence of trauma and discomfort to human bodies during measurement; the method does not need calibration and is convenient to use; the method adopts an automatic supervision transfer learning framework, can carry out a representation learning model of fine-grained BP prediction from 11 converted PPG sequences, can promote the model to utilize more discriminative characteristics in the learning process, and has richer information contained in input signals, so that the measured blood pressure result is more stable, and higher prediction precision is achieved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a schematic flow chart of a non-invasive blood pressure estimation method based on self-supervised transfer learning according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of an experimental design flow of a non-invasive blood pressure estimation method based on self-supervised transfer learning according to an embodiment of the present invention;
FIG. 3 is a schematic structural diagram of a noninvasive blood pressure estimation model;
fig. 4 is a schematic diagram of a signal conversion method.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a non-invasive blood pressure estimation method based on self-supervision transfer learning, which can realize non-invasive blood pressure estimation through photoplethysmography, improve the estimation precision and provide reference for blood pressure ill detection.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
The invention adopts a Self-supervision transfer learning framework based on a Transformer, which is named as STP (Self-powered transferred based on Transformer for non-invasive blood Pressure estimation) and is used for continuous intelligent non-invasive blood Pressure estimation, and signal characterization learned from a large amount of unlabelled PPG data can be used for establishing a steady learning target. Specifically, the STP model provided by the invention inputs a pre-training model through randomly selecting 11 sequence transformations of the self-supervision PPG signal to learn a valuable signal representation of a low-dimensional embedded feature space, and then an encoder structure of the pre-training model is transferred to two progressive transfer learning sub-modules for fine tuning. The first migration weight is represented by a coarse-grained class that learns BP in a manner that counters optimization. The framework provided by the invention shows high-precision BP estimation performance, in addition, paired data with different proportions are used, the invention proves that the learning representation of the framework provided by the invention is superior to a supervised deep learning method optimized for monitoring a fine-grained BP value, the invention does not make any pre-assumption on a PPG signal for training, and trains the model through a main body paradigm, so that the STP model provided by the invention can be accurately generalized to individual data which does not exist in a training set.
As shown in fig. 1, the noninvasive blood pressure estimation method based on the self-supervised transfer learning according to the embodiment of the present invention includes the following steps:
step 1: collecting photoelectric volume pulse wave signals and invasive continuous blood pressure waveform signals of a human body;
step 2: constructing a non-invasive blood pressure estimation model based on self-supervision transfer learning;
and step 3: training a noninvasive blood pressure estimation model according to the photoplethysmography signals and invasive continuous blood pressure waveform signals of the monitored human body;
and 4, step 4: collecting the photoplethysmogram signals of a user, inputting the photoplethysmogram signals into a noninvasive blood pressure estimation model, and performing noninvasive blood pressure estimation.
In step 1, collecting photoplethysmographic signals and invasive continuous blood pressure waveform signals of a human body, specifically comprising the following steps:
collecting a photoplethysmography signal, namely a PPG signal, and an invasive continuous blood pressure waveform signal, namely an invasive BP signal, of a human body, wherein the invasive BP signal is used as a reference signal;
the invention adopts the public MIMIC III data set, WESAD data set, PPG-DaLiA data set and private data set, and the four data sets are trained and verified, and the four data sets are introduced respectively:
1) MIMIC (Multi-parameter Intelligent Monitoring in Intelligent Care) dataset. This data set was provided by the bess israel medical practice centre and philips medical, and 67,830 records including electrocardiogram, blood pressure, respiration and PPG waveforms were collected from nearly 3 million ICU patients. Since the signal record of the data set is incomplete, the invention finally screens the paired PPG-BP data of 200 patients for the supervised mode adaptation task and the BP estimation task. However, 300 patients with an absence of invasive blood pressure signals were used as PPG signal recordings available in the aforementioned tasks.
2) WESAD data set. The wearable device collects WESAD data sets for detecting and differentiating emotional states (neutral, stress, entertainment). The data set recorded physiological data for about 100 minutes each for 15 subjects (24-35 years old) via a wrist-worn device (including PPG, accelerometer, electrodermal activity and body temperature) and a chest-worn device (including ecg, accelerometer, electromyogram, respiration and body temperature). Therefore, this data set does not include a blood pressure signal, and is only used for the pre-training task.
3) PPG-DaLiA dataset. It is a multi-modal dataset used for motion compensation and heart rate estimation in daily life activities. The data set recorded by the wrist-worn device (PPG, three-axis acceleration, electrodermal activity and body temperature) and the chest-worn device (electrocardiogram, respiration and three-axis accelerometer) included 15 subjects (21-55 years old), each approximately 150 minutes. This data set does not include a blood pressure signal and is only used for the pre-training task.
4) A private data set. The inventive method collected data sets of PPG, ECG and invasive blood pressure recordings from 683 ICU patients from 7 different hospitals. The private data set has an electronic health record of the patient anonymously processed including demographic information, medication history, complications, and the like. The data sets are high-quality paired PPG-BP signals and are very suitable for supervised learning tasks.
The non-invasive blood pressure estimation method based on the self-supervision transfer learning can be regarded as a time series regression problem, and the space X of the original input PPG signal belongs to the R t×d×s And the corresponding target BP space Y ∈ R 2×s Giving paired samples
Precisely, for a data set comprising pairs of s segments, the input PPG signal X comprises t time steps (equivalent to the sampling rate) and d dimensions, where the input dimensions can be chosen among PPG, VPG and APG, and even other physiological signals, such as ECG, whose corresponding blood pressure likewise comprises values in both the SBP and DBP dimensions, the proposed STP model of the invention learns a mapping from the PPG signal X to the corresponding BP value y:
F:{x}→y (1)
which can be described by minimizing the Mean Square Error (MSE) between the reference BP and the output.
As shown in fig. 3, in step 2, a non-invasive blood pressure estimation model based on the self-supervised transfer learning is constructed, specifically:
the invention provides a process taking an STP model as a core, and the performance of BP estimation is improved by a mode-adaptive transfer learning model, wherein the STP model consists of two gradual transfer learning processes. The first weight migration is through optimization in antagonismLearning the coarse-grained class characterization of the BP in the process. Subsequently, the coarse-grained BP category representation is migrated to the fine-grained BP estimation sub-module to improve its discriminative power. Accordingly, the present invention defines three groups of tasks, the pre-training task, the adaptive pre-training and the downstream blood pressure estimation task, which are labeled as T respectively p 、T a And T d Wherein T is p Are PPG signal reconstruction or conversion identification tasks by which unpaired PPG representations, T, are learned a Is a PPG signal pattern recognition task, T d The BP value is estimated through learning, so that the non-invasive blood pressure estimation model comprises a data preprocessing module, an auto-supervision learning module, a mode adaptation module and a blood pressure estimation module;
a data preprocessing module: denoising and normalizing a clinically obtained PPG signal and a paired invasive BP signal thereof to form an automatic supervision data set, a BP mode adaptation data set and a BP estimation data set;
an automatic supervision module: the source model in the migration learning framework designed by the invention is trained by using a large number of unlabeled PPG signal sets so as to learn physiological characterization meaningful to photoplethysmography signal information and improve the characterization capability of hemodynamic information;
a mode adaptation module: the method plays a transitional role in the constructed STP architecture, and aims to enable the signal characterization capability learned by an automatic monitoring module to be better adapted to the discrimination of the BP mode of the coarse-grained category;
blood pressure estimation module, BP estimation module: the method is a target model established on a pre-training model of the two sub-modules, and better realizes a fine-grained BP estimation task by using knowledge learned from unmarked source data.
Constructing a data preprocessing module, which specifically comprises the following steps:
the present invention considers that unsupervised data sets are collected from different devices, the activity states of subjects are very different, more motion interference may exist in the data collected by wearable systems, and furthermore, a major challenge of clinical data sets such as MIMIC is that the record length of each patient is different and there is a great amount of missing data, therefore, the model construction needs to fully consider the noise interference and individual difference in signals, wherein the specific data preprocessing process is as follows:
acquiring the acquired PPG signal and the invasive BP signal, and resampling the PPG signal and the invasive BP signal:
the experiment of the invention adopts a plurality of data sets, and the sampling rates of the signals recorded in different databases are greatly different, for example, the sampling rate of the PPG signal in the PPG-DaLiA data set is 64Hz, and the sampling rate of all the signals in the MIMIC data set is 250Hz, so that the invention resamples all the signals to 125Hz in order to facilitate the uniform processing of the data flow;
after the sampling is finished, removing noise of the sample:
the frequency and amplitude of the PPG signal are low, and the noise is mainly baseline drift and motion interference, therefore, the invention uses the wavelet based on db8, decomposes the PPG signal into nine levels by discrete wavelet transform, and extracts approximate coefficients representing low-frequency components and detail coefficients representing high-frequency components, herein, the invention sets CA of the first stage and CD of the 7 th to 9 th stages as zero, and reconstructs the signal according to the decomposition coefficients to remove the baseline drift and a small amount of high-frequency noise in the PPG signal, and for the invasive BP signal, band-pass filtering (0.5-35 Hz) can be carried out by a finite impulse response filter to eliminate the power frequency interference and baseline drift of 50-60Hz in the BP signal;
after the noise removal is finished, performing outlier removal on the noise:
due to the complexity of the multi-source data set and unavoidable interference in the actual environment, there are still many abnormal parts in the filtered signal, including horizontal line signals (missing signals), BP values that are extremely low or even less than zero, and noise interference that overwhelms the original signal, which cannot be filtered out by the filter, and accordingly, the missing signals can be removed by finding the corresponding first order difference value, the invention sets a minimum threshold (for example, DBP is 25 mmHg) to remove abnormal BP values, and removes excessive abnormal waveforms by creating an average template;
specifically, for segments belonging to the same section (considered as the same patient), calculating the average difference and the correlation coefficient of the PPG segments to obtain an average template of the section, and finally, segments with too large deviation or too low correlation coefficient are considered as abnormal signals to be filtered out, and the threshold value of the filtering is set according to 3 times of standard deviation of normal distribution;
after the removal is finished, carrying out normalization processing on the PPG signal:
since the amplitude of the PPG signal is affected by the ambient light intensity, the intensity of the light emitted by the sensor, the sensitivity of the sensor photodiode, the blood oxygen concentration, the venous volume, the skin color of the subject, and other factors, the present invention uses a normalization approach to reduce individual differences, scaling the PPG signal over the [0,1] interval;
after the processing is finished, the PPG signal and the matched invasive BP signal are subjected to phase alignment processing, so that the phases of the PPG signal and the invasive BP signal are consistent:
in MIMIC datasets there may be delays of up to 500ms between the PPG signal and the invasive BP signal, therefore, in downstream tasks of BP estimation, the invention needs to ensure that the phases of the PPG signal and the invasive BP signal are identical, where the maximum correlation coefficient between the PPG signal and the invasive BP signal is used as a reference point, and the PPG signal is shifted to make the phases of the two signals identical;
after the phase alignment is finished, segmenting the PPG signal and the invasive BP signal:
in order to obtain successive BP values, it is necessary to detect the positions of the peaks and valleys of the PPG and BP signals, so as to segment them. There are various methods to divide the signal, for example, beat by beat and fixed length, because of the particularity of the signal in MIMIC, there may be time delay as high as 500 milliseconds between the signals, the beat by beat division method will result in incorrect matching and large error, the division method of the fixed length is easy to operate, but it will often destroy the periodic characteristic of the paired signal, therefore, the invention adopts the multicycle sliding division method to the paired signal, namely each segment of the divided signal contains multiple cycles, there is some period overlap between segment and segment, considering the sampling rate of the data, the invention sets the period number of each segment as 5 here, there is two period overlap between segment and segment, in addition, because the signal has changed the time domain after the data enhancement, the invention has adopted the fixed length segmentation to the PPG signal after conversion;
generating a BP mode adaptation data set and a BP estimation data set according to the segmented PPG signal and the invasive BP signal, and taking the PPG signal without the matched invasive BP signal as an automatic supervision data set:
the present invention requires three subsets to be set up to accommodate the respective tasks, most studies use sample-based randomly shuffled dataset partitions, however, this paradigm does not take into account the similarity of samples within individuals, which may lead to identical patient datasets between datasets, a simple example is that during the test, the model is likely to encounter individual data fitting in the training phase, which may lead to overly optimistic results, the present invention uses a subject data partitioning paradigm to avoid the risk of such data leakage, specifically, generating a BP mode adaptation dataset and a BP estimate dataset from the segmented PPG signal and the invasive BP signal, wherein the BP mode adaptation dataset accounts for 70%, and using the PPG signal without the matched invasive BP signal as the self-supervision dataset.
In particular, the mode label in the BP mode adaptation data set is formulated according to AAMI standards, specifically, if SBP ≧ 130mmHg and/or DBP ≧ 80mmHg, the present invention classifies the blood pressure value as high blood pressure, if SBP <90mmHg and/or DBP <60mmHg, low blood pressure, otherwise normal blood pressure, and furthermore, the AAMI standards require that the distribution of invasive blood pressure data should satisfy the following conditions: consisting of at least 150 measurements of 85 subjects, 10% of SBP was above 160mmhg,10% was below 100mmhg,10% was above 90mmhg,10% was below 60mmHg.
The method comprises the following steps of constructing an automatic supervision learning module, specifically:
constructing a self-supervision learning module and a corresponding target optimization function, acquiring a self-supervision data set, carrying out signal transformation on the self-supervision data set, inputting a transformed signal into the self-supervision learning module for signal reconstruction or signal transformation recognition, and comparing the reconstructed or transformed signal with a signal before transformation to realize pre-training of the self-supervision learning module;
wherein a PPG signal set in an auto-supervised data set is acquired
Where s represents the number of samples, the PPG signal set is signal transformed by a signal transformation function G (·), i.e.:
finally obtaining the transformed signal
In the course of the signal conversion, the invention defines that the input PPG signal->
Paired tuples +>
And their corresponding transformed pseudo labels P e R s As shown in fig. 4, the specific signal transformation includes:
gaussian noise addition: this transformation adds random noise of gaussian distribution of amplitude a and frequency θ to the original PPG signal X with probability ρ;
adding power frequency noise: adding power frequency interference with amplitude alpha and frequency of 50Hz to an original PPG signal X by probability rho;
adding motion disturbance: adding motion interference noise with amplitude alpha and frequency theta into an original PPG signal X with the duration of 1-10 percent;
addition of baseline drift: a baseline drift noise of 0.05Hz is added to the original PPG signal X with a probability ρ;
performing respiratory sinus arrhythmia frequency modulation: the transformation adds a sine modulation signal with the mean value of 1 and the amplitude of 0.05Hz to the original PPG signal;
and (3) carrying out random mask: the transformation is performed for each PPG sample x j ∈R t×d Randomly generating a binary mask sequence m j ∈R t ×d The converted sample is obtained by multiplying the two
The present invention does not use the conventional bernoulli distribution to randomly generate a mask sequence, considering that the conventional mean interpolation method can easily recover the masked signal. The present invention assumes the value t m The average length of the mask sequence generated for the scale is p. Therefore, the average length of the unmasked sequences is t m (1- ρ)/ρ, where both sequences are geometrically distributed parameters;
hard clipping is performed: the PPG sensor can adjust the driving current according to individual difference, so that the amplifier can amplify the PPG signal in the whole positive and negative offset, unreasonable direct current bias can cause clipping at the wave crest and the wave trough of the obtained PPG signal waveform, and the increased clipping times are defined as beta;
and (3) amplitude overturning: the transformation inverts the amplitude of the original PPG signal, i.e. the amplitude scaling factor is-1;
and (3) carrying out time scale turnover: the transformation reverses the temporal order of the original PPG signals, e.g., for the jth PPG sample x j ∈R t×d In the time sequence of
And its changed time sequence is->
Time scale interchange is carried out: the transformation randomly decomposes and recombines the individual sub-signal segments of the original PPG signal, e.g. for the jth PPG sample x j ∈R t×d Can be equally dividedDivided into ζ sub-signal sections, i.e.
Thus, a processed PPG signal->
Can be expressed as
Wherein m =1,2.., ζ, the sequence is randomly broken up;
performing time scale scaling: the original PPG sequence is randomly resampled, i.e. stretched or compressed in the time direction, for a scaling factor β and the jth PPG sample x j ∈R t×d Transformed PPG sample
Will be resampled to
Each signal transformation mode transforms the PPG signal in each input self-supervision data set according to a preset probability;
for PPG signal reconstruction, the signal recovery function is learned by the Transformer encoder architecture for recovering the transformed signal to the original PPG signal, as:
wherein, the first and the second end of the pipe are connected with each other,
for the reconstructed PPG signal sample set, MSE is defined as training loss @>
For PPG signal transform identification, the pseudo labels of signal variations are also classified by the Transformer encoder architecture:
in the formula, gamma i Is the loss factor of the ith conversion task, the corresponding loss
Comprises the following steps:
in the formula, p i e.P is the label automatically generated from the ith converted PPG, κ i Is the predicted probability of the model for that class.
In particular, the transform-based encoder-decoder architecture is illustrated by the self-supervised characterization learning sub-module in fig. 3. The specific data flow is as follows: the model inputs PPG signals which are subjected to signal conversion in an automatic supervision set
Mapping the characteristic vector into a one-dimensional convolution layer, then stacking the characteristic vector with a position code to mark time sequence information, inputting the characteristic vector with the time sequence information into a transform encoder-decoder to reconstruct a PPG signal before transformation, and recording the output as ^ er>
Wherein the transform encoder passes each sample point in the input transformed PPG signal in parallel. Position embedding is employed and position information is added to provide timing information for the context. Correlations with other timing information are then learned through the multi-head attention layer, thereby generating a plurality of attention vectors. These vectors are then averaged and a normalization layer is applied to simplify the optimization. These vectors are in turn passed to a feed-forward network that converts the signal values into dimensions that can be learned by the next encoder or encoder-decoder attention layer.
Wherein, the structure of the Transformer decoder is similar to that of the encoder. The attention layer with masking is passed to learn attention between the current sample point and all sample points in its previous timing in the output signal and not allow for upcoming sample points. Then, normalization operation is carried out through a residual connecting and normalizing layer, the output of the encoder layer is used as a key and a value vector to the next attention layer, and the next layer of the decoder uses the attention value as a query. This self-attention mechanism enhances the information interaction between the input and output PPG signals, thereby making the algorithm better understand the signal reconstruction task.
The method comprises the following steps of constructing a mode adaptation module and a blood pressure estimation module, and specifically comprises the following steps:
the method comprises the steps of constructing a mode adaptation module and a corresponding target optimization function, wherein the mode adaptation module comprises a mode feature extractor and a mode discriminator, the mode feature extractor is a Transformer encoder which is arranged in an auto-supervised learning module and is used for removing a Transformer decoder, and the mode discriminator consists of PatchGAN. Also, the present invention employs a pattern antagonism optimization method to achieve adaptive learning, specifically, in a pattern feature extractor and patternA Gradient Reverse Layer (GRL) is added between the discriminators, and the GRL performs constant transformation in the forward propagation process and simultaneously performs D a To G a In the reverse propagation process of
The training mode discriminator distinguishes the characteristic representation between the target abnormal blood pressure mode and the normal blood pressure mode by minimizing the mode loss, and the mode characteristic extractor and the mode discriminator carry out antagonistic learning by the game theory principle of the minimum-maximum optimization task:
wherein the loss L is resisted a Comprises the following steps:
in the formula, y to p target (Y) represents a target label, loss of pattern adaptive training L a (j) Comprises the following steps:
in the formula, y j Is the mode label x of the jth PPG signal j ∈R t×d ；
The invention removes a Transformer decoder from a self-supervision network, takes the other Transformer encoders as feature extractors of a BP mode, converts a final layer of feature graph into a feature vector with 1 xN dimension by utilizing a Global Average Pooling (GAP) layer, and sends the feature vector into a discriminator consisting of PatchGAN, thereby increasing a perception domain of discriminant loss, wherein the specific structure is shown in figure 3;
in particular, the mode adaptation module is formed by a mode feature extractor G a And a mode discriminator D a Form, as illustrated by the mode-adapted block diagram in FIG. 3。
Wherein, the pattern feature extractor G a The transform encoder is the same as the transform encoder, and the weight of the transform encoder is obtained by pre-training of an unsupervised learning module.
Wherein, the mode discriminator D a Is composed of a PatchGAN model, for learning the blood pressure classes in the pattern adaptation set (i.e., three classes of hypertension, normal blood pressure, and hypotension). The PatchGAN model is composed of a fully connected convolutional neural network, and in blood pressure class prediction, the region-based receptive field is scaled to the region length, i.e., the entire signal is convolved by a discriminator, and the final blood pressure class is predicted using the average response of 1 × N samples, unlike the original PatchGAN.
Constructing a blood pressure estimation module and a corresponding target optimization function, wherein after the pre-training tasks of the self-supervision learning module and the mode adaptation module are completed, the BP value is accurately estimated through transfer learning, and a pre-trained encoder G is used a Feature extractor G migrating to an estimated BP value d Above, using the MSE loss as the regression training loss L d ：
Specifically, the blood pressure estimation learning sub-module is composed of a Transformer encoder, as illustrated by the blood pressure estimation learning sub-module in fig. 3. The transform encoder is the same as the transform encoder described above, and the weights thereof are obtained by pre-training the blood pressure mode adaptation learning submodule.
In step 3, training a noninvasive blood pressure estimation model according to the monitored photoplethysmographic signals and invasive continuous blood pressure waveform signals of the human body, specifically comprising the following steps:
the data after data preprocessing is obtained and divided according to the proportion of 7.5.
In step 4, collecting the photoplethysmogram signals of the user, inputting the photoplethysmogram signals into a noninvasive blood pressure estimation model for noninvasive blood pressure estimation, specifically:
collecting a PPG signal of a user, and sending a signal waveform as an input into a non-invasive blood pressure estimation model to obtain a blood pressure value corresponding to the user so as to realize non-invasive blood pressure estimation.
The present invention performs validation on a data set containing 1213 subjects from four different sources using an individual-based data partitioning paradigm. The experiment of the present invention avoids the problem of data leakage that may result from the random shuffling paradigm. In addition, the present invention sets up a test set of >85 subjects strictly in accordance with AAMI standards, as compared to existing small sample studies. Experimental test results show that the estimation errors of systolic pressure (SBP) and diastolic pressure (DBP) are respectively 0.85 +/-4.21 mmHg and 0.49 +/-2.76 mmHg, and the method meets the clinical standard (AAMI standard) of medical instruments.
The invention provides a noninvasive blood pressure estimation method based on self-supervision transfer learning, which comprises the steps of collecting photoplethysmography signals and invasive continuous blood pressure waveform signals of a human body, constructing a noninvasive blood pressure estimation model based on self-supervision transfer learning, training the noninvasive blood pressure estimation model according to the photoplethysmography signals and the invasive continuous blood pressure waveform signals of the monitored human body, collecting photoplethysmography signals of a user, inputting the photoplethysmography signals into the noninvasive blood pressure estimation model, and performing noninvasive blood pressure estimation; the input of the method only consists of photoplethysmography, so that the acquisition circuit only needs to acquire PPG signals, and compared with the traditional method based on the propagation velocity of the pulse wave, the method omits the step of acquiring electrocardiosignals, does not need excessive derivation calculation and characteristic engineering, is convenient to integrate into devices such as a bracelet and the like, does not need blood pressure measurement devices such as a cuff and the like, breaks away from the constraint of the cuff, and enables the devices to be more portable; the method can realize continuous blood pressure estimation and long-term blood pressure monitoring, can be used for measuring the blood pressure in daily life, and can not bring the influence of trauma and discomfort to human bodies during measurement; the method does not need calibration and is convenient to use; the method adopts an automatic supervision transfer learning framework, can carry out a representation learning model of fine-grained BP prediction from 11 converted PPG sequences, can promote the model to utilize more discriminative characteristics in the learning process, and has richer information contained in input signals, so that the measured blood pressure result is more stable, and higher prediction precision is achieved.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the foregoing, the description is not to be taken in a limiting sense.
Claims (8)
1. A non-invasive blood pressure estimation method based on self-supervision transfer learning is characterized by comprising the following steps:
step 1: collecting photoelectric volume pulse wave signals and invasive continuous blood pressure waveform signals of a human body;
step 2: constructing a non-invasive blood pressure estimation model based on self-supervision transfer learning;
and step 3: training a noninvasive blood pressure estimation model according to the photoplethysmography signals and invasive continuous blood pressure waveform signals of the monitored human body;
and 4, step 4: collecting the photoplethysmogram signals of a user, inputting the photoplethysmogram signals into a noninvasive blood pressure estimation model, and performing noninvasive blood pressure estimation.
2. The non-invasive blood pressure estimation method based on the self-supervised transfer learning as claimed in claim 1, wherein in step 1, the photoplethysmographic pulse wave signal and the invasive continuous blood pressure waveform signal of the human body are collected, specifically:
collecting a photoplethysmography signal, namely a PPG signal, and an invasive continuous blood pressure waveform signal, namely an invasive BP signal, of a human body, wherein the invasive BP signal is used as a reference signal.
3. The noninvasive blood pressure estimation method based on the self-supervised transfer learning of claim 2, wherein in the step 2, a noninvasive blood pressure estimation model based on the self-supervised transfer learning is constructed, specifically:
the method comprises the steps of constructing a non-invasive blood pressure estimation model based on self-supervision transfer learning, wherein the non-invasive blood pressure estimation model comprises a data preprocessing module, a self-supervision learning module, a mode adaptation module and a blood pressure estimation module.
4. The non-invasive blood pressure estimation method based on the self-supervised transfer learning of claim 3, wherein a data preprocessing module is constructed, specifically:
the method comprises the steps of acquiring collected PPG signals and invasive BP signals, resampling the PPG signals and the invasive BP signals, removing noise after sampling is finished, removing outliers of the PPG signals after the noise is removed, normalizing the PPG signals after the PPG signals are removed, performing phase alignment processing on the PPG signals and the matched invasive BP signals after the processing is finished, ensuring that the phases of the PPG signals and the invasive BP signals are consistent, segmenting the PPG signals and the invasive BP signals after the phase alignment is finished, generating a BP mode adaptive data set and a BP estimation data set according to the segmented PPG signals and the invasive BP signals, and taking the PPG signals without the matched invasive BP signals as an auto-supervision data set.
5. The noninvasive blood pressure estimation method based on self-supervised transfer learning of claim 4, wherein a self-supervised learning module is constructed, specifically:
constructing a self-supervision learning module and a corresponding target optimization function, acquiring a self-supervision data set, carrying out signal transformation on the self-supervision data set, inputting a transformed signal into the self-supervision learning module for signal reconstruction or signal transformation recognition, and comparing the reconstructed or transformed signal with a signal before transformation to realize pre-training of the self-supervision learning module;
wherein a PPG signal set in an auto-supervised data set is acquired
Where s represents the number of samples, the PPG signal set is signal transformed by a signal transformation function G (·), i.e.:
finally obtaining the transformed signal
The specific signal transformation comprises Gaussian noise addition, power frequency noise addition, motion interference addition, baseline drift addition, respiratory tract sinus arrhythmia frequency modulation, random mask masking, hard clipping, amplitude inversion, time scale interchange and time scale scaling, wherein each signal transformation mode transforms PPG signals in each input self-supervision data set according to preset probability;
for PPG signal reconstruction, the signal recovery function is learned by the Transformer based encoder-decoder architecture for recovering the transformed signal to the original PPG signal, as:
wherein the content of the first and second substances,
for the reconstructed PPG signal sample set, MSE is defined as the training loss
For PPG signal transform identification, the pseudo labels of signal variations are also classified by the Transformer encoder architecture:
in the formula, gamma i Is the loss factor of the ith conversion task, the corresponding loss
Comprises the following steps:
in the formula, p i e.P is the label automatically generated from the ith converted PPG, κ i Is the predicted probability of the model for that class.
6. The noninvasive blood pressure estimation method based on the self-supervised transfer learning of claim 5, wherein a mode adaptation module and a blood pressure estimation module are constructed, and specifically:
constructing a mode adaptation module and a corresponding objective optimization function, wherein the mode adaptation module comprises a mode feature extractor G a And a mode discriminator D a In the pattern feature extractor G a And a mode discriminator D a A gradient inversion layer is added in between, the gradient inversion layer performs constant transformation in the forward propagation process, and meanwhile, D is added a To G a Does not adversely propagate gradient ^ L during back propagation a Mode feature extractor G a Mode discrimination for the remaining Transformer encoder after removing the Transformer decoder from the supervised learning moduleDevice D a Is composed of PatchGAN, wherein, a pattern discriminator D a A pattern feature extractor G for distinguishing a feature representation between a target abnormal blood pressure pattern and a normal blood pressure pattern with a minimized pattern loss a And a mode discriminator D a Performing antagonistic learning through the game theory principle of the min-max optimization task:
wherein the loss L is resisted a Comprises the following steps:
in the formula, y to p target (Y) represents a target label, loss of pattern adaptive training L a (j) Comprises the following steps:
in the formula, y j Mode label x being the jth PPG signal j ∈R t×d ；
And constructing a blood pressure estimation module and a corresponding target optimization function, wherein the blood pressure estimation module consists of a Transformer encoder after two times of progressive pre-training. Specifically, after the pre-training task of the self-supervision learning module and the mode adaptation module is completed, the pre-trained encoder G is processed a Feature extractor G migrating to estimated BP values d In the above, accurate estimation of BP value is realized by transfer learning, and MSE loss is used as regression training loss L d ：
7. The method for noninvasive blood pressure estimation based on self-supervised transfer learning of claim 6, wherein in step 3, the noninvasive blood pressure estimation model is trained according to the photoplethysmography signals and the invasive continuous blood pressure waveform signals of the monitored human body, which are specifically as follows:
the data after data preprocessing is acquired and divided according to the proportion of 7.5.
8. The method of claim 7, wherein in step 4, the photoplethysmographic signals of the user are collected and input into a noninvasive blood pressure estimation model for noninvasive blood pressure estimation, specifically:
collecting a PPG signal of a user, and sending a signal waveform as an input into a non-invasive blood pressure estimation model to obtain a blood pressure value corresponding to the user so as to realize non-invasive blood pressure estimation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211609290.0A CN115836846A (en) | 2022-12-14 | 2022-12-14 | Non-invasive blood pressure estimation method based on self-supervision transfer learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211609290.0A CN115836846A (en) | 2022-12-14 | 2022-12-14 | Non-invasive blood pressure estimation method based on self-supervision transfer learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115836846A true CN115836846A (en) | 2023-03-24 |
Family
ID=85578651
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211609290.0A Pending CN115836846A (en) | 2022-12-14 | 2022-12-14 | Non-invasive blood pressure estimation method based on self-supervision transfer learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115836846A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116211316A (en) * | 2023-04-14 | 2023-06-06 | 中国医学科学院阜外医院 | Type identification method, system and auxiliary system for multi-lead electrocardiosignal |
| CN116304777A (en) * | 2023-04-12 | 2023-06-23 | 中国科学院大学 | Self-adaptive electrocardiosignal denoising method and system based on reference signal during rest |
-
2022
- 2022-12-14 CN CN202211609290.0A patent/CN115836846A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116304777A (en) * | 2023-04-12 | 2023-06-23 | 中国科学院大学 | Self-adaptive electrocardiosignal denoising method and system based on reference signal during rest |
| CN116304777B (en) * | 2023-04-12 | 2023-11-03 | 中国科学院大学 | Self-adaptive electrocardiosignal denoising method and system based on reference signal during rest |
| CN116211316A (en) * | 2023-04-14 | 2023-06-06 | 中国医学科学院阜外医院 | Type identification method, system and auxiliary system for multi-lead electrocardiosignal |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Rong et al. | A multi-type features fusion neural network for blood pressure prediction based on photoplethysmography | |
| Zhu et al. | Learning your heart actions from pulse: ECG waveform reconstruction from PPG | |
| CN115836846A (en) | Non-invasive blood pressure estimation method based on self-supervision transfer learning | |
| Chaudhuri et al. | Ambulation analysis in wearable ECG | |
| Zhao et al. | An IoT-based wearable system using accelerometers and machine learning for fetal movement monitoring | |
| Roy et al. | On-device reliability assessment and prediction of missing photoplethysmographic data using deep neural networks | |
| Sonawani et al. | Biomedical signal processing for health monitoring applications: a review | |
| Rajput et al. | Automated detection of hypertension using wavelet transform and nonlinear techniques with ballistocardiogram signals | |
| Liu et al. | Motion artifact detection in ppg signals based on gramian angular field and 2-d-cnn | |
| Qin et al. | Advances in cuffless continuous blood pressure monitoring technology based on PPG signals | |
| Priyadarshini et al. | Review of PPG signal using machine learning algorithms for blood pressure and glucose estimation | |
| Pan et al. | Improved blood pressure estimation using photoplethysmography based on ensemble method | |
| Zhang et al. | PPG-based Heart Rate Estimation with Efficient Sensor Sampling and Learning Models | |
| Naeini et al. | A Deep Learning–based PPG Quality Assessment Approach for Heart Rate and Heart Rate Variability | |
| Ma et al. | STP: Self-supervised transfer learning based on transformer for noninvasive blood pressure estimation using photoplethysmography | |
| Suleman et al. | Respiratory Events Estimation From PPG Signals Using a Simple Peak Detection Algorithm | |
| US20210117782A1 (en) | Interpretable neural networks for cuffless blood pressure estimation | |
| Zhou et al. | OSA-CCNN: obstructive sleep apnea detection based on a composite deep convolution neural network model using single-lead ECG signal | |
| Wadhwani et al. | IOT based biomedical wireless sensor networks and machine learning algorithms for detection of diseased conditions | |
| Bassiouni et al. | Combination of ECG and PPG Signals for Healthcare Applications: A Survey | |
| Sayem et al. | A novel 1D generative adversarial network-based framework for atrial fibrillation detection using restored wrist photoplethysmography signals | |
| Surapaneni et al. | Graph Signal Processing Based Classification of Noisy and Clean PPG Signals Using Machine Learning Classifiers for Intelligent Health Monitor | |
| De Giovanni | System-Level Design of Adaptive Wearable Sensors for Health and Wellness Monitoring | |
| Lu et al. | Multiple Feature Extraction Based on PPG Signal to Realize Blood Pressure Measurement of Composite Multi-channel Convolutional Neural Network Model | |
| Ghosal et al. | Improved method for motion artifact reduction from finger photoplethysmogram signal |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |