CN111528814A - Method for monitoring blood pressure through machine learning based on LSTM neural network - Google Patents

Abstract

The invention discloses a method for monitoring blood pressure by machine learning based on an LSTM neural network, which comprises the steps of obtaining electrocardio pulse wave signals and pressure pulse wave signal data of a monitored object by a signal acquisition device; preprocessing the acquired signals to remove noise and baseline drift, and filtering to obtain a cardiac electric wave and a PPW waveform; segmenting and processing the electrocardio waveform and PPW signal waveform signal data; extracting the characteristics of the segmented data into a characteristic vector sequence, and inputting the characteristic vector sequence into an LSTM neural network model; the output information of the LSTM neural network is the blood pressure information of the monitored object, including the systolic pressure SBP and the diastolic pressure DBP. The invention can effectively utilize PTT which is a key parameter, and is difficult to identify aiming at a plurality of characteristics of pulse waves of partial crowds, and improves the universality of blood pressure monitoring by only identifying necessary characteristics.

Description

Method for monitoring blood pressure through machine learning based on LSTM neural network

Technical Field

The invention belongs to the field of medical treatment, and particularly relates to a continuous blood pressure monitoring method, in particular to a method for extracting multi-feature mixture by utilizing signals of electrocardio and pressure pulse waves and continuously monitoring systolic pressure and diastolic pressure by utilizing an LSTM neural network.

Background

Blood pressure is the pressure of blood against the side of the vessel wall, which is generated by blood flowing through the vessel during systole, and is a comprehensive reflection of hemodynamic factors such as the amount of circulating blood, cardiac output, and elasticity of the arterial vessel wall. The systolic pressure and the diastolic pressure are important physiological indexes which can reflect the functional conditions of the heart and the blood vessel of the human body, and are one of four basic characteristics of human health detection.

At present, the conventional blood pressure measuring methods are an auscultation method and an oscillography method, but both the methods cannot carry out real-time continuous monitoring, and the equipment has large volume and is inconvenient. The blood pressure monitoring method capable of continuous measurement is mainly based on pulse wave signals for detection. There are three main methods for measuring blood pressure based on pulse wave signals, namely a method for measuring Pulse Wave Velocity (PWV), a method for measuring pulse wave transit time (PTT), and a method for measuring pulse wave characteristic parameters (PWP).

Because the shape of the photoplethysmography (PPG) waveform varies greatly among different people, the PPG waveform is difficult to extract for subjects with cardiovascular problems, and therefore, only extracting the PPG waveform features has a large range limitation, and the accuracy is also low. The blood pressure is conveniently detected and fitted by using the PTT, but the problems that (1) the pulse wave of a human body is easily interfered by a plurality of factors including respiration, impedance, measurement posture and the like, and the accuracy of blood pressure calculation by using only one parameter of the PTT is low are also existed. (2) PTT has a high correlation with systolic pressure and a low correlation with diastolic pressure.

The Pressure Pulse Wave (PPW) detected by the piezoelectric sensor can reflect the change of the blood vessel pressure more accurately than the photoplethysmographic pulse wave because of its high sensitivity and directly reflects the change of the blood vessel pressure, and is more accurate for measuring the blood pressure. However, in the conventional measurement method based on the pressure pulse wave, too many features are extracted from the waveform, and the features are easily interfered, so that the measurement accuracy is influenced, and the application range is narrow.

A Recurrent Neural Network (RNN) is a Neural Network for processing sequence data, each layer of which outputs not only a next layer but also a hidden state. A long and short Term Memory Network (LSTM) is an improved recurrent neural Network, and solves the problems of gradient loss and gradient explosion in the RNN long sequence training process. The method has better performance in a long-time sequence, and reduces the influence caused by characteristic change along with time.

Disclosure of Invention

Aiming at the importance of blood pressure to health in the background technology and the problems of the current blood pressure monitoring method, the invention aims to provide a method for monitoring blood pressure by machine learning based on an LSTM neural network, which utilizes multi-feature mixed signals of electrocardio and pressure pulse waves.

Therefore, the technical scheme adopted by the invention is as follows: the method for monitoring the blood pressure based on the machine learning of the LSTM neural network is characterized by comprising the following steps: the method comprises the following steps:

1) acquiring electrocardio pulse wave signals and pressure pulse wave signal data of a monitored object through a signal acquisition device;

2) preprocessing the signals obtained in the step 1) for removing noise and baseline drift, and filtering to obtain a cardiac electric wave and a PPW waveform;

3) segmenting electrocardiographic waveform and PPW signal waveform signal data, identifying R wave peaks of an electrocardiogram, and dividing the ECG into a cardiac cycle according to adjacent R-R wave peaks; identifying the wave crest of the PPW wave, dividing the PPW into a cardiac cycle according to the adjacent wave crest, and then carrying out normalization processing on the cardiac cycle;

4) and performing feature extraction on the segmented data, wherein the feature extraction comprises the following features:

a) PTT, PTT is the pulse wave transmission time, the wave crest of the PPW pulse wave and the R wave of the heart wave are identified at the same time, and the pulse wave transmission time can be obtained by calculating the time difference between the wave crest of the R wave and the wave crest of the PPW wave;

b) the distance from the starting point to the rising peak is the rising time t 1;

c) the distance between the starting point and the dicrotic wave is the second rise time t 2;

d) reflection Index (RI), ratio of contraction peak to rebuffering peak, i.e. h₁/h₂；

e) Heart rate HR, namely identifying R waves of the electrocardiographic waveform, calculating the time difference between the electrocardiographic waveform and the adjacent R waves, and then calculating a heart rate value according to the time of the electrocardiographic waveform;

5) recording the characteristic sequence obtained in the step 4) as a characteristic vector sequence

x₁Representing PTT, x₂Represents a rise time t1, x₃Representing a second rise time t2, x₄Denotes the index of Reflection (RI), x₅A data set representing the heart rate HR is shown,

is x₁Transposing the matrix; inputting the characteristic vector sequence into an LSTM neural network model, wherein the neural network model consists of a CNN convolution layer, a batch standardization BN layer, a bidirectional LSTM layer, a dropout layer and a full connection Dense layer;

6) the output information of the LSTM neural network is the blood pressure information of the monitored object, including the systolic pressure SBP and the diastolic pressure DBP.

In the first part, the raw signal data includes the electrocardiograph pulse wave signal and the pressure pulse wave data.

The two electrodes can transmit electric signals on the surface of a human body, the signals are filtered and amplified through the AFE electrocardio analog front end, and are converted into digital signals through the high-resolution ADC, so that the single-lead electrocardiogram can be obtained.

The pressure pulse wave signal is obtained by using a high-sensitivity pressure sensor, placing the sensor and an artery, and adopting a piezoresistive sensor or a piezoelectric sensor, wherein the circuits of the two types of sensors are slightly different, the piezoresistive sensor has resistivity along with the pressure change, the resistance change needs to be converted into the voltage change, and the change of the voltage signal is detected so as to obtain the pressure pulse wave signal. The piezoelectric pressure sensor changes electric charge, so that the change of the electric potential can be directly detected through an amplifier and a filter to obtain a pressure pulse wave signal.

In the second section above, the pre-processing of the signal includes removing noise and baseline wander. In the signal acquisition process, due to factors such as human respiration and movement, human impedance and the like, the obtained signal contains various noises, so that interferences such as noise, baseline drift and the like are removed through preprocessing. Since the effective frequency of the pulse wave signal is generally between 0.3-30Hz, the raw data is filtered by using Butterworth 5-order bandpass filtering, and the cut-off frequencies are 0.3Hz and 30Hz respectively. The baseline drift in the signal is mainly caused by low-frequency interference such as respiration of the measured object and sensor movement. The pulse wave and the electrocardiosignal contain abundant low-frequency components, and the baseline drift can cover useful information. Morphological filtering is widely applied to the field of signal processing and image analysis, and morphological characteristics of signals can be greatly reserved. In order to filter baseline drift, the invention uses a larger structure to respectively carry out the operation of opening first and closing second and the operation of closing first and opening second on the original signal, and then the average value is taken for filtering.

Further, the feature vector sequence X is input to the LSTM neural network model, and the calculation steps of the neural network model are as follows:

i) firstly inputting the characteristic vector sequence X into the convolution layer to extract a characteristic layer;

ii) inputting a batch of standardized BN layers to carry out standardized processing on the characteristics of each layer;

iii) inputting the bidirectional LSTM network by taking M as a time step;

iv) adding a dropout layer after the step to prevent the network from being over-fitted;

v) finally outputting Y ═ SBP, DBP, SBP is systolic pressure and DBP is diastolic pressure by the fully connected Dense layer.

Long time of LSTMThe memory neural network is commonly used for semantic analysis, machine translation and the like, and the LSTM can be applied to blood pressure detection to reduce the accuracy reduction caused by the change of characteristics along with time. LSTM uses the current input x^TAnd h passed from the previous state^t-1Training results in 4 states, three state gates and one memory cell. Wherein x is^TRepresenting the current layer, input data set, h^t-1Representing the output of the previous cell. The three control gates are respectively an input gate, a forgetting gate and an output gate. The input of the previous node is selectively forgotten, the playback point is selectively memorized, and a proper memory is selected for output.

Further, the loss function of the neural network is Mean Square Error (MSE), and the formula is as follows:

the Adam optimization is adopted in the gradient optimization, the adaptive learning rate of each parameter can be automatically and adaptively calculated, and the method not only stores the exponential decay average value of the previous square gradient, but also maintains the exponential decay average value of the previous gradient. In regularization, the L2 norm is reasonably set to prevent overfitting. MSE is used to detect the deviation between the predicted and true values of the model. M denotes the total number of samples tested, M denotes the specific sample index, y_mThe true value of the mth sample is represented,

representing the predicted value of the m-th sample.

Output y of the neural network^TI.e. the measured systolic pressure SBP, diastolic pressure DBP are included. y is^TThe transpose of the blood pressure result dataset is output for the neural network.

Compared with the prior art, the invention has the beneficial effects that: the invention relates to a blood pressure calculation method based on electrocardio and pressure pulse wave signals, which is used for calculating to obtain a key parameter PTT related to blood pressure, mixing necessary characteristics extracted from photoelectric pulse waves and pressure pulse waves, and reducing the influence of the time change of the characteristics on a measurement result through an LSTM neural network. The method can effectively utilize PTT which is a key parameter, is difficult to identify aiming at a plurality of characteristics of pulse waves of partial crowds, and improves the universality of blood pressure monitoring by only identifying necessary characteristics. The invention can carry out continuous blood pressure detection, and has convenient detection and high comfort.

Drawings

The invention is further described below with reference to the accompanying drawings.

FIG. 1 is a flow chart of a pulse wave multi-feature-based acquisition method according to the present invention;

FIG. 2 is a diagram of a pressure pulse waveform and its characteristic partitions; the distance from the starting point of t1 to the rising peak is the rising time; t2 is the distance between the starting point and the dicrotic wave, i.e. the second rise time; h1 is the systolic peak, h2 is the dicrotic peak;

FIG. 3 is a diagram of the structure of the LSTM;

FIG. 4 is a diagram of a signal acquisition unit;

FIG. 5 is a diagram of a morphological filter structure;

fig. 6 is a diagram of a blood pressure prediction neural network model.

Detailed Description

The present invention will be described in further detail with reference to the following detailed description and accompanying drawings.

As shown in figure 1, the present invention first acquires the photoplethysmography, the electrocardiographic wave, the cardiac wave and the pressure pulse wave signals.

The acquisition of the signals mainly comprises the following units: a photoelectric sensor unit, an electrocardio unit and a pressure pulse wave unit.

The photoelectric signal can be obtained by a photoelectric sensor, the principle is that a light-emitting diode is used for irradiating blood vessels of a human body, light is green light with the wavelength of 530nm, the penetration of the green light is stronger than that of red light and infrared light, reflection interference of human tissue light is smaller, and the obtained photoelectric volume wave waveform is better. The photoelectric sensor may be placed at the wrist, finger, or the like. When the green light is reflected to the photodiode, the internal resistance of the photodiode is changed, and the voltage at two ends of the photodiode is changed.

During the heart beat, the electric potential on the blood vessel is polarized and depolarized along with the contraction and relaxation of the ventricle, so that a slight point position change is caused in the human body, and an electrocardiogram can be obtained by detecting the change of the electric potential. Because the PTT signal is obtained, only the position of the electrocardio R wave needs to be positioned, and the single-lead electrocardio is adopted for the convenience of equipment and test.

The two electrodes are attached to two parts of a human body, and are recommended to be respectively placed on fingers and wrists for collection. Through which potential variations can be transmitted to a post-processing unit.

The pressure pulse wave signals are acquired through a pressure sensor, and the pressure sensor used for detecting the pulse waves commonly used is a piezoresistive sensor and a piezoelectric sensor. When the pressure changes, the resistance of the pressure sensor changes in direct proportion to the measured pressure, and then a corresponding voltage output signal can be obtained through circuit design. The piezoelectric pressure sensor utilizes the positive piezoelectric effect, when the sensor is stressed by pressure, the sensor generates electric charges, the magnitude of the electric charge is related to the magnitude of the pressure, and the change of the arterial pressure can be measured by detecting the magnitude of the electric charge to obtain a pressure pulse wave. The pressure sensor may be placed at the radial artery or carotid artery of the wrist.

The obtained pulse wave signals are weak and contain much noise. Filtering and amplification by the filtering and amplifying unit are required. As the effective signal of the human body is between 0.3Hz and 30Hz, the cut-off frequency of the band-pass filtering is 0.3Hz and 30Hz respectively. Including 50Hz power frequency interference. Meanwhile, the method for carrying out differential amplification by using the amplifier is used for carrying out the method on the acquired signal.

And obtaining the analog signal after the processing of the filtering and amplifying unit in the last step. The signal is then fed to an AD conversion unit.

And converting the analog signal into a digital signal through the AD conversion unit for processing by a processor.

As shown in fig. 1, based on the pulse wave signals obtained in the previous step, further filtering processes are performed, which mainly include band-pass filtering and morphological filtering. The band-pass filtering is software filtering, the software band-pass filtering adopts Butterworth filtering, the order is 5 orders, and the cut-off frequency is 0.3Hz to 30 Hz.

Meanwhile, the baseline drift in the signal is mainly caused by low-frequency interference such as respiration of a measured object and sensor movement. The pulse wave and the electrocardiosignal contain abundant low-frequency components, and the baseline drift can cover useful information. Morphological filtering is widely applied to the field of signal processing and image analysis, and morphological characteristics of signals can be greatly reserved. In this embodiment, to filter the baseline drift, a larger structure is used to perform the first-to-close operation and the first-to-close and then-to-open operation on the original signal, and then an average value is taken to perform filtering. The structure is shown in figure 5.

Before being input into the neural network, the electrocardiogram signal waveform and the PPW signal waveform signal data are segmented, and because the signal frequency of each person is different, if the same sampling number is simply used, the accuracy of the model is influenced. The R wave peak of the electrocardiogram is first identified and the ECG is divided into one cardiac cycle based on the adjacent R-R wave peaks. Similarly, the peak of the PPW wave is identified, the PPW is divided into one cardiac cycle according to the adjacent peaks, and then normalization processing is performed on the cardiac cycle.

And (3) carrying out feature extraction on the segmented data, and mainly extracting the following features:

1) PTT, PTT is the pulse wave transmission time, the wave crest of PPW pulse wave and the R wave of the heart wave are identified at the same time, and the pulse wave transmission time can be obtained by calculating the time difference between the wave crest of the R wave and the wave crest of the PPW wave.

2) The distance from the starting point to the rising peak, i.e., the rise time, t 1;

3) the distance between the starting point and the dicrotic wave, i.e. the second rise time, t 2;

4) reflection Index (RI), shrinkage peak (h)₁) With the heavy Bofeng (h)₂) Ratio of (i.e. h)₁/h₂；

5) Heart rate HR, the R-wave of the electrocardiographic waveform is identified, the time difference between the electrocardiographic waveform and the adjacent R-wave is calculated, and then the heart rate value is derived from the time.

There are 5 features in total, as described above. Respectively extract the above-mentioned characteristics from the continuous cardiac cycles to form

A data set. x is the number of₁Representing PTT, x₂Represents a rise time t1, x₃Representing a second rise time t2, x₄Denotes the index of Reflection (RI), x₅A data set representing the heart rate HR is shown,

are respectively x₁，x₂，x₃，x₄，x₅Transposing of the matrix.

The above-mentioned sequence of features is respectively a sequence of feature vectors

And inputting the characteristic vector sequence into the neural network model by taking the characteristic vector sequence as input. The calculation steps of the neural network model are as follows:

a) firstly inputting the feature matrix X into the convolution layer to extract a feature layer;

b) inputting a batch standardized BN layer to carry out standardized processing on the characteristics of each layer;

c) inputting the M as a time step into a bidirectional LSTM network;

d) then adding a dropout layer to prevent the network from being over-fitted;

e) finally, the full-connection Dense layer outputs Y ═ SBP, DBP, SBP is systolic pressure, DBP is diastolic pressure, and Y refers to the output scalar of the neural network, namely the predicted blood pressure result data set output.

The LSTM long-time memory neural network is commonly used for semantic analysis, machine translation and the like, and can be used for detecting blood pressure, so that the accuracy reduction caused by the change of characteristics along with time is reduced. LSTM uses the current input x^TAnd h passed from the previous state^t-1Training results in 4 states, three state gates and one memory cell. Wherein x is^TRepresenting the current layer, input data set, h^t-1Representing a preceding sheetAnd (4) outputting the element. The three control gates are respectively an input gate, a forgetting gate and an output gate. The input of the previous node is selectively forgotten, the playback point is selectively memorized, and a proper memory is selected for output.

The loss function of the neural network is Mean Square Error (MSE), and the formula is as follows:

the Adam optimization is adopted in the gradient optimization, the adaptive learning rate of each parameter can be automatically and adaptively calculated, and the method not only stores the exponential decay average value of the previous square gradient, but also maintains the exponential decay average value of the previous gradient. In regularization, the L2 norm is reasonably set to prevent overfitting. MSE is used to detect the deviation between the predicted and true values of the model. M denotes the total number of samples tested, M denotes the specific sample index, y_mThe true value of the mth sample is represented,

representing the predicted value of the m-th sample.

Output y of the neural network^TI.e. the measured systolic pressure SBP, diastolic pressure DBP are included. y is^TThe transpose of the blood pressure result dataset is output for the neural network.

Claims (2)

1. The method for monitoring the blood pressure based on the machine learning of the LSTM neural network is characterized by comprising the following steps: the method comprises the following steps:

1) acquiring electrocardio pulse wave signals and pressure pulse wave signal data of a monitored object through a signal acquisition device;

2) preprocessing the signals obtained in the step 1) for removing noise and baseline drift, and filtering to obtain a cardiac electric wave and a PPW waveform;

3) segmenting electrocardiographic waveform and PPW signal waveform signal data, identifying R wave peaks of an electrocardiogram, and dividing the ECG into a cardiac cycle according to adjacent R-R wave peaks; identifying the wave crest of the PPW wave, dividing the PPW into a cardiac cycle according to the adjacent wave crest, and then carrying out normalization processing on the cardiac cycle;

4) and performing feature extraction on the segmented data, wherein the feature extraction comprises the following features:

a) PTT, PTT is the pulse wave transmission time, the wave crest of the PPW pulse wave and the R wave of the heart wave are identified at the same time, and the pulse wave transmission time can be obtained by calculating the time difference between the wave crest of the R wave and the wave crest of the PPW wave;

b) the distance from the starting point to the rising peak is the rising time t 1;

c) the distance between the starting point and the dicrotic wave is the second rise time t 2;

d) reflection Index (RI), ratio of contraction peak to rebuffering peak, i.e. h₁/h₂；

e) Heart rate HR, namely identifying R waves of the electrocardiographic waveform, calculating the time difference between the electrocardiographic waveform and the adjacent R waves, and then calculating a heart rate value according to the time of the electrocardiographic waveform;

5) recording the characteristic sequence obtained in the step 4) as a characteristic vector sequence

x₁Representing PTT, x₂Represents a rise time t1, x₃Representing a second rise time t2, x₄Denotes the index of Reflection (RI), x₅A data set representing the heart rate HR is shown,

is x₁Transposing the matrix; inputting the characteristic vector sequence into an LSTM neural network model, wherein the neural network model consists of a CNN convolution layer, a batch standardization BN layer, a bidirectional LSTM layer, a dropout layer and a full connection Dense layer;

6) the output information of the LSTM neural network is the blood pressure information of the monitored object, including the systolic pressure SBP and the diastolic pressure DBP.

2. The LSTM neural network-based machine learning method of monitoring blood pressure of claim 1, wherein: the feature vector sequence X is used as input to an LSTM neural network model, and the calculation steps of the neural network model are as follows:

i) firstly inputting the characteristic vector sequence X into the convolution layer to extract a characteristic layer;

ii) inputting a batch of standardized BN layers to carry out standardized processing on the characteristics of each layer;

iii) inputting the bidirectional LSTM network by taking M as a time step;

iv) adding a dropout layer after the step to prevent the network from being over-fitted;

v) finally outputting Y ═ SBP, DBP, SBP is systolic pressure and DBP is diastolic pressure by the fully connected Dense layer.

Images