CN108498089B

CN108498089B - Noninvasive continuous blood pressure measuring method based on deep neural network

Info

Publication number: CN108498089B
Application number: CN201810430346.3A
Authority: CN
Inventors: 袁学光; 张阳安; 杨帆
Original assignee: Beijing University of Posts and Telecommunications
Current assignee: Beijing University of Posts and Telecommunications
Priority date: 2018-05-08
Filing date: 2018-05-08
Publication date: 2022-03-25
Anticipated expiration: 2038-05-08
Also published as: CN108498089A

Abstract

The invention provides a noninvasive continuous blood pressure measuring method based on deep learning, which mainly comprises a pulse wave signal preprocessing method, a deep neural network model establishing method and a blood pressure measuring method. The pulse wave signals input into the neural network are preprocessed and subjected to filtering, denoising and normalization processing, so that the established model is more accurate and has better adaptability. The method applies the deep neural network, avoids the problems of subjective feature extraction and complex mathematical modeling, and simultaneously leads the network to be capable of measuring the blood pressure of the pulse waves of different forms of different individuals through a large amount of data training networks, and can be applied to a wide blood pressure prediction scene after one-time training. Compared with the prior art, the method has the advantages of high objectivity and strong model robustness, is suitable for being integrated in household medical monitoring equipment, and is also suitable for monitoring the blood pressure of wearable equipment.

Description

Noninvasive continuous blood pressure measuring method based on deep neural network

Technical Field

The invention relates to the field of digital signal processing, in particular to a method for performing noninvasive continuous blood pressure calculation based on digital signal processing.

Background

Blood Pressure (BP) is an important parameter reflecting the performance of the circulatory system of the human body, and examination of blood pressure is a large way to clinically judge diseases and observe treatment effects. Humans suffer from a number of diseases related to Arterial Blood Pressure (ABP), including hypertension and heart disease. Therefore, monitoring of blood pressure is of great significance.

Currently, the blood pressure measurement methods can be broadly divided into two major types, namely direct measurement and indirect measurement: direct measurement requires the catheter connected to a pressure sensor to be inserted directly into the aorta or heart, which is the most accurate measurement, but is only suitable for critically ill patients due to the high technical requirements and invasiveness; indirect measurement methods can be further classified into intermittent measurement methods and continuous measurement methods:

(I) the intermittent measurement method is typically a sphygmomanometer based on the korotkoff method, which is configured to determine diastolic and systolic pressures by blocking arterial blood flow through an air cuff and then monitoring the sound of arterial blood vessels through an earpiece. The disadvantage is that the measured blood pressure is a blood pressure value at a specific moment, which is not necessarily representative, and the compression and discomfort caused by the air sleeve can cause psychological stress to some patients.

(II) continuous measurement has a great advantage in observing continuous changes in blood pressure. The currently mature measuring methods include an arterial tonometry, a volume compensation method, a pulse wave velocity measurement method, a photoplethysmography and the like:

1) the arterial tonometry method is to make the pressure in the blood vessel equal to the external force by compressing the blood vessel and to obtain the arterial blood pressure by measuring the external pressure. The artery tension measuring method has higher precision and can basically meet the requirement of long-time noninvasive continuous blood pressure measurement. The disadvantage is that the pressure sensor must be small enough and accurately positioned just above the collapsed portion of the artery and the sensor position needs to be held fixed for a long period of time.

2) The volume compensation method is characterized in that the artery is in a load-shedding state by presetting reference pressure, and the external pressure is adjusted by adopting a fast-response servo control system according to the fluctuation time of the blood pressure to enable the artery blood vessel to be in a constant volume state all the time, and at the moment, the dynamic arterial blood pressure value can be obtained by measuring the external pressure. The method can continuously measure the blood pressure of each stroke and can measure the blood pressure waveform without distortion. The disadvantage is that the accuracy is insufficient and long-term measurement leads to venous congestion.

3) Pulse wave velocity measurement method for indirectly calculating arterial blood pressure by measuring Pulse Wave Velocity (PWV)^[1]. The advantage is that no air bag is needed, so that the problem caused by long-time measurement can not be caused. The disadvantage is that different mathematical models need to be established and the blood flow must be measured accurately.

4) The photoplethysmography (PPG) method, referred to as pulse wave method for short, measures the change of photoelectric signals to obtain pulse wave signals, and derives corresponding blood pressure values through the pulse wave signals. The method is easy to implement, high in measurement accuracy and free of any discomfort, and therefore the method becomes a hot point of research. Currently, there are three main methods of pulse wave therapy: a combination method of a cardiac electric wave (SSG) and a pulse wave (PPG), a combination method of two pulse waves and a blood pressure measuring method of pulse wave characteristic parameters.

5) Method for combining electrocardiographic waveform with photoplethysmography^[2][3][4]: the combined electrocardio-plethysmography and photoplethysmography estimates blood pressure by using a time interval between the transmission of the same arterial pulse wave from an electrocardio R wave to a pulse wave characteristic point. The formula of the blood pressure estimation by inference calculation is as follows:

where a is the thickness of the arterial wall, d is the internal diameter, g is the acceleration of gravity, E₀Is the elastic system of the arterial wall. K represents the distance traveled by the pulse wave, and T represents the travel time. The method cannot detect the change condition of the pulse wave in the pre-ejection period, so that the estimated blood pressure is not accurate enough. In addition, the method also needs to design a set of electrocardio sensor, which can increase the burden for the tested person and lead the measuring system to become complicated and the portability to be influenced.

6) Method for combining two-path photoelectric volume pulse waves^[5]: the method utilizes two different parts of the human body, such as fingers and wrists, and estimates the blood pressure according to the time interval of the characteristic points of the two pulse wave signals. The method of combining the two pulse waves keeps the consistency of arterial blood vessels and avoids the influence of the heart pre-shooting period. However, the pulse wave propagation time obtained by the method only has a certain relation with the variation of the systolic pressure, the variation of the diastolic pressure has little correlation, and the diastolic pressure cannot be accurately measured. And two paths of pulse waves need to be measured, which brings great inconvenience to the testee.

7) Pulse wave feature parameter method^{[6][7][8][9][10]}: the method is characterized in that the relationship between the pulse wave characteristic parameters and the blood pressure is established. This method only requiresOne path of pulse wave is measured, and diastolic pressure and systolic pressure are estimated through the correlation between characteristic parameters of the waveform and blood pressure. The disadvantage is that the stability requirement for the pulse wave signal is high, but the blood pressure measurement is still highly practical considering that the blood pressure measurement is generally in a quiet condition. Another disadvantage is that the characteristic parameter method is highly subjective and empirical and the mathematical process is complex.

8) Neural network and characteristic parameter combination method^{[11][12][13][14]}: the method still belongs to a characteristic parameter method, but the extracted characteristic parameters are not used for solving a derivation formula through a mathematical modeling method, but are learned through a neural network. Thus, although the mathematical reasoning reduces the workload, the feature selection still avoids subjectivity.

9) End-to-end method based on neural network^[15]: the method directly inputs data into a network and outputs the data as a predicted blood pressure value. Literature reference^[15]The wavelet neural network is adopted, the nonlinear wavelet basis is used as the nonlinear excitation function of the neuron in the established neural network, but the method only uses the neural network with 3 layers, and deeper pulse wave characteristics cannot be obtained for blood pressure measurement.

The common application of the pulse wave characteristic parameter method is that a parameter equation of characteristic parameters corresponding to blood pressure is established by extracting the characteristics of a single PPG signal and applying a traditional mathematical modeling method, so that the blood pressure can be measured. In general, researchers select characteristics deemed appropriate through experience and a large number of experiments, and a series of signal processing methods and mathematical modeling methods are applied to construct a satisfied mathematical formula. However, the method is very dependent on manual feature selection, and the quality of feature selection is directly related to prediction accuracy. In addition, the selected characteristics may have large differences for pulse waves of different forms acquired by different measuring devices. This shows that the method is too subjective and not robust enough.

The invention adopts the deep neural network, the processing model structure is more complex, thus the characteristics of higher dimension of pulse wave can be learned, meanwhile, the network can learn various characteristics by self, thereby avoiding artificial characteristic selection, and the invention can be applied to a wide blood pressure prediction scene after one training. In addition, the invention carries out preprocessing on the pulse wave signals input into the neural network, and carries out filtering, denoising and normalization processing, thereby enabling the established model to be more accurate and better in adaptability.

Reference documents:

[1]Miyauchi Y，Koyama S，Ishizawa H.Basic experiment of bloodpressure measurement which uses FBG sensors.In：Proceedings of the IEEE International Conference on Instrumentation and Measurement Technology，Minneapolis，2013.1767-1770。

[2]Newlin D B.Relationships of pulse transmission times to pre-ejection period and blood pressure.Phychophysiology，1981，18：316-321。

[3]Lane J D，Greenstadt L，Shapiro D.Pulse transit time and blood pressure：an intensive analysis.Phychophysiology，1983，20：45-49。

[4] guo Li Hua. study of sleeveless continuous blood pressure measurement method based on PPG signals [ D ]. Hangzhou: zhejiang university, 2011.

[5]Chen Y，Wen C，Tao G，et a1.Continuous and noninvasive measurement of systolic and diastolic blood pressure by one mathematical model with the same model parameters and two separate pulse wave velocities.Ann Biomed Eng，2012，40：871-882。

[6]Li YJ，Wang ZL，Zhang L，et al.Characters available in photoplethysmogram for blood pressure estimation：beyond the pulse transit time.Australas Phys Eng Sci，2014，37∶367-376。

[7]Teng XF，Zhang Y T.Continuous and noninvasive estimation of arterial blood pressure using a photoplethysmographic approach.In：Proceedings of the IEEE International Conference on Medicine and Biology Society，Cancun，2003.3153-3156。

[8]Yoon Y，Yoon G.Nonconstrained blood pressure measurement by photoplethysmography.J Opt Soc Korea，2006，10：91-95。

[9]Fortino G，Giamp｀a V.PPG-based methods for non invasive and continuous BP measurement：an overview and development issues in body sensor networks.In：Proceedings of the IEEE International Conference on Medical Measur，Ottawa，2010.10-13。

[10] Sun Jian pulse wave feature extraction algorithm and application study thereof [ D ]. university of major graduates, 2007.

[11] Popyang. study of predicting essential hypertension using an artificial neural network model [ D ]. university of chinese medical, 2010.

[12] Neural network blood pressure measurement algorithm [ J ] based on Gaussian fitting, sensor and microsystem, 2014, 33 (4): 132, 134, 138. DOI: 10.3969/j.issn.1000-9787.2014.04.039.

[13]Kurylyak Y，Lamonaca F，Grimaldi D.A neural network-based method for continuous blood pressure estimation from a PPG signal.In：Proceedings of the IEEE International Conference on Instrumentation and Measurement Technology，Minneapolis，2013.280-283。

[14] Wuyudong, shuncong, shenglauchun, etc. non-invasive continuous blood pressure measurement methods based on SWT and ANN study [ J ] chinese medical devices, 2017, 32 (5): 22-27. D0I: 10.3969/j.issn.1674-1633.2017.05.006.

[15]Peng Li，Ming Liu，Xu Zhang，Xiaohui Hu，Bo Pang，Zhaolin Yao，Hongda Chen.Novel wavelet neural network algorithm for continuous and noninvasive dynamic estimation of blood pressure from photoplethysmography[J].Science China Information Sciences，2016，59(4)。

Disclosure of Invention

The invention provides a noninvasive continuous blood pressure measuring method based on deep learning, aiming at the problem of a traditional pulse wave characteristic parameter method.

The signal preprocessing method comprises the following steps of firstly, cutting an acquired time domain pulse wave signal into a series of equal-length periodic segments, and processing the time domain pulse wave signal into a plurality of equal-length periodic segments, wherein the processing method comprises the following steps:

1.1) outputting a small segment of waveform, and estimating the period length L of the waveform through observation_predictPoint;

1.2) starting from the first point of the input waveform signal, take L_predictPoint: finding the maximum value of the data

And maximum position

Then from the first point to

Minimum value calculation

And position of minimum value

Then from

To L_predictPoint minimum

And position of minimum value

The first signal period is

To

Period of timeLength of

Then storing

To

Is the first segment data, and records

1.3) from

And then repeating the following process to obtain segment data of each period until the end: wherein k is 1, 2.. n, n is a natural number;

a) will be dotted

As the starting point of the segment, its position is noted

Value is recorded as

b) Get

To

To find the maximum value

And maximum position

c) To find

To

Minimum value of point

And position of minimum value

d) Preservation of

To

The data of (1) is the (k + 1) th cycle fragment data with the length of

And record

The following steps are then performed for all periodic segments: 1.4) noise filtering, and 1.5) normalization processing.

Further, the specific method for filtering noise in step 1.4) is as follows:

1.4) calculating a first quartile and a third quartile of the cycle length for all the processed cycle segments, and then directly discarding all the segments of which the cycle length is less than the first quartile or greater than the third quartile as noise; for all remaining period segments, the average maximum is calculated

Mean value of starting point

Mean value of end point

Where m is the number of remaining period segments, then, setting a threshold epsilon, for each period segment, if:

d_max＞(d_{ave_max}+εd_{ave_max}) Or d_max＜(d_{ave_max}-εd_{ave_max}) Or

d_{min_start}＞(d_{ave_start}+εd_{ave_start}) Or d_{min_start}＜(d_{ave_start}-εd_{ave_start}) Or

d_{mmin_end}＞(d_{ave_end}+εd_{ave_end}) Or d_{min_end}＜(d_{ave_end}-εd_{ave_end})，

The periodic segment is regarded as noise and is directly discarded;

then, for all remaining period segments, considered as noise-free or small-noise data, a denoising process may be performed by a filter. Preferably, the filtering is performed using a band-pass filter.

Further, the normalization processing method in step 1.5) comprises the following steps: finding the maximum value L of all current cycle lengths_{cycle_max}For each period segment, if the period length L_cycle＜L_cycleMax, then interpolate it using cubic spline interpolation to make its length equal to L_cycle__max；

Further, the pulse wave sampling frequency is 125 Hz.

Further, before or after the normalization process, a translation process is performed, in which all the waves are translated to a range of 0 to 1 in order to make the finally obtained pulse wave range consistent, since the pulse wave amplitude range acquired by different devices is not necessarily the same. The specific method comprises the following steps:

A) some range intervals are established first, such as: [ (-2, 0), (-1, 0), (-0.5, 0.5), (0, 1), (1, 2) ];

B) judging the interval of each pulse wave;

C) the minimum value of the interval is then added or subtracted from all the sampling points of the pulse wave, so that the pulse wave is translated to the range of 0-1 without deformation of the waveform.

The invention also provides a noninvasive continuous blood pressure measuring device which comprises an input device, a deep neural network model calculating device and an output device, wherein the deep neural network calculating model comprises a neural network model with an input layer of pulse wave signals, a convolutional layer, a batch normalization layer, a pooling layer, a full-connection layer and a blood pressure measuring result layer. Between the input device and the deep neural network model calculation device, a data preprocessing device is further included, which performs the pulse wave signal preprocessing method according to any one of claims 1 to 8.

The invention also provides a training method of the deep neural network model, which comprises the following steps:

1) dividing pulse wave signal data into a training set, a verification set and a test set, and inputting the pulse wave signals of the training set into an input layer as one-dimensional data;

2) performing convolution operation through one or more convolution layers to extract local characteristics of signals;

3) normalizing the features by one or more batch normalization layers;

4) increasing the network receptive field through one or more pooling layers, merging small local features into larger local features;

5) repeating the steps 2) -4) for 3-10 times;

6) inputting the characteristic data obtained in the step 5) into a full-connection layer, and carrying out nonlinear mapping on the characteristics extracted from the previous convolutional layer by the full-connection layer to map the characteristics into a value to be predicted;

7) after passing through the full connection layer of 3-5 layers, the predicted blood pressure value y is output_predict；

8) By predicted blood pressure value y_predictAnd the actual blood pressure value y_trueComparing to obtain the absolute error

e＝|y_predict-y_true|；

9) The model loss function is the mean absolute error, the mean absolute error L of the training data_trainThe calculation method is

Wherein N is₁Is the total number of training data used during each training, i is the sequence number of the training data, e_iIs the absolute error of the data with sequence number i, y_{predict_i}Is the predicted value of data with sequence number i, y_{true_i}Is the actual blood pressure value of the data with sequence number i, A (-) is the activation function, ω_iIs a training parameter for data with sequence number i, x_iInput data with sequence number i;

10) inputting the verification set data, calculating the average absolute error L of the verification set data_valThe calculation method is

Wherein N is₂Is the total number of all verification data, j is the serial number of the verification data, e_jIs the absolute error of the data with sequence number j;

11) calculating a loss function L of a training set_trainGradient for weight matrix omega

Wherein

For the partial derivative operator, the weight matrix ω is the training parameter ω for each datum_iIn the form of a matrix;

12) update the value of the weight matrix:

wherein: as an update operator, α is an update step,

is the gradient calculated in step 11), and omega is a random value initially;

13) repeat steps 1) -12) until the validation set error L_valTo the extent that the range is satisfied: at least less than 5mmHg, or the number of training sessions reaches a specified value: generally the number of times does not exceed 10 ten thousand.

Wherein the pulse wave signal input in the step 1) is the pulse wave signal processed by the signal preprocessing method.

Further, the original data is divided into a training set, a verification set and a test set, and the proportion of the training set is as follows: and (4) verification set: the test set is 8: 1, and when the data volume is large, the proportion of the training set is increased. Wherein the input in the step 1) is the pulse wave signal of the training set.

Further, the method also comprises the following steps:

14) inputting test set data for testing, and testing the obtained Mean Absolute Error (MAE) L_testAnd Root Mean Square Error (RMSE) M_testIs the true error of the model, if L of the model_testLess than 5mmHg and M_testThe standard is reached when the mmHg is less than 8 mmHg.

If the error does not reach the standard, but L in the step 13)_valWhen the standard is reached, the representative model is overfit, and some methods for reducing overfit are needed, such as: reducing convolutional and fully-connected layer depths reduces model complexity and thus overfitting, or adding a random inactivating layer in the fully-connected layer allows some neurons to be randomly deleted and thus overfitting, and then retraining 1) -13).

If the error does not reach the standard, and L in the step 13)_valAnd (3) the model does not reach the standard, namely, the representative model is under-fitted, and some methods for reducing the under-fitting are adopted, such as: increaseContinuing with step 13) after a large number of trainings, or increasing the convolutional layer and fully-connected layer depths so that the model complexity increases to reduce under-fitting, and then retraining 1) -13).

Further, the pulse wave signal input into the input layer is a signal processed by a pulse wave signal preprocessing method.

A noninvasive continuous blood pressure measuring method based on a deep neural network comprises the steps of 1.1) -1.5) processing data, then 1) -13) training the deep neural network, and 14) judging whether a trained model reaches the standard, wherein blood pressure measurement can be carried out on new data after the model reaches the standard.

The method provided by the invention can effectively solve the problem of noninvasive continuous blood pressure measurement, does not cause physical trauma to a user in the measurement process, collects the whole pulse wave, and can always measure the blood pressure value corresponding to each second from the 4 th second; meanwhile, the method can estimate the blood pressure of the pulse waves of different forms of different individuals, does not perform classification calculation on the waveform data of different forms any more, thereby reducing the data requirement, avoiding the problems of subjective feature extraction and complex mathematical modeling, improving the blood pressure measurement speed based on the pulse waves, and realizing the blood pressure measurement within the required precision for most individuals after one-time training of the model. Compared with the prior art, the method has the advantages of high objectivity and strong model robustness, is suitable for being integrated in household medical monitoring equipment, and is also suitable for monitoring the blood pressure of wearable equipment.

Drawings

Fig. 1 is a block diagram of a noninvasive blood pressure measuring method based on a deep neural network.

Fig. 2 is a pulse wave signal and its power spectrum.

Fig. 3 is a flow chart of the building of the deep neural network model.

FIG. 4 is a schematic diagram of a single-layer fully-connected neural network.

Fig. 5 is a schematic diagram of a neural network model structure.

Detailed Description

The invention is described in detail below with reference to the figures and the embodiments.

The technical scheme of the invention is as shown in figure 1, and mainly comprises the steps of preprocessing the acquired original pulse wave signals, and then inputting the preprocessed original pulse wave signals into a deep neural network model to obtain a blood pressure value.

The sampling rate of the pulse wave is usually 125Hz, and the collected time-domain pulse wave signal needs to be cut into a series of periodic segments in the collection of the original pulse wave signal and the pretreatment part, then the filtering treatment is carried out, and finally the pulse wave signal is processed into the equal-length periodic segments through an interpolation method. The cycle fragment data first needs to be processed:

1.1) outputting a small segment of waveform, and estimating the period length L of the waveform through observation_predictPoint, the estimated value should be larger than the observed cycle length, i.e. the estimated value should be larger;

And maximum position

Then from the first point to

Minimum value calculation

And position of minimum value

Then from

To L_predictPoint minimum

And position of minimum value

The first signal period is

To

Length of cycle

Then storing

To

Is the first segment data, and records

1.3) from

a) will be dotted

As the starting point of the segment, its position is noted

Value is recorded as

b) Get

To

To find the maximum value

And maximum position

c) To find

To

Minimum value of point

And position of minimum value

d) Preservation of

To

The data of (1) is the (k + 1) th cycle fragment data with the length of

And record

Then, the step 1.4) filtering processing needs to be carried out on the processed periodic segment data, and the data with larger noise can be directly discarded, wherein the method comprises the following steps:

1.4) calculating a first quartile and a third quartile of the cycle length for all the processed cycle segments, and then directly discarding all the segments of which the cycle length is less than the first quartile or greater than the third quartile as noise; the Quartile (Quartile) refers to a numerical value which is obtained by arranging all numerical values from small to large in statistics, dividing the numerical values into four equal parts and locating at the positions of three dividing points; sorting all period segments from small to large according to length, and discarding the data of the first 25 percent and the data of the second 25 percent;

for all remaining period segments, the average maximum is calculated

Mean value of starting point

Mean value of end point

d_max＞(d_{ave_max}+εd_{ave_max}) Or d_max＜(d_{ave_max}-εd_{ave_max}) Or

d_{mmin_start}＞(d_{ave_start}+εd_{ave_start}) Or d_{mmin_start}＜(d_{ave_start}-εd_{ave_start}) Or

d_{mmin_end}＞(d_{ave_end}+εd_{ave_end}) Or d_{mmin_end}＜(d_{ave_end}--εd_{ave_end})，

The periodic segment is regarded as noise and is directly discarded;

the rest data with no noise or less noise can be denoised by a filter. Here, a simple band-pass filter can be used for filtering, and generally, the frequency of the normal pulse signal of a human is 0.5-4Hz, as shown in fig. 2. Noise signals outside this range can be filtered out. Then, the normalization processing of the step 1.5) is uniformly carried out on all the pulse wave signals subjected to the denoising processing, so that the subsequent prediction is facilitated.

In addition, the range of the pulse wave amplitude value acquired by different devices is not necessarily required, such as the range of 0-1, the range of-0.5-0.5 and the like. To make the range of pulse waves consistent, we choose to shift all waves to the range of 0-1 by:

1) some range intervals are established first: [ (-2, 0), (-1, 0), (-0.5, 0.5), (0, 1), (1, 2) ];

2) judging the interval of each pulse wave;

3) the minimum value of the interval is then added or subtracted from all the sampling points of the pulse wave, so that the pulse wave is translated to the range of 0-1 without deformation of the waveform.

Finally, in order to make the lengths of the period segments equal, interpolation processing needs to be performed on the normalized data, and the method includes:

1.5) determining the maximum value L of the remaining period length_{cycle_max}For each period segment, if the period length L_cycle＜L_{cycle_max}Then it is interpolated using cubic spline interpolation to make its length equal to L_{cycle_max}。

As shown in fig. 3, the deep neural network model is built by training, specifically, raw pulse wave data is obtained by a PPG method, a data set is built by noise filtering, a true blood pressure value corresponding to each data is obtained by a sphygmomanometer, and the built data set is introduced into the neural network model to obtain a predicted blood pressure value. In order to train the neural network model, a training set (training set), a relatively small test set (test set), and a verification set (validation set) with a large amount of data are prepared. The training set data is used for training the neural network, the verification set is used for evaluating the performance of the network model after certain training, the test set is used for testing the performance of the network when the network training is determined to be completed, and the data obtained by the test set can be regarded as the actual performance of the network model. In general, the training set, the verification set and the test set can be divided according to the ratio of 8: 1, and when the data volume is large, the proportion of the training set can be further increased.

The training of the neural network model is the key of the whole system, and the quality of the network model determines whether the measured error can reach a satisfactory range. As shown in fig. 4, it is mainly composed of an input layer, a hidden layer, and an output layer, and the neural network structure shown in fig. 4 is the most basic fully-connected network structure, which is represented by regarding each input as an input neuron, and then connecting each neuron with all neurons in the layer above and the layer below it. A deep neural network is a kind of neural network, which mainly shows that the number of hidden layers is large, and generally varies from ten layers to hundred layers. Meanwhile, the most common network structure in the deep neural network is a convolutional neural network, which is widely applied to image processing, voice processing and other problems, and is represented by performing a translation convolution operation on the whole input signal through a convolution kernel.

The model processing and training process is as follows:

1) dividing pulse wave signal data into a training set, a verification set and a test set, and inputting the pulse wave signals of the training set into an input layer as one-dimensional data; the input pulse wave signals are regarded as one-dimensional data;

2) performing convolution operation through a plurality of convolution layers to extract local characteristics of signals;

3) normalizing the features through a plurality of batches of normalization layers;

4) the network receptive field is increased through a plurality of pooling layers, and small local features are converged into larger local features;

5) repeating the steps 2) -4) for 3-10 times;

6) inputting the obtained feature data into a full-connection layer, and carrying out nonlinear mapping on the features extracted by the convolutional network by the full-connection network to obtain a value to be predicted;

7) after passing through the full-connection network of several layers, the predicted blood pressure value y is output_predict；

8) By predicted blood pressure value y_predictThe actual blood pressure value y corresponding to each segment of training data_trueComparing to obtain absolute error e ═ y_predict-y_true|；

The model loss function selects the Mean Absolute Error (MAE) by the calculation method

Where N is the total number of training data, i is the sequence number of the training data, e_iIs the absolute error of the data with sequence number i, y_predictiIs the predicted value of data with sequence number i, y_{true_i}Is the actual blood pressure value of the data with sequence number i, A (-) is an activation function, and the common activation functions are sigmoid function, tanh function, relu function, etc., x_iIs input data with sequence number i, omega_iTraining parameters of data with sequence number i, representing the weight of each input data, the model has a plurality of layers, wherein each layer of convolution layer and full-connection layer has its own weight, the training of the whole model actually finds out the correct omega through iteration, the matrix formed by the training parameters is the weight matrix, and omega is initially_iIs a random value;

9) after each training, inputting the data of the verification set, and calculating the average absolute error L of the data of the verification set_valThe calculation method is

Where N is the total number of verification data, i is the serial number of the verification data, e_iIs the absolute error of the data with sequence number i;

10) calculating a loss function L of a training set_trainGradient for weight matrix omega

Wherein

For the partial derivative operator, the weight matrix ω is the training parameter ω for each datum_iA matrix is formed;

11) update the value of the weight matrix:

wherein: the calculation result of the left side ω is updated to the calculation result of the right side, and α is the update step size, which is a parameter that needs to be set artificially, and may be 0.0001-0.1,

is the gradient calculated from step 11);

12) repeat steps 1) -12) until the validation set error L_valTo the extent that the range is satisfied: at least less than 5mmHg, or the number of training sessions reaches a specified value: the number of times is not more than 10 ten thousand;

13) inputting test set data for testing, and testing the obtained Mean Absolute Error (MAE) L_testAnd Root Mean Square Error (RMSE) M_testIs the true error of the model, if L of the model_test＜5mmHg、M_testThe standard is reached when the mmHg is less than 8 mmHg. If the error does not reach the standard, but L in the step 13)_valWhen the standard is reached, the representative model is overfit, and some methods for reducing overfit are needed, such as: reducing the depth of the convolutional network and the fully-connected network so that the complexity of the model is reduced and the overfitting is reduced, or adding a random inactivation layer in the fully-connected layer so that some neurons are randomly deleted and the overfitting is reduced, and then retraining the steps 1) -13); if the error does not reach the standard, and L in the step 13)_valAnd (3) the model does not reach the standard, namely, the representative model is under-fitted, and some methods for reducing the under-fitting are adopted, such as: continuing to perform step 13) after increasing the number of trainings, or increasing the depths of the convolutional network and the fully-connected network so that the complexity of the model is increased to reduce under-fitting, and then retraining steps 1) -13).

When the model is trained through the training process, test set data is input for testing, the average absolute error and the mean square error obtained through testing can be regarded as the real error of the model, and the model after being trained to reach the standard can measure the blood pressure of new data.

The model established by the invention uses a deep neural network model combined by various structures, as shown in fig. 5, the model sequentially comprises an input layer of pulse wave signals, a convolutional layer, a batch normalization layer and a pooling layer, and finally a full connection layer and a blood pressure measurement result layer, the model is established by the deep neural network model training method, and by using the model, the corresponding blood pressure value can be measured only by inputting pulse wave data of several seconds, and continuous measurement can be carried out by continuous input, so that accurate, rapid and noninvasive continuous blood pressure measurement can be realized.

The invention also comprises a noninvasive continuous blood pressure measuring device which comprises an input device, a deep neural network model calculating device and an output device, wherein the deep neural network model calculating device comprises a neural network model with an input layer of pulse wave signals, a convolution layer, a batch normalization layer, a pooling layer, a full-connection layer and a blood pressure measuring result layer. Between the input device and the deep neural network model calculation device, a data preprocessing device is further included, and the data preprocessing device executes the pulse wave signal preprocessing method. The output device is a display device.

The results of the experimental establishment of the deep neural network model according to the present invention are shown in the following table:

through 4 tests, a deep neural network model is effectively established, and the feasibility and the effectiveness of a data preprocessing method and a modeling method are verified.

It will be understood by those skilled in the art that the foregoing is only exemplary of the present invention, and is not intended to limit the invention to the particular forms disclosed, since any modifications, equivalents, improvements and the like which fall within the spirit and scope of the invention are deemed to fall within the scope and spirit of the invention.

Claims

1. A noninvasive continuous blood pressure measuring device based on deep learning is characterized by comprising an input device, a deep neural network model calculating device and an output device, wherein the deep neural network model calculating device comprises a neural network model with an input layer, a convolutional layer, a batch normalization layer, a pooling layer, a full-link layer and a blood pressure measuring result layer of pulse wave signals; the deep neural network calculation model is obtained through model training, and the training method comprises the following steps:

1) dividing the pulse wave signals into a training set, a verification set and a test set, and inputting the pulse wave signals of the training set into an input layer as one-dimensional data;

5) repeating the steps 2) -4) for 3-10 times;

6) inputting the characteristic data obtained in the step 5) into a full connection layer;

7) after passing through a full connection layer of 3-5 layers, outputting a predicted blood pressure value ypredict;

8) comparing the predicted blood pressure value ypredict with the actual blood pressure value ytrue to obtain an absolute error e ═ ypredict-ytrue |;

9) the model loss function is the average absolute error, the average absolute error Ltrain of the training data, and the calculation method is

Wherein N1 is the total number of training data used in each training, i is the serial number of the training data, ei is the absolute error of the data with serial number i, ypredict _ i is the predicted value of the data with serial number i, ytrue _ i is the actual blood pressure value of the data with serial number i, a (-) is the activation function, ω i is the training parameter of the data with serial number i, xi is the input data with serial number i;

10) inputting verification set data, and calculating the average absolute error Lval of the verification set data by the method

Where N2 is the total number of all verification data, j is the serial number of the verification data, and ej is the absolute error of the data with serial number j;

11) calculating the gradient of the loss function Ltrain of the training set to the weight matrix omega

Wherein, the weight matrix omega is a matrix formed by training parameters omega i of each data;

12) update the value of the weight matrix:

wherein: as an update operator, α is an update step,

is the gradient calculated in step 11), and omega is a random value initially;

13) repeating steps 1) -12) until the validation set error Lval is reduced to meet the range or the training times reach the specified value.

2. The apparatus of claim 1, wherein the satisfaction ranges of step 13) are that the validation set error Lval is at least less than 5mmHg, and that the number of training sessions is specified to be no more than 10 ten thousand sessions.

3. The apparatus of claim 1, wherein the update step size α ranges from 0.0001 to 0.1.

4. The apparatus of any one of claims 1-3, wherein the training method further comprises the steps of:

14) inputting test set data for testing, and testing the obtained average absolute error (MAE) Ltest and Root Mean Square Error (RMSE) Mtest, wherein if the Ltest of the model is less than 5mmHg and the Mtest is less than 8mmHg, the error reaches the standard;

if the error does not reach the standard and Lval reaches the standard in the step 13), indicating that the representative model is over-fitted, and needing to adopt a method for reducing over-fitting and then carrying out the training of the steps 1) -13);

if the error does not reach the standard and Lval does not reach the standard in the step 13), namely, the model is represented to be under-fitted, the step 13) is trained continuously or the steps 1) to 13) are retrained again by adopting a method for reducing the under-fitting.

5. The apparatus of claim 1, wherein the raw data is divided into a training set, a validation set, and a test set, wherein the ratio of the training set, the validation set, and the test set is 8: 1, and when the amount of data is large, the ratio of the training set is increased.

6. The apparatus of claim 1, further comprising a data preprocessing device between the input device and the deep neural network device, the data preprocessing device being configured to preprocess the pulse wave signals, the preprocessing method comprising:

the method comprises the following steps of cutting collected time domain pulse wave signals into a plurality of equal-length periodic segments, wherein the processing method comprises the following steps:

1.1) outputting a small segment of waveform, and estimating the cycle length of the waveform to be Lpredict points through observation;

1.2) starting from the first point of the input waveform signal, taking Lpredict point: finding the maximum value of the data

And maximum position

Then from the first point to

Minimum value calculation

And the position of the minimum value is determined,

then from

Minimum value calculation to Lpredict point

And the position of the minimum value is determined,

the first signal period is

To

Length of cycle

Then storing

To

Is the first segment data, and records

1.3) from

And then repeating the following process to obtain segment data of each period until the end: wherein k is 1, 2 … n;

a) will be provided with

The next point of the point is taken as the starting point of the segment, and the position of the next point is recorded as

Value is recorded as

b) Get

To

To find the maximum value

And maximum position

c) To find

To

Minimum value of point

And position of minimum value

d) Preservation of

To

The data of (1) is the (k + 1) th cycle fragment data with the length of

And record

7. The apparatus of claim 6, wherein the 1.4) processing method for filtering noise is:

1.4) calculating a first quartile and a third quartile of the cycle length for all the processed cycle segments, and then directly discarding the cycle segments which are smaller than the first quartile or larger than the third quartile in all the cycle lengths as noise; for the remaining period segments, the average maximum is calculated

Mean value of starting point

Mean value of end point

dmax > (dave _ max + ε dave _ max) or dmax < (dave _ max- ε dave _ max) or

dmmin _ start > (dave _ start + ε dave _ start) or dmmin _ start < (dave _ start- ε dave _ start) or

dmin _ end > (dave _ end + ε dave _ end) or dmmin _ end < (dave _ end- ε dave _ end),

the periodic segment is regarded as noise and is directly discarded;

then, regarding all the remaining period segments as noiseless or small-noise data, denoising processing is performed through a filter.

8. The apparatus of claim 6, wherein the step of 1.5) normalizing process is:

1.5) the maximum value of all the current cycle lengths, Lcycle _ max, is found, and for each cycle segment, if the cycle length Lcycle < Lcycle _ max, the cycle segment is interpolated to a length equal to Lcycle _ max.

9. The apparatus of claim 8, wherein: the interpolation adopts a cubic spline interpolation method.

10. The apparatus of claim 6, wherein the Lpredict estimate in step 1.1) is greater than the observed period length, i.e., the estimate is greater.

11. The apparatus of claim 6, wherein the sampling frequency of the pulse wave is 125 Hz.

12. The apparatus of claim 7, wherein the filter is a band pass filter.

13. The apparatus according to any of claims 6 to 12, wherein before or after the 1.5) normalization processing step, a translation step is further performed:

A) some range intervals are established first: [ (-2, 0), (-1, 0), (-0.5, 0.5), (0, 1), (1, 2) ];

B) judging the interval of each pulse wave;