Continuous blood pressure measurement method based on remote photoplethysmography and radar sensor

Abstract

The invention discloses a continuous blood pressure measuring method based on a remote photoelectric volume method and a radar sensor. In order to solve the problem that a traditional blood pressure measurement method cannot meet the non-inductive and continuous measurement requirement, the system uses a camera to collect face pulse wave signals, uses a radar sensor to collect chest pulse wave signals, uses Fwave and Fphase as input feature sets, uses real blood pressure values (diastolic pressure DBP and systolic pressure SBP) as labels, trains a machine learning model, and obtains a blood pressure measurement result. And forming a mapping relationship between the Fwave and Fphase feature sets and the corresponding DBP and SBP, and calculating a time difference between the two. The system is optimized in links of signal acquisition, processing, calculation, estimation and the like, adopts an advanced algorithm to extract pulse wave features and uses a machine learning model to map blood pressure values, has the advantages of non-contact, continuous monitoring, convenience and the like, can provide reliable data support for cardiovascular health monitoring, and meets clinical and family medical care requirements.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical signal processing, and specifically relates to a continuous blood pressure measurement method based on the fusion of remote photoplethysmography and radar sensors. Technical Background

[0002] In some hypertensive patients, blood pressure may be within the relatively normal range during the day, but it will abnormally increase during nighttime sleep, that is, "dipper blood pressure" becomes "non-dipper blood pressure" or "reverse-dipper blood pressure". Continuous blood pressure measurement during sleep helps to detect such occult conditions in a timely manner. Patients with obstructive sleep apnea hypopnea syndrome (OSAHS) repeatedly experience apnea or shallow breathing during sleep, which can cause blood pressure fluctuations. Continuous blood pressure measurement can capture the association between these blood pressure changes and apnea events, and is of great significance for the diagnosis, treatment, and prognosis assessment of OSAHS.

[0003] Currently, common continuous non-invasive blood pressure measurement methods are mostly contact-based, such as the volume compensation method, which is easily affected by factors such as temperature and human circulating blood volume. Prolonged compression of blood vessels will cause great discomfort to the human body and does not meet international standards. The arterial tension method is mainly applicable to the measurement of superficial skin arteries, requires high precision of the sensor, and cannot measure for a long time. Intermittent blood pressure measurements such as the Korotkoff sound auscultation method and the oscillometric method, although convenient to use and simple to operate, are suitable for home measurement, but have obvious defects. Traditional methods use the human ear to judge through a stethoscope, which is prone to subjective errors and is not suitable for measuring anytime and anywhere.

[0004] To solve this problem, the present invention proposes a continuous blood pressure measurement method based on the fusion of remote photoplethysmography and radar sensors, which can perform continuous blood pressure monitoring at night and provide relatively accurate physiological signal extraction results. Summary of the Invention

[0005] The system of the present invention measures the pulse wave signals on the face and chest, calculates the time difference between the two, and this time difference is closely related to the pulse transit time (PTT) and the pulse wave velocity (PWV). This method uses the feature extraction ability of neural networks and applies machine learning to more accurately estimate SBP and DBP, ensuring the extraction accuracy of signals.

[0006] Specifically, the innovation points of the present invention are reflected in the following aspects:

[0007] 1. Integrate facial video and radar chest pulse wave measurement. For the face, a camera is used to capture the subtle skin color changes caused by blood flow with rPPG technology to obtain the pulse wave. For the chest, a radar sensor is used to detect the minute movements on the skin surface in the millimeter wave band to collect the pulse wave. This combination not only enables non-contact measurement, avoiding the discomfort and skin irritation of contact measurement, but also can comprehensively obtain pulse wave information from multiple dimensions, greatly improving the measurement accuracy.

[0008] 2. Use a neural network for signal feature extraction. Taking F wave and F phase as the input feature set, and the real blood pressure values (diastolic blood pressure DBP, systolic blood pressure SBP) as labels, train a machine learning model to form the mapping relationship between the F wave and F phase feature sets and the corresponding DBP and SBP.

[0009] 3. In signal synchronization and time difference calculation, use high-precision clock synchronization technology to calibrate the clocks of the camera and the radar sensor to ensure that the signal acquisition times are consistent, reducing the time difference measurement error. When calculating, combine the pulse wave feature points and signal phase information to improve the calculation accuracy. In the face of noise and signal quality problems, use advanced algorithms such as wavelet denoising and adaptive filtering to enhance and denoise the signal, further improving the accuracy of time difference calculation.

[0010] Through the method of the present invention, continuous and non-intrusive blood pressure monitoring can be carried out during the night sleep state, and to a certain extent, the existing monitoring means are technically improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 FIG. 1 is a flowchart of blood pressure signal processing provided by the present invention.

[0012] FIG. 2 shows two extracted pulse wave signals.

[0013] Figure 3 FIG. 3 shows the names and dimensions of the signal feature extraction.

[0014] Figure 4 FIG. 4 is a neural network structure diagram for mapping SBP and DBO.

[0015] Figure 5 FIG. 5 shows the statistical results of the estimation model based on the BHS protocol. DETAILED DESCRIPTION OF THE INVENTION

[0016] Based on the arterial wall mechanical model and the arterial wave propagation model, its propagation speed is related to the elasticity of the blood vessel. The compliance (the change in blood vessel cross-sectional area divided by the blood vessel pressure) can be used to represent the blood vessel elasticity. Under certain assumptions, the pulse wave velocity (PWV) is related to the compliance (C) and arterial inertia (L), that is The pulse transit time (PTT) is inversely proportional to the pulse wave velocity, and the formula is (where | is the blood vessel length and P is the blood pressure). Therefore, by measuring the time difference between the pulse waves on the face and chest, the blood pressure value can be effectively estimated.

[0017] 1. Signal acquisition

[0018] The camera uses remote photoplethysmography (rPPG) to measure the pulse wave signal S in the facial area (forehead or cheek). face . Under normal light conditions, a conventional RGB camera (frame rate greater than or equal to 30 frames per second, 60 frames per second is recommended) is used, and a near-infrared camera with infrared fill light is used during the night sleep stage; automatic face recognition is performed on the video frames to obtain the region of interest (ROI); the time stamps of the picture frames during video shooting are recorded, and the video frames with an accurate time axis are restored through linear interpolation; a band-pass filter with a frequency range of 0.5 - 4 Hz is used to filter the extracted pulse wave signal; when the RGB camera is adopted, any dimensionality reduction method including but not limited to Green, Chrom, POS, ICA, APON_Vector, APON_Angle, etc. is used to reduce the three-channel RGB signals into a single-channel pulse wave signal, and the signal is shown in Figure 2(a); when the infrared camera is adopted, signals in multiple ROI regions are collected to improve the quality of the pulse wave signal.

[0019] A linear frequency modulation continuous wave (LFMCW) radar is used to detect the pulse wave in the chest S chest . Its transmitting synthesizer generates a linear frequency modulation signal, the echo signal is captured by the quadrature receiver, mixed with the transmitted signal in quadrature, the high-frequency part is filtered out by a low-pass filter to obtain an intermediate frequency signal, and then the intermediate frequency signal is sampled by the ADC. In order to obtain an accurate carotid artery pulse wave waveform (the carotid artery pulse wave can be regarded as the aortic pressure wave and can better reflect the blood pressure situation), a series of processing is performed on the collected signal, and a method based on non-convex optimization is used to remove the DC bias, and the accurate pulse wave signal is selected through operations such as body movement judgment and filtering, as shown in Figure 2(b).

[0020] 2. Feature extraction

[0021] Based on the processing results of the signal, feature extraction is performed on the extracted pulse wave F wave , and the extracted features are mainly divided into two categories, one is physiological parameters and the other is information parameters. Physiological parameters are calculated from feature points with physical meanings (such as systolic time), while information parameters are the overall information representation of the signal characteristics. In this scheme, two physiological parameters are extracted from medical features, and twenty-six information parameters are extracted from the time domain, entropy, wavelet domain, and derivative domain for continuous blood pressure estimation. As Figure 3 shown

[0022] Time-domain features:

[0023]

[0024] where signal is the carotid artery pulse wave signal, MN is the signal mean value, and SD is the signal standard deviation

[0025] Entropy features:

[0026] If it is defined and The approximate entropy is expressed as:

[0027] ApEn = Φ m (r) - Φ m+1 (r)

[0028] The sample entropy is expressed as:

[0029] SampEn(m, r, N) = -log(B m+1 (r) / B m (r))

[0030] Calculate the time-domain difference of the corresponding key points of S face and S chest to obtain the pulse wave phase difference feature F phase , using a method that combines feature points and signal phases; realizing time synchronization with the camera through high-precision clock synchronization technology to ensure the time consistency of the collected signals and reduce the time difference measurement error. Facing noise and signal quality problems, wavelet denoising and adaptive filtering algorithms are used to enhance and denoise the signals, further improving the accuracy of time difference calculation.

[0031] 3. Neural network model training

[0032] The neural network training flowchart is as shown in Figure 4 . In the present invention, a residual neural network is used to map SBP and DBP. ResNet is a residual neural network containing 17 convolutional layers and 1 fully connected layer. This model includes a preprocessing block, eight residual blocks, and 1 fully connected layer. There is 1 convolutional layer, 1 batch normalization layer, 1 activation layer (ReLU), and 1 pooling layer in the preprocessing block. The input signal of the layer is one-dimensional, the number of channels generated by convolution is 64, the size of the convolution kernel is 7, the convolution stride (stride) is 2, and the number of layers for padding 0 to each input side is 3. The batch normalization layer is for small batches of training data, introducing two learnable model parameters (scale parameter and offset parameter), mapping the standard normal distribution to an isomorphic space, enhancing the network's expression ability, and improving the model accuracy. In this section, the scale parameter is set to 64, and the activation function uses the ReLU (Rectified Linear Unit) function, and its formula is:

[0033]

[0034] The last layer of the network is a fully connected layer with 2048 input feature numbers and 1000 output feature numbers. Subsequently, after passing through the ReLU activation function, the output layer outputs 2 feature numbers, namely the systolic blood pressure (SBP) value and the diastolic blood pressure (DBP) value.

[0035] After the training of the residual neural network model is completed, different segments are tested, and the schematic diagram of the results is as Figure 5 shown. According to the BHS standard formulated by the British Hypertension Society, the cumulative error percentage of the errors in the estimation of SBP and DBP by this model can be seen. The proposed model is Class A in both the estimation of SBP and DBP based on the BHS standard, showing good accuracy.

[0036] Through the above steps, the present invention can measure blood pressure values based on videos and radar signals. This method has advantages in signal processing, feature extraction, etc., and is suitable for physiological state monitoring in various complex scenarios. Compared with traditional methods, the method of the present invention has a certain degree of improvement in signal-to-noise ratio, anti-interference ability, and real-time performance, and has great application prospects compared with existing methods.

Claims

1. A continuous blood pressure measurement method based on remote photoplethysmography and radar sensor, the main features of which include: S1. Measure the pulse signal at the facial position by using the remote photoplethysmography method to obtain the pulse signal S of the target human face face ; S2. Measure the mechanical wave of the target human chest movement using a radar sensor and process it to obtain the pulse signal S of the chest position chest ; S3. Process S face and S chest respectively for signal processing, obtain key points such as peaks and valleys of their respective time-domain waveforms, and calculate the feature set F wave ; S4. Calculate S face and S chest to obtain the time-domain difference of the corresponding key points to get the pulse wave phase difference feature F phase ; S5. With F wave and F phase as the input feature set, and the true blood pressure values (diastolic blood pressure DBP, systolic blood pressure SBP) as the labels, train a machine learning model to form the mapping relationship between the F wave and F phase feature sets and the corresponding DBP and SBP. S6. For the actual measurement target, use a video and a radar sensor to measure the facial and chest positions respectively, extract the corresponding feature sets according to the steps of S1 - S4, and input them into a pre-trained model to complete non-contact continuous blood pressure measurement.

2. In the continuous blood pressure measurement method according to claim 1, for the acquisition of the pulse wave signal based on remote photoplethysmography in step S1, it specifically includes using a conventional RGB camera (frame rate greater than or equal to 30 frames per second, preferably 60 frames per second) under normal light conditions, and using a near-infrared camera with infrared fill light during the night sleep stage; performing automatic face recognition on the video frames to obtain the region of interest (ROI); recording the time stamps of the picture frames during video shooting, and restoring the video frames with an accurate time axis through linear interpolation; filtering the extracted pulse wave signal with a band-pass filter of 0.5 - 4 Hz; when the RGB camera is used, using any dimensionality reduction method including but not limited to Green, Chrom, POS, ICA, APON_Vector, APON_Angle, etc., to reduce the RGB three-channel signals into a single-channel pulse wave signal; when the infrared camera is used, collecting multi-ROI region signals to improve the quality of the pulse wave signal.

3. In the continuous blood pressure measurement method according to claim 1, for the acquisition of the pulse signal based on the radar sensor in step S2, use a linear frequency modulation continuous wave (LFMCW) radar sensor, whose transmitting synthesizer generates a linear frequency modulation signal, and the echo signal is captured by an orthogonal receiver and mixed, filtered, and sampled with the transmitted signal; it has signal processing functions such as removing DC bias, body movement judgment, and filtering on the collected signals to obtain an accurate chest pulse wave signal.

4. The signal processing of the pulse wave in step S3 of the continuous blood pressure measurement method according to claim 1 includes filtering, smoothing, outlier screening, peak detection, valley detection, etc. For the feature set extraction in step S3, it specifically includes extracting features from multiple dimensions such as the time domain, entropy, wavelet domain, derivative domain, and medical analysis of the pulse waveform. The extracted features are mainly divided into two categories, one is physiological parameters, and the other is information parameters. Physiological parameters are calculated from feature points with physical meanings (such as systolic time), while information parameters are the overall information representation of signal characteristics. In this solution, two physiological parameters are extracted from medical features, and twenty-six information parameters are extracted from the time domain, entropy, wavelet domain, and derivative domain for continuous blood pressure estimation.

5. F in step S4 of the continuous blood pressure measurement method according to claim 1 phase The calculation specifically includes using a method that combines feature points and signal phase; achieving time synchronization with the camera through high-precision clock synchronization technology to ensure the time consistency of the acquired signals and reduce the measurement error of the time difference. In the face of noise and signal quality problems, wavelet denoising and adaptive filtering algorithms are used to enhance and denoise the signals, further improving the accuracy of the time difference calculation.

6. For the model training in step S5 of the continuous blood pressure measurement method according to claim 1, the machine learning model includes, but is not limited to, classical regression methods such as support vector regression (SVR), etc. The deep learning model adopts a neural network architecture, including constructing a residual neural network with 17 convolutional layers and 1 full connection layer. The model includes a preprocessing block, eight residual blocks and a full connection layer. There is a network with a convolutional layer, a batch normalization layer, an activation layer (ReLU) and a pooling layer in the preprocessing block. The network uses the Adam algorithm for parameter optimization to improve the learning ability of the model and the accuracy of blood pressure estimation.

7. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for extracting blood pressure signals from pulse wave signals as described in claim 1.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it causes the processor to execute the steps of the method as described in claim 1.

Images