Heart rate variation parameter prevention and blood pressure monitoring method based on terahertz radar electrocardiogram

Abstract

The invention relates to a method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variation parameters, and belongs to the field of terahertz radar signal processing. The method comprises the steps that a human thoracic cavity micro-motion signal is collected through a terahertz radar, an intermediate frequency signal is obtained through ADC sampling, distance dimension Fourier transform and phase unwrapping processing are carried out, and a target object is obtained; extracting a chest displacement signal containing heartbeat information; band-pass filtering is adopted to preliminarily separate respiration and heartbeat signals, and variation mode decomposition optimized by an ant colony algorithm is introduced to effectively suppress respiration harmonic interference; inputting the heartbeat signal into a multi-modal deep learning framework fusing a convolutional neural network, a long-short-term memory network and a multi-head attention mechanism to realize end-to-end high-precision reconstruction from a radar mechanical vibration signal to a standard electrocardio waveform; r peak detection is conducted on the reconstructed radar ECG signals, and time domain and frequency domain heart rate variability parameters are calculated and used for evaluating the autonomic nerve function state. The method has the effect of improving the monitoring comfort of special crowds.

Description

Technical Field

[0001] This application relates to the field of terahertz radar signal processing technology, and in particular to a method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters. Background Technology

[0002] Hypertension, as one of the major chronic diseases worldwide, has a high prevalence. If left untreated and untreated, it can lead to severe damage and functional decline in multiple organs and systems, including hypertensive heart disease (HHD), stroke, and renal insufficiency. In HHD patients, continued deterioration of cardiac structure and function significantly increases the risk of adverse cardiovascular events such as arrhythmias. Heart rate variability (HRV) analysis based on traditional Holter monitoring can reflect and assess the function of the autonomic nervous system and has been widely used in patients with hypertension and other cardiovascular diseases. Studies on the role of the autonomic nervous system in the occurrence of hypertension and blood pressure regulation have confirmed a positive correlation between autonomic nervous system dysfunction and hypertension. Therefore, monitoring and following up on HRV in HHD patients is particularly important and can help detect and predict the occurrence of arrhythmias. However, traditional ECG monitoring is time-consuming, which is not conducive to outpatient monitoring and follow-up. Some clinical studies have investigated the impact of autonomic nervous function on cardiac regulation and the occurrence of arrhythmias by analyzing short-term HRV data from ECG monitoring. Summary of the Invention

[0003] To address the issues of long processing times and unsuitability for outpatient monitoring and follow-up using traditional electrocardiogram (ECG) monitoring, this application provides a method for preventing and monitoring blood pressure based on terahertz radar ECG heart rate variability parameters.

[0004] This application provides a method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters, employing the following technical solution: Methods for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters include: S1: The intermediate frequency signal is obtained by ADC sampling of the terahertz radar signal; S2: Range-dimensional Fourier transform and phase extraction based on intermediate frequency signals; S3: Based on the reference heartbeat signal frequency range, perform a bandpass filter and ACO-VMD algorithm to distinguish between respiratory and heartbeat signals; S4: Heartbeat signal input encoder, introducing a multi-head attention mechanism module to enhance feature extraction and robustness; S5: Perform a one-dimensional transposed convolution on the attention output signal to recover features while preserving correlation and resolution; S6: The model uses a two-level long short-term memory network to model the features at a deep level and introduces a fully connected layer to capture the dynamic changes in the time dimension of the heartbeat signal; S7: Detect the R peak of the reconstructed radar ECG signal output and analyze parameters such as the standard deviation of sinus interval; S8: Analyze the frequency domain characteristics of radar ECG signals and detect the neural activity of the sympathetic and parasympathetic nervous systems.

[0005] Optionally, in step S1, the intermediate frequency signal model obtained by ADC sampling of the terahertz radar signal is represented as follows: ;in, The amplitude of the intermediate frequency signal. The starting frequency of the signal. For frequency modulation period, For frequency modulation bandwidth, This is the time delay between the radar echo signal and the transmitted signal.

[0006] Optionally, step S2, which involves performing a range-dimensional Fourier transform and phase extraction based on the intermediate frequency signal, further includes: The acquired intermediate frequency signals are arranged into a two-dimensional data matrix consisting of the sampling interval and the frequency modulation period. A distance-dimensional Fourier transform is performed on the sampling interval dimension, and the FFT signal is subjected to phase arctangent and phase unwrapping. The resulting phase signal model is expressed as: ;in, For the pleural cavity micro-displacement signal within the range determined by the distance-dimensional Fourier transform in step S2, extract its phase. .

[0007] Optionally, in step S3, a preset reference heartbeat signal frequency range lookup table is used. Based on the reference heartbeat signal frequency range, a bandpass filter and signal processing are performed to distinguish between respiratory and heartbeat signals. This also includes... Based on the characteristics of vital signs, the phase signal is transformed into the frequency domain, with the respiratory frequency domain ranging from 0.1 to 0.5 Hz and the heart rate frequency domain ranging from 0.8 to 2 Hz. Simple separation is achieved using two 8th-order infinite impulse response bandpass filters, while variational mode decomposition and ant colony optimization algorithms are incorporated. The ACO algorithm is used to optimize the parameter setting range of the VMD algorithm. The probabilistic model for calculating the envelope entropy of multiple permutations and combinations is expressed as follows: ;in For pheromone weights, As a heuristic factor weight, the probability of the path with the minimum envelope entropy is determined by the pheromone concentration. and heuristic factors Joint decision; After each round of iterative calculation, the parameter combination is updated to obtain the parameter combination with minimum envelope entropy. This combination is then used to decompose the modal components, ensuring that each modal component has a center frequency and finite bandwidth, and that the sum of the bandwidths of all modes is minimized. The model is expressed as follows: ;in For each modal function, For the center frequencies of each mode, The unit impulse function, It is a convolution symbol, where the constraint is: the sum of all modes equals the original heartbeat signal. .

[0008] Optionally, to improve the signal reconstruction accuracy and the strictness of the constraints, a quadratic penalty factor can be introduced. and Lagrange multiplier This constrained variational problem is transformed into an unconstrained variational problem, and then iteratively updated. , , After multiple iterations, determine whether the change in modality or center frequency after each iteration meets the convergence condition of being less than a threshold. .

[0009] Optionally, in steps S4, S5, and S6, a short-time Fourier transform is used to convert the frequency domain signal to a time domain signal with a signal length of T, which is then used as the network input. Convolutional layers, batch normalization, Tanh activation function, and max pooling are used to extract and downsample the input signal, obtaining stable and discriminative low-level features. All the obtained output features are stacked into a feature matrix H. The model is represented as follows: ;in, For the first Output features of the output after a convolutional operation It is the sequence length. It is a location Feature dimensions; Perform a linear projection on matrix H. Assume there are M heads. For each head, calculate the query, key, and value. Calculate the scaling dot product attention for each head. The model is represented as: ;in, Let m be the learnable parameter matrix for the m-th head. It is the dimension of the query key in each header; The weighted and fused M head outputs are concatenated along one dimension to integrate their diverse features into a joint model, which is represented as: ; Mapping the concatenated vector back to the original target signal feature dimension, and then multiplying it by a projection matrix, the model is represented as: ; This is the output feature matrix after the multi-head attention mechanism, where These are learnable parameters in a multi-head attention layer; Feature tensors output by the attention mechanism module Feature recovery and upsampling are performed on the signal through batch normalization and Tanh activation function. The convolutional kernel size model is then represented as follows: ;in, Indicates the kernel size. Input the number of channels. Number of output channels; The output model of the transposed convolution is then represented as: Where s is the model step size; This represents the reconstructed output at time t. The model uses a long short-term memory network to perform deep temporal modeling of features, input... The update process is performed within the Long Short-Term Memory (LSTM) network. A fully connected layer maps high-level temporal features to specific ECG signal values. The model is represented as follows: ;in, and These are the weight matrix and bias vector of the fully connected layer. Let be the predicted value of the ECG signal at time t.

[0010] Optionally, in steps S7 and S8, the signal processing of the ECG signal mainly includes two parts: removing various noises present in the ECG signal and detecting the R-peak. The algorithm followed in the R-peak detection is as follows: Let the data to be denoised be a set of n{x} t y t A group of points is defined as (j=1, ..., n), where x is a variable (independent), and y is a variable (independent). t These are observations; a set of m convolution coefficients C i Reapply to the data; After denoising, the peak value of the signal R is found by using the tangential function of the signal. The amplitude axis is redefined from its original range to a new range, namely from 0 to... The multiplier is called the scaling factor; this scaling factor can be defined as: ; This indicates that the value is slightly larger than the maximum value on the signal's y-axis; by multiplying the y-axis by this factor, the y-axis will be transformed into another scale, where large values ​​become larger and small values ​​become smaller; For the detected peak positions R, the peak difference between them is obtained by subtraction, and samples are plotted based on this difference. The resulting vector yields changes in heart rate, which can be used to construct a heart rate variability signal. After constructing the heart rate variability signal, it is analyzed; linear and nonlinear features are extracted from it, and the linear features are further divided into time domain and frequency domain.

[0011] Optionally, the time-domain feature is defined as the feature calculated over time; the average of the RR changes for all samples, the standard deviation of the peak-to-peak distance for all samples; and the RR distance greater than 50 milliseconds. The number of consecutive samples; The value is calculated using the following expression: In frequency domain analysis, the heart rate variability signal, initially defined in the time domain, is first transformed into the frequency domain; this is accomplished by performing a Fourier transform. When any signal is converted to the frequency domain, in that domain, the signal is represented by multiple components of different frequencies; When plotting heart rate variability signals in the frequency domain, the frequency domain plot is divided into three frequency bands: extremely low frequency ≤ 0.04 Hz; low frequency 0.04-0.4 Hz; and high frequency > 0.4 Hz. At this time, the proportion of extremely low frequency and low frequency in the heart rate variability signal is greater than that of the other frequencies.

[0012] In summary, this application includes at least one of the following beneficial technical effects: By monitoring and analyzing 5-minute electrocardiograms of HHD patients, we aim to understand the role of various time-domain and frequency-domain parameters of heart rate variability in assessing patients' autonomic nervous function, and to explore whether they are consistent with heart rate variability, so as to understand whether more flexible short-term HRV detection has the potential to be applied in assessing the regulation of autonomic nervous function in hypertensive patients. Combining terahertz radar signal processing methods, the original radar signal is first preprocessed to suppress respiratory harmonic interference, achieving high-precision separation of the heartbeat signal. Secondly, a multimodal deep learning framework is constructed to achieve end-to-end mapping from radar signal to ECG signal. Finally, the R-peak of the obtained radar ECG signal is detected, and its time-domain and frequency-domain HRV characteristic parameters are calculated for comparison with those of normal individuals, enabling early prevention and rapid monitoring of disease. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 Here is a flowchart example illustrating the general implementation of this application.

[0015] Figure 2This is an example of a flowchart for R-peak detection in the embodiments of this application.

[0016] Figure 3 The diagrams show the KubiosHRV results and algorithm implementation results in the embodiments of this application. Detailed Implementation

[0017] The following is in conjunction with the appendix Figure 1-3 This application will be described in further detail.

[0018] This application discloses a method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters.

[0019] A non-contact measurement device and signal reconstruction method for human electrocardiogram based on terahertz radar, the method including the following steps: S1: The intermediate frequency (IF) signal is obtained by sampling the terahertz radar signal using an ADC. In this implementation, the terahertz radar module transmits a radio frequency (RF) signal and receives the radar echo signal reflected from the target human body. After passing through a mixer, the signal is input to a low-pass filter for product and difference operations to remove high frequencies and retain low frequencies, thus obtaining the IF signal.

[0020] In this implementation scheme, the intermediate frequency signal model is represented as: ;in The amplitude of the intermediate frequency signal. The starting frequency of the signal. For frequency modulation period, For frequency modulation bandwidth; This is the time delay between the radar echo signal and the transmitted signal.

[0021] S2: Perform range-fourth ...

[0022] In this implementation, multiple sets of intermediate frequency signals are arranged into a two-dimensional matrix according to the sampling interval (fast time) and the frequency modulation period (slow time). A range-fourth transform (Range-FFT) is performed on the fast time dimension to obtain the energy distribution of the target in the range dimension, thereby determining the range cell where the human chest cavity is located. To further enhance the weak heartbeat signal, the complex signal of this range cell is subjected to arctangent operation to extract the phase, and phase unwrapping processing is performed to eliminate distortion caused by phase jumps.

[0023] Specifically, the acquired multiple sets of intermediate frequency signals are arranged into a two-dimensional data matrix consisting of the sampling interval and the frequency modulation period. A range-fourth transform (Range-FFT) is performed on the sampling interval dimension, and the FFT signal is subjected to phase arctangent and phase unwrapping. The resulting phase signal model is expressed as follows: ;in, For the pleural micro-displacement signal within the range determined by the Range-FFT in step S2, extract its phase. .

[0024] S3: Based on the reference heartbeat signal frequency range, perform a bandpass filter and ACO-VMD algorithm to distinguish between respiratory and heartbeat signals.

[0025] In this implementation, the unwound phase signal is transformed to the frequency domain to identify its frequency components. Based on physiological characteristics, respiratory signals are concentrated in the 0.1–0.5 Hz range, while heartbeat signals are concentrated in the 0.8–2 Hz range. First, two 8th-order IIR bandpass filters are used to initially separate the respiratory and heartbeat signals. To further eliminate the interference of respiratory harmonics on the heartbeat signal, variational mode decomposition (ACO-VMD) optimized by the ant colony algorithm is introduced. The ACO algorithm is used to search for the optimal combination of parameters in VMD (such as the number of modes K and the penalty factor α), and its objective function is to minimize the envelope entropy.

[0026] Specifically, based on the characteristics of vital signs signals, the phase signals are transformed into the frequency domain, with the respiratory frequency domain ranging from 0.1 to 0.5 Hz and the heart rate frequency domain ranging from 0.8 to 2 Hz. Simple separation is achieved using two 8th-order infinite impulse response (IIR) bandpass filters, while variational mode decomposition and ant colony optimization algorithms are also incorporated. The ACO algorithm is used to optimize the parameter setting range of the VMD algorithm. The probabilistic model for calculating the envelope entropy of multiple permutations and combinations is expressed as follows: ;in For pheromone weights, As a heuristic factor weight, the probability of the path with the minimum envelope entropy is determined by the pheromone concentration. and heuristic factors Joint decision; After each round of iterative calculation, the parameter combination is updated to obtain the parameter combination with minimum envelope entropy. This combination is then used to decompose the modal components, ensuring that each modal component has a center frequency and finite bandwidth, and that the sum of the bandwidths of all modes is minimized. The model is expressed as follows: ;in For each modal function, For the center frequencies of each mode, The unit impulse function, It is a convolution symbol, where the constraint is: the sum of all modes equals the original heartbeat signal. .

[0027] To improve the accuracy of signal reconstruction and the strictness of constraints, a quadratic penalty factor is introduced. and Lagrange multiplier This constrained variational problem is transformed into an unconstrained variational problem, and then iteratively updated. , , After multiple iterations, determine whether the change in modality or center frequency after each iteration meets the convergence condition of being less than a threshold. .

[0028] S4: Heartbeat signal input encoder, introducing a multi-head attention mechanism module to enhance feature extraction and robustness.

[0029] In this implementation, the separated heartbeat signal is converted into a time-frequency graph using a short-time Fourier transform (STFT) and used as input to the deep learning model. The input signal first passes through a one-dimensional convolutional layer (Conv1D), batch normalization (BatchNorm), Tanh activation function, and max pooling to extract local features, which are then used as input to the multi-head attention mechanism.

[0030] Multi-head attention mechanisms compute in parallel using multiple attention heads, each learning the dependencies at different positions in the input sequence. Each head independently computes a weighted sum of the query, key, and value, and finally, the outputs of all heads are concatenated and projected back into the original feature space.

[0031] S5: Perform a one-dimensional transpose convolution on the attention output signal to recover features while maintaining correlation and resolution.

[0032] In this implementation, the feature tensor output by the attention mechanism is processed by batch normalization (BatchNorm) and the Tanh activation function before being input into a one-dimensional transposed convolutional layer (Conv1D Transpose). The transposed convolution restores the low-resolution feature map to the original signal length through a learnable upsampling operation, while preserving local spatial correlations.

[0033] S6: The model uses a two-level long short-term memory network (LSTM) to model the features at a deep level and introduces a fully connected layer to capture the dynamic changes in the time dimension of the heartbeat signal.

[0034] In this implementation, the feature sequence output from the transposed convolution is input into a Long Short-Term Memory (LSTM) network. The LSTM network effectively captures long-term dependencies in the time series through a gating mechanism, avoiding the vanishing gradient problem. The hidden state at each time step is passed to the next layer of the LSTM network, ultimately outputting the feature sequence.

[0035] Specifically, in S4-S6, the Short-Time Fourier Transform (STFT) is used to convert the frequency domain signal to a time domain signal with a signal length of T, which is then used as the network input. Convolutional layers, batch normalization (BatchNorm), Tanh activation function, and max pooling (MaxPool) are used to extract and downsample the input signal, obtaining stable and discriminative low-level features. All the obtained output features are then stacked into a feature matrix H. The model is represented as follows: ;in, For the first Output features of the output after a convolutional operation It is the sequence length. It is a location The feature dimensions.

[0036] Linear projection is performed on matrix H. In this implementation, assuming there are M heads, the query, key, and value are calculated for each head, and the scaling dot product attention is calculated for each head. The model is represented as follows: ;in, Let m be the learnable parameter matrix for the m-th head. It is the dimension of the query key in each header; The weighted and fused M head outputs are concatenated along one dimension to integrate their diverse features into a joint model, which is represented as: ; Mapping the concatenated vector back to the original target signal feature dimension, and then multiplying it by a projection matrix, the model is represented as: ; This is the output feature matrix after the multi-head attention mechanism, where These are learnable parameters in a multi-head attention layer; Feature tensors output by the attention mechanism module Feature recovery and upsampling are performed on the signal using batch normalization (BatchNorm) and Tanh activation function. At this point, the convolutional kernel size model representation... ;in, Indicates the kernel size. Input the number of channels. This represents the number of output channels.

[0037] The output model of the transposed convolution is then represented as: Where s is the model step size; This represents the reconstructed output at time t. The model uses a long short-term memory network to perform deep temporal modeling of features, input... The update process is performed within the Long Short-Term Memory (LSTM) network. A fully connected layer maps high-level temporal features to specific ECG signal values. The model is represented as follows: ;in, and These are the weight matrix and bias vector of the fully connected layer. Let be the predicted value of the ECG signal at time t.

[0038] S7: Detect the R peak of the reconstructed radar ECG signal output and analyze parameters such as the standard deviation of sinus interval (SDNN).

[0039] In this implementation scheme, the radar ECG signal output by the model is preprocessed to remove residual noise. A sliding window convolution smoothing method is used to enhance signal quality, and then the signal amplitude is scaled using a tangent function to make the R-peak more prominent. The RR interval is calculated based on the time interval between adjacent R-peaks, and then the time-domain HRV features are extracted, including: the average RR interval; the standard deviation of the RR interval; and the proportion of adjacent RR interval differences greater than 50ms.

[0040] S8: Analyze the frequency domain characteristics of radar ECG signals and detect the neural activity of the sympathetic and parasympathetic nervous systems.

[0041] In this implementation scheme, the RR interval sequence is transformed to the frequency domain using Fourier transform to obtain the power spectral density of the HRV signal. Based on the frequency band division criteria, the energy of the following frequency bands is analyzed: VLF: reflects long-term regulatory mechanisms; LF: represents the combined action of the sympathetic and parasympathetic nervous systems; HF: mainly reflects parasympathetic activity. By calculating the LF / HF ratio, the balance between the sympathetic and parasympathetic nervous systems is assessed, thereby determining the degree of autonomic dysfunction in hypertensive patients. This analysis provides an important reference for early warning and dynamic monitoring.

[0042] Specifically, in S7-S8, the signal processing of the ECG signal mainly includes two parts: removing various noises present in the ECG signal and detecting the R-peak. The algorithm followed in the R-peak detection is as follows: Let the data to be denoised be a set of n{x} t y t A group of points is defined as (j=1, ..., n), where x is a variable (independent), and y is a variable (independent). t These are observations; a set of m convolution coefficients C i Reapply to the data; After denoising, the peak value of the signal R is found by using the tangential function of the signal. The amplitude axis is redefined from its original range to a new range, namely from 0 to... The multiplier is called the scaling factor; this scaling factor can be defined as: ; This indicates that the value is slightly larger than the maximum value on the signal's y-axis; by multiplying the y-axis by this factor, the y-axis will be transformed into another scale, where large values ​​become larger and small values ​​become smaller.

[0043] For the detected peak positions R, the peak difference between them is obtained by subtraction, and samples are plotted based on this difference. The resulting vector yields changes in heart rate, which can be used to construct a heart rate variability signal. After constructing the heart rate variability signal, it is analyzed; linear and nonlinear features are extracted from it, and the linear features are further divided into time domain and frequency domain.

[0044] Temporal features are defined as features calculated over time; the mean of the RR changes for all samples; the standard deviation of the peak-to-peak distance for all samples (SDNN); some features can be calculated based on reference values, for example, in this implementation, an RR distance greater than 50 milliseconds (...). The number of consecutive samples; The value is calculated using the following expression: In frequency domain analysis, the heart rate variability signal, initially defined in the time domain, is first transformed into the frequency domain; this is accomplished by performing a Fourier transform. When any signal is converted to the frequency domain, in that domain, the signal is represented by multiple components of different frequencies; When plotting the heart rate variability signal in the frequency domain, the frequency domain plot is divided into three frequency bands: very low frequency (VLF) ≤ 0.04 Hz; low frequency (LF) 0.04-0.4 Hz; and high frequency (HF) > 0.4 Hz. At this time, the proportion of very low frequency and low frequency in the heart rate variability signal is greater than that of the other frequencies. In this embodiment, the HRV signal is mainly composed of very low frequency and low frequency, and the high frequency component is less than that of the other two frequency components.

[0045] The implementation scheme of this application is as follows: A non-contact terahertz radar device generates a triangular wave through a digital signal processor, inputs it into a voltage-controlled oscillator to generate a high-frequency continuous wave transmission signal, which is then transmitted through an amplifier and a transmitting antenna. At the same time, a portion of the transmission signal is sent to a mixer as a local oscillator signal through a power divider.

[0046] After the transmitted signal encounters the target, it is scattered back and enters the mixer through the receiving antenna and low-noise amplifier. After being mixed with the local oscillator signal, a lower frequency intermediate frequency signal is generated, which includes the required distance information and irrelevant interference. After A / D sampling, the target distance information is extracted by the digital signal processor.

[0047] In the actual system measurement process, the target detection distance range was set to 0.5 meters to 1 meter. Real-time data acquired by the radar was transmitted at high speed to the PC via UART serial protocol for signal processing. Simultaneously, to facilitate the non-contact radar ECG signal experiment, Qt was used to develop the host computer interface, enabling control and data acquisition of the front-end electronics system. This allowed for PC control, sending specified commands to the driver program via USB interface, the radar chip command parsing module, and finally, the configuration of radar parameters.

[0048] The complex signals acquired by the radar serve as input for subsequent core algorithms, used for target detection and feature analysis. After data processing is complete, the entire system automatically clears the memory buffer to ensure the real-time performance and continuity of data acquisition in the next cycle.

[0049] The algorithm extracts the phase signal from the radar complex signal and incorporates ACO-VMD to separate respiration and heartbeat. Subsequently, an end-to-end structure of CNN+LSTM is selected. CNN+LSTM is a hybrid neural network architecture that combines the advantages of convolutional neural networks (CNN) and long short-term memory networks (LSTM), specifically designed to process data that simultaneously possesses spatial structure (or local patterns) and temporal order.

[0050] After selecting the end-to-end structure of CNN+LSTM, a multi-head attention mechanism is added. Finally, the output radar ECG signal is analyzed by detecting the position of the R peak, and its time-domain and frequency-domain characteristic parameters are calculated. The accuracy of the results is confirmed by KubiosHRV, a software tool used to analyze heart rate variability (HRV).

[0051] This implementation scheme employs non-contact monitoring, improving patient comfort and applicability. Traditional monitoring requires direct skin contact through electrodes, which presents problems such as poor comfort, skin inflammation, and inconvenience for burn patients, infants, people with skin allergies, or those requiring long-term dynamic monitoring. This invention utilizes the strong penetrating power and non-contact characteristics of terahertz radar to reconstruct electrocardiograms and analyze heart rate variability (HRV) by detecting weak mechanical vibrations on the surface of the chest cavity, completely avoiding physical electrode contact. This not only significantly improves the monitoring comfort and safety for special populations but also provides a seamless and unrestricted monitoring method for scenarios such as sleep monitoring, home-based elderly care, and telemedicine.

[0052] A multimodal deep learning framework enables high-precision end-to-end signal reconstruction: This invention constructs an end-to-end deep learning model that integrates convolutional neural networks, long short-term memory networks, and multi-head attention mechanisms. This framework can automatically extract global and local correlation information from radar time-frequency features, effectively suppress noise and respiratory harmonic interference, and achieve accurate mapping from chest cavity mechanical vibration signals to standard electrocardiogram waveforms, significantly improving the accuracy and robustness of signal reconstruction.

[0053] This method possesses both early prevention and dynamic monitoring potential for clinical application. It achieves non-contact acquisition of ECG waveforms and further performs R-peak detection and time-frequency domain heart rate variability (HRV) characteristic analysis on the reconstructed signal, enabling quantitative assessment of the balance between the sympathetic and parasympathetic nervous systems. By analyzing key indicators such as SDNN, PNN50, and LF / HF, it can assist in assessing the degree of autonomic nerve function impairment in hypertensive patients, providing objective evidence for early risk warning. Simultaneously, this method supports rapid short-term analysis and long-term dynamic follow-up. Compared to traditional 24-hour Holter monitoring, it is more suitable for rapid outpatient screening and daily home monitoring, demonstrating significant clinical application value. High-precision validation and clinical application potential are also demonstrated.

[0054] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters, characterized in that, include: S1: The intermediate frequency signal is obtained by ADC sampling of the terahertz radar signal; S2: Range-dimensional Fourier transform and phase extraction based on intermediate frequency signals; S3: Based on the reference heartbeat signal frequency range, perform a bandpass filter and ACO-VMD algorithm to distinguish between respiratory and heartbeat signals; S4: Heartbeat signal input encoder, introducing a multi-head attention mechanism module to enhance feature extraction and robustness; S5: Perform a one-dimensional transposed convolution on the attention output signal to recover features while preserving correlation and resolution; S6: The model uses a two-level long short-term memory network to model the features at a deep level and introduces a fully connected layer to capture the dynamic changes in the time dimension of the heartbeat signal; S7: Detect the R peak of the reconstructed radar ECG signal output and analyze parameters such as the standard deviation of sinus interval; S8: Analyze the frequency domain characteristics of radar ECG signals and detect the neural activity of the sympathetic and parasympathetic nervous systems.

2. The method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters according to claim 1, characterized in that: In step S1, the intermediate frequency signal model obtained by ADC sampling of the terahertz radar signal is represented as follows: ;in, The amplitude of the intermediate frequency signal. The starting frequency of the signal. For frequency modulation period, For frequency modulation bandwidth, This is the time delay between the radar echo signal and the transmitted signal.

3. The method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters according to claim 1, characterized in that: In step S2, the distance-dimensional Fourier transform and phase extraction based on the intermediate frequency signal also include, The acquired intermediate frequency signals are arranged into a two-dimensional data matrix consisting of the sampling interval and the frequency modulation period. A distance-dimensional Fourier transform is performed on the sampling interval dimension, and the FFT signal is subjected to phase arctangent and phase unwrapping. The resulting phase signal model is expressed as: ;in, For the pleural cavity micro-displacement signal within the range determined by the distance-dimensional Fourier transform in step S2, extract its phase. .

4. The method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters according to claim 3, characterized in that: In step S3, a preset reference heartbeat signal frequency range lookup table is used. Based on the reference heartbeat signal frequency range, a bandpass filter and signal processing are performed to distinguish between respiratory and heartbeat signals. This also includes... Based on the characteristics of vital signs, the phase signal is transformed into the frequency domain, with the respiratory frequency domain ranging from 0.1 to 0.5 Hz and the heart rate frequency domain ranging from 0.8 to 2 Hz. Simple separation is achieved using two 8th-order infinite impulse response bandpass filters, while variational mode decomposition and ant colony optimization algorithms are incorporated. The ACO algorithm is used to optimize the parameter setting range of the VMD algorithm. The probabilistic model for calculating the envelope entropy of multiple permutations and combinations is expressed as follows: ;in For pheromone weights, As a heuristic factor weight, the probability of the path with the minimum envelope entropy is determined by the pheromone concentration. and heuristic factors Joint decision; After each round of iterative calculation, the parameter combination is updated to obtain the parameter combination with minimum envelope entropy. This combination is then used to decompose the modal components, ensuring that each modal component has a center frequency and finite bandwidth, and that the sum of the bandwidths of all modes is minimized. The model is expressed as follows: ;in For each modal function, For the center frequencies of each mode, The unit impulse function, It is a convolution symbol, where the constraint is: the sum of all modes equals the original heartbeat signal. .

5. The method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters according to claim 4, characterized in that: To improve the accuracy of signal reconstruction and the strictness of constraints, a quadratic penalty factor is introduced. and Lagrange multiplier This constrained variational problem is transformed into an unconstrained variational problem, and then iteratively updated. , , After multiple iterations, determine whether the change in modality or center frequency after each iteration meets the convergence condition of being less than a threshold. .

6. The method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters according to claim 4, characterized in that: In steps S4, S5, and S6, a short-time Fourier transform is used to convert the frequency domain signal to a time domain signal. The signal length is T, which is used as the network input. Convolutional layers, batch normalization, Tanh activation function, and max pooling are used to extract and downsample the input signal, obtaining stable and discriminative low-level features. All the obtained output features are stacked into a feature matrix H. The model is represented as follows: ;in, For the first Output features of the output after a convolutional operation It is the sequence length. It is a location Feature dimensions; Perform a linear projection on matrix H. Assume there are M heads. For each head, calculate the query, key, and value. Calculate the scaling dot product attention for each head. The model is represented as: ;in, Let m be the learnable parameter matrix for the m-th head. It is the dimension of the query key in each header; The weighted and fused M head outputs are concatenated along one dimension to integrate their diverse features into a joint model, which is represented as: ; Mapping the concatenated vector back to the original target signal feature dimension, and then multiplying it by a projection matrix, the model is represented as: ; This is the output feature matrix after the multi-head attention mechanism, where These are learnable parameters in a multi-head attention layer; Feature tensors output by the attention mechanism module Feature recovery and upsampling are performed on the signal through batch normalization and Tanh activation function. The convolutional kernel size model is then represented as follows: ;in, Indicates the kernel size. Input the number of channels. Number of output channels; The output model of the transposed convolution is then represented as: Where s is the model step size; This represents the reconstructed output at time t. The model uses a long short-term memory network to perform deep temporal modeling of features, input... The update process is performed within the Long Short-Term Memory (LSTM) network. A fully connected layer maps high-level temporal features to specific ECG signal values. The model is represented as follows: ;in, and These are the weight matrix and bias vector of the fully connected layer. Let be the predicted value of the ECG signal at time t.

7. The method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters according to claim 6, characterized in that: In steps S7 and S8, the signal processing of the ECG signal mainly includes two parts: removing various noises present in the ECG signal and detecting the R-peak. The algorithm followed in the R-peak detection is as follows: Let the data to be denoised be a set of n{x} t y t A group of points is defined as (j=1, ..., n), where x is a variable (independent), and y is a variable (independent). t These are observations; a set of m convolution coefficients C i Reapply to the data; After denoising, the peak value of the signal R is found by using the tangential function of the signal. The amplitude axis is redefined from its original range to a new range, namely from 0 to... The multiplier is called the scaling factor; this scaling factor can be defined as: ; ; This indicates that the value is slightly larger than the maximum value on the signal's y-axis; by multiplying the y-axis by this factor, the y-axis will be transformed into another scale, where large values ​​become larger and small values ​​become smaller; For the detected peak positions R, the peak difference between them is obtained by subtraction, and samples are plotted based on this difference. The resulting vector yields changes in heart rate, which can be used to construct a heart rate variability signal. After constructing the heart rate variability signal, it is analyzed; linear and nonlinear features are extracted from it, and the linear features are further divided into time domain and frequency domain.

8. The method for preventing and monitoring blood pressure based on terahertz radar electrocardiogram heart rate variability parameters according to claim 7, characterized in that: Time-domain features are defined as features calculated over time; the average of the RR changes for all samples, and the standard deviation of the peak-to-peak distance for all samples; RR distance greater than 50 milliseconds ( The number of consecutive samples; The value is calculated using the following expression: In frequency domain analysis, the heart rate variability signal, initially defined in the time domain, is first transformed into the frequency domain. This is accomplished by performing a Fourier transform; When any signal is converted to the frequency domain, in that domain, the signal is represented by multiple components of different frequencies; When plotting heart rate variability signals in the frequency domain, the frequency domain plot is divided into three frequency bands: extremely low frequency ≤ 0.04 Hz; low frequency 0.04-0.4 Hz; and high frequency > 0.4 Hz. At this time, the proportion of extremely low frequency and low frequency in the heart rate variability signal is greater than that of the other frequencies.

Images