A single-channel ppg non-invasive blood pressure monitoring system based on generative ecg enhancement

Abstract

The application relates to the field of noninvasive blood pressure monitoring and relates to a single-channel PPG noninvasive blood pressure monitoring system based on generative ECG enhancement, which comprises a data preprocessing module, an ECG signal and a PPG signal are preprocessed to obtain a PPG-ECG signal sample and a PPG signal sample with a blood pressure label; an ECG generation model is used for realizing cross-modal physiological signal generation from PPG to ECG to obtain a time-domain-aligned generative ECG signal; and a blood pressure prediction model is used for constructing a feature extractor based on an improved U-Net architecture, embedding a multi-scale convolution module and a cross-modal attention fusion module, inputting extracted space-time features into a mapping regressor, outputting continuous prediction values of systolic pressure and diastolic pressure, and realizing end-to-end blood pressure regression. The application realizes cross-modal data enhancement by constructing a time-domain-aligned generative ECG signal, and simultaneously integrates a multi-scale convolution and a cross-modal attention fusion module in the prediction model, thereby breaking through the precision limitation of single-channel PPG blood pressure prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-invasive blood pressure monitoring, in particular to a single-channel PPG non-invasive blood pressure monitoring system based on generative ECG enhancement. BACKGROUND

[0002] Hypertension, as a common chronic disease, is an important risk factor for cardiovascular diseases. Regular blood pressure monitoring plays a key role in the early detection, prevention and treatment of cardiovascular diseases. Current clinical monitoring methods are mainly divided into two categories: invasive and non-invasive. Invasive monitoring has the risk of infection and is complex to operate, making it difficult to apply to daily monitoring. Traditional non-invasive monitoring relies on inflatable cuff compression of blood vessels, which not only affects blood circulation and wearing comfort, but also cannot achieve continuous blood pressure monitoring. This makes it an important research direction to develop non-invasive continuous blood pressure monitoring algorithms based on sleeveless devices.

[0003] Current non-invasive blood pressure monitoring algorithms have a double bottleneck: at the signal acquisition level, electrocardiogram (ECG) can provide accurate heart cycle information, but is subject to the application threshold of professional medical settings, making it difficult to apply to daily monitoring; photoplethysmogram (PPG) is simple to collect and low in cost, but its lack of cardiac electrophysiological characteristics limits the accuracy of single-channel PPG prediction algorithms. At the algorithm level, generating electrocardiogram (ECG) to supplement key hemodynamic parameter information can effectively improve the performance of single-channel PPG blood pressure monitoring. However, the PPG-ECG generation network based on CycleGAN can ensure the accuracy of the RR interval feature of ECG, but the generated signal has a pulse wave transmission time (PAT) delay problem, which makes the blood pressure monitoring model unable to learn the correct hemodynamic characteristics, thereby seriously affecting the accuracy of the final blood pressure monitoring. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a single-channel PPG non-invasive blood pressure monitoring system based on generative ECG, which integrates a PPG-ECG cross-modal generation network constrained by hemodynamic medical knowledge to construct a time-domain aligned generative ECG signal for cross-modal data generation, and integrates multi-scale convolution and cross-modal attention fusion modules to break through the accuracy limitations of single-channel PPG blood pressure prediction.

[0005] In an embodiment of the present application, a single-channel PPG non-invasive blood pressure monitoring system based on generative ECG includes the following modules:

[0006] A data preprocessing module pre-processes ECG signals used to train an ECG generation model and PPG signals used to train a blood pressure prediction model to obtain PPG-ECG signal samples and PPG signal samples with blood pressure labels;

[0007] An ECG generation model is a PPG-ECG cross-modal generation network fused with hemodynamic medical knowledge constraints, used to realize cross-modal physiological signal generation from PPG to ECG; a PPG signal sample with a blood pressure label is input into the trained cross-modal generation network to obtain a time-domain aligned generated ECG signal;

[0008] A blood pressure prediction model is constructed based on an improved U-Net architecture to construct a feature extractor, and a multi-scale convolution module and a cross-modal attention fusion module are embedded in the encoding-decoding structure thereof, time-space features extracted are input into a mapping regressor, and finally continuous prediction values of systolic pressure and diastolic pressure are output; the blood pressure prediction model realizes end-to-end blood pressure regression by jointly inputting a PPG signal sample with a blood pressure label and a generated ECG signal thereof.

[0009] Compared with the prior art, the main technical effects achieved by the present application include:

[0010] The single-channel PPG non-invasive blood pressure monitoring system based on the generated ECG of the present application realizes cross-modal data enhancement by constructing a time-domain aligned generated ECG signal, and breaks through the precision limitation of single-channel PPG blood pressure prediction by integrating a multi-scale convolution and a cross-modal attention fusion module in the prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is the architecture diagram of the non-invasive blood pressure monitoring system in the embodiment of the present application;

[0012] Figure 2 is the architecture diagram of the ECG generation model, wherein (a) is the architecture diagram of the generator G E , and (b) is the architecture diagram of the generator G P ;

[0013] Figure 3 is the architecture diagram of the blood pressure monitoring model;

[0014] Figure 4 is the structure diagram of the multi-scale time convolution block;

[0015] Figure 5 is the structure diagram of the time convolution block. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] EMBODIMENT

[0018] Referring to Figure 1 The embodiment provides a single-channel PPG non-invasive blood pressure monitoring system based on generative ECG, comprising a data preprocessing module, an ECG generation model and a blood pressure prediction model.

[0019] The data preprocessing module preprocesses the ECG signal used for training the ECG generation model and the PPG signal used for training the blood pressure prediction model to obtain PPG-ECG signal samples and PPG signal samples with blood pressure labels.

[0020] The ECG generation model is a PPG-ECG cross-modal generation network that fuses blood flow dynamics medical knowledge constraints, and can realize cross-modal physiological signal generation from PPG to ECG. The ECG generation model is based on the CycleGAN architecture, adopts a WaveNet-style generator, a PatchGAN-style frequency domain discriminator and a standard time domain discriminator, and combines PPG-ECG signal samples for model training to realize cross-modal physiological signal generation from PPG to ECG; the PPG signal samples with blood pressure labels are input into the trained PPG-ECG cross-modal generation network to obtain time-domain aligned generative ECG signals.

[0021] The blood pressure prediction model constructs a feature extractor based on an improved U-Net architecture, embeds a multi-scale convolution module and a cross-modal attention fusion module in the encoding-decoding structure thereof, inputs the extracted spatio-temporal features into a mapping regressor, and finally outputs continuous prediction values of systolic blood pressure (SBP) and diastolic blood pressure (DBP). In the embodiment, the blood pressure prediction model realizes end-to-end blood pressure regression by jointly inputting the PPG signal samples with blood pressure labels and their generative ECG signals.

[0022] The data preprocessing module also performs sampling rate synchronization, filtering and denoising and normalization processing on the original PPG-ECG data set and the PPG-blood pressure data set, and finally outputs PPG-ECG signal samples and PPG signal samples with blood pressure labels. The ECG generation model is trained in combination with the PPG-ECG signal samples to obtain an optimized network that can stably generate ECG. The PPG signal samples with blood pressure labels are input into the trained ECG generation model to output generative ECG signals that are time-domain aligned with the original PPG signals. Finally, the blood pressure prediction model is trained in combination with the generative ECG signals and the PPG signal samples with blood pressure labels to output systolic blood pressure (SBP) and diastolic blood pressure (DBP) prediction values.

[0023] More specifically, the preprocessing steps of the data preprocessing module are as follows:

[0024] (1) Synchronize the sampling rates of different data sets to 128 Hz using cubic spline interpolation.

[0025] (2) The PPG signal samples are filtered by a fourth-order Butterworth filter with a frequency range of 0.5-8 Hz, and the ECG signal is filtered by a band-pass filter with a frequency range of 3-45 Hz. Finally, a zero-phase filter is used to preserve the timing accuracy of the PPG signal and ECG signal, avoiding the influence of phase delay on the signal morphology.

[0026] (3) The ECG and PPG signal segments are normalized by z-score, and then each signal segment is normalized to the interval [-1, 1] by min-max normalization.

[0027] The structure of the ECG generation model is shown in Figure 2 , which includes two generators and four discriminators. G E is a generator for generating ECG from PPG, as shown in Figure 2 (a), and G P is a generator for generating PPG from ECG, as shown in Figure 2 (b). For generator G E , there are corresponding time-domain discriminators and frequency-domain discriminators For generator G P , there are corresponding time-domain discriminators and frequency-domain discriminators The input of the time-domain discriminators is the time series signal, and the input of the frequency-domain discriminators is the frequency spectrum obtained by frequency domain conversion of the PPG signal and ECG signal using short-time Fourier transform (STFT).

[0028] Generators G E and G PAll adopt the WaveNet architecture, which specifically includes a starting convolution module, a core residual module, and an output module. The starting convolution module is composed of a 1x1 pointwise convolution layer, which receives a single-channel one-dimensional time-series input signal and linearly maps it to a high-dimensional feature space of 64 channels, preparing for subsequent residual learning. The residual module includes ten sequentially stacked residual blocks, each of which includes a one-dimensional dilated convolution layer, a nonlinear activation layer, and a normalization layer connected in turn. The one-dimensional dilated convolution layer has a kernel size of 3, and the input and output channel numbers are both 64, and the dilated factor increases exponentially with the model depth (1, 2, 4, 8, …, 512). This design allows the receptive field of the convolution network to expand rapidly, effectively capturing the long-term dependencies of the signal. After the dilated convolution, a ReLU activation function is applied to increase the non-linear representation capability of the model. Then, a batch normalization layer is applied to stabilize the training process and accelerate model convergence. The input and output of each residual block are added element by element to form a residual connection, ensuring effective gradient propagation. Finally, the output module projects the features extracted by the residual module back to a single-channel feature through a ReLU activation function and a 1x1 pointwise convolution layer, and finally generates the target time-series waveform.

[0029] In this embodiment, the input of the time domain discriminator is the PPG signal and the ECG signal, and the input of the frequency domain discriminator is composed of the joint frequency spectrum generated by the short-time Fourier transform (STFT) of the PPG and ECG time series signals. This embodiment defines x[n] as a time series, and the short-time Fourier transform STFT(x[n]) of the time series can be defined as the following formula:

[0030]

[0031] where m represents the step size, ω represents the frequency parameter, w[n] represents the Hamming window function, and the final frequency spectrum is represented as f(x[n]) = log(|X(m, ω)| + θ), θ = 1e -10 for avoiding infinite conditions.

[0032] Time domain discriminator Responsible for processing one-dimensional time-series waveforms, its structure includes four down-sampling modules and a final classification module. Each down-sampling module is composed of a one-dimensional convolution layer with a convolution kernel size of 15 and a stride of 2, an instance normalization layer, and a Leaky-ReLU activation function, where the first module does not have a normalization layer. After four times of down-sampling, a scalar probability is finally output through the Sigmoid function to realize the global judgment of the entire waveform time dependence.

[0033] Frequency domain discriminator The PatchGAN architecture is adopted, which consists of four fully convolutional modules. Each module comprises a 2D convolutional layer with a kernel size of 4, a batch normalization layer, and a Leaky-ReLU activation function. The first three modules have a convolution stride of 2 to achieve downsampling, and the first module does not have a normalization layer. The frequency domain discriminator finally outputs a 13×13 2D confidence map. Each element in the confidence map corresponds to the realism score of a specific local region on the input spectrum, enabling the model to accurately discriminate the local texture and high-frequency details of the signal, guiding the generator to generate more realistic waveform signals.

[0034] The loss function of the ECG generation model is a composite loss function. To address the common ECG waveform timing delay problem in generative models, this embodiment introduces a PAT-based loss term in addition to the existing adversarial loss and cycle consistency loss to constrain the timing consistency of the generated ECG signal. This loss term incorporates clinical knowledge of hemodynamics and can effectively constrain the timing consistency of the generated ECG signal, ensuring the effectiveness of the generated ECG for blood pressure prediction. This embodiment also defines the peak time point of the real ECG signal as... The peak time point of the generative ECG signal is K pairs of matching peaks are established through dynamic time warping. The PAT loss function is then constructed as follows:

[0035]

[0036] The PAT loss function measures the offset of the peaks of the generative ECG signal relative to the peaks of the real ECG signal, thereby optimizing timing alignment performance. This is achieved by minimizing... The model can effectively correct the output delay caused by network structure or training strategy, making the generated ECG more consistent with the real physiological process.

[0037] The final objective loss function of the GAN network is calculated as follows:

[0038]

[0039] in and These are the adversarial losses of the time-domain discriminator and the frequency-domain discriminator against the generator, respectively, with α and β being the corresponding coefficients; is the cycle consistency loss, λ is the corresponding coefficient; ω is the PAT loss coefficient. The coefficients of the adversarial loss and the cycle consistency loss are relatively fixed, while the weight coefficient of the PAT loss is dynamically adjusted using the Logstic function with the training round. In the early stage of training, the PAT loss weight coefficient is close to 0, so that the model pays more attention to the generation of the overall waveform, and in the later stage of training, as the waveform structure is initially formed, the weight of the PAT loss is gradually increased to fine-tune the R wave peak position in the generated ECG signal. The coefficient dynamic adjustment function is defined as follows:

[0040]

[0041] wherein epoch is the current training round, ω max is the final PAT loss coefficient, m is the curve center, i.e. approximately in the mthepoch coefficient grows to the middle value, k is used to control the growth rate.

[0042] In the ECG generation model training process, the generator receives the paired PPG and ECG samples and outputs ECG and PPG generated signals respectively, and then inputs the generated signals into the generator again for inverse reconstruction to obtain reconstructed PPG and ECG signals. The cycle consistency loss is calculated by comparing the difference between the original sample and the reconstructed signal, and the PAT loss is calculated based on the R wave time difference between the real ECG and the generated ECG. The discriminator discriminates the authenticity of the original PPG / ECG sample and its generated signal and outputs the adversarial loss. In the training process, the generator parameters are updated and optimized by the cycle consistency loss, the PAT loss and the adversarial loss, while the discriminator parameters are updated independently according to the respective modal adversarial loss.

[0043] The architecture of the blood pressure prediction model is shown in Figure 3 The feature extractor uses an encoder and decoder architecture based on U-Net, and the CNN layer of the decoder is replaced by a multi-scale convolution layer. The skip connection between the decoder and the encoder embeds a cross-modal attention fusion module to weight the extracted PPG and ECG signal features, thereby narrowing the semantic gap.

[0044] The PPG sample with blood pressure label and the ECG sample generated by the ECG generation model are first subjected to feature extraction by the encoder. The encoder uses multi-scale convolution instead of CNN layer to enhance the standard U-Net structure, and the entire encoder includes four multi-scale temporal convolution modules (Multiscale Temporal Convolution Block). The structure of the multi-scale temporal convolution module is as follows: Figure 4As shown, each convolution module realizes multi-scale feature fusion through a double-branch parallel convolution structure. The local feature branch adopts a convolution layer with a kernel of 3, and the global feature branch uses a convolution layer with a kernel of 7. The features extracted by the double-branch parallel convolution structure are added and fused after batch normalization and ReLU activation function, and finally down-sampling is performed using maximum pooling. The number of channels of the convolution module increases layer by layer.

[0045] In the present embodiment, the cross-modal attention fusion module of the feature extractor performs weighted calculation on the extracted PPG and ECG signal features, and the specific steps include:

[0046] (1) The PPG and ECG features of each layer extracted from the multi-scale time convolution module are converted into feature tokens through average pooling and deep convolution with a kernel size of 1, (P is the fixed patch number); for each feature token T i , a query Q i is obtained through deep convolution projection; C The feature tokens of all stages are spliced along the channel dimension to generate T i , and the keys K and values V are obtained through deep convolution projection; the cross-channel attention (CCA) is calculated through the following formula:

[0047]

[0048] where Q c is the query, K and V are global channel information, C i is the total number of channels, is the transpose of the query Q T , and V i is the transpose of the value V; the output is the weighted channel feature. The feature weighted by cross-channel attention is obtained through residual connection.

[0049] (2) Spatial cross-attention (SCA) is used to capture the long-distance dependence of multi-scale features in the spatial dimension, and the features output by CCA are spliced along the channel dimension to and the query Q and key K are generated through deep convolution, and V is obtained by deep convolution from the features output by CCA c .

[0050]

[0051] where h T is the number of attention heads; K is the transpose of the key K; the output is the weighted spatial feature. The feature weighted by spatial attention is obtained through residual connection.

[0052] (3) The features weighted by cross-modal attention are processed by deep convolution, layer normalization and GELU activation, and finally connected to the decoder by upsampling.

[0053] The decoder of the feature extractor includes four temporal convolution blocks. The input of each temporal convolution block is obtained by concatenating the features obtained by the skip connection and the up-sampled features output by the previous temporal convolution block. The structure of the temporal convolution block is shown in Figure 5 Each temporal convolution block includes two one-dimensional convolution layers with a kernel size of 3. The first convolution layer expands the channel dimension, and the second convolution layer remains unchanged. The temporal convolution block adopts a layer-by-layer channel decreasing design (128, 64, 32, 16). The final extracted feature shape is (16, 848).

[0054] The mapping regressor of the blood pressure prediction module mainly includes a one-dimensional global average pooling layer and a multi-layer perception head (MLPHead). The output of the feature extractor is first aggregated into a multi-dimensional (e.g., 16-dimensional) global feature vector in the time sequence dimension by the one-dimensional global average pooling, and then input into the multi-layer perception head for nonlinear transformation, mapped to a single output node, to realize the prediction of DBP and SBP.

[0055] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application are equivalent replacement methods and are included in the protection scope of the present application.

Claims

1. A single-channel PPG non-invasive blood pressure monitoring system based on generative ECG enhancement, characterized in that, Includes the following modules: The data preprocessing module preprocesses the ECG signals used to train the ECG generation model and the PPG signals used to train the blood pressure prediction model to obtain PPG-ECG signal samples and PPG signal samples with blood pressure labels. The ECG generation model is a PPG-ECG cross-modal generation network that integrates hemodynamic medical knowledge constraints. It is used to generate cross-modal physiological signals from PPG to ECG. PPG signal samples with blood pressure labels are input into the well-trained cross-modal generation network to obtain time-domain aligned generative ECG signals. The blood pressure prediction model is based on a feature extractor built on an improved U-Net architecture. Multi-scale convolutional modules and cross-modal attention fusion modules are embedded in its encoder-decoder structure. The extracted spatiotemporal features are input into a mapping regressor, and finally, continuous predicted values ​​of systolic and diastolic blood pressure are output. The blood pressure prediction model achieves end-to-end blood pressure regression by jointly inputting PPG signal samples with blood pressure labels and their generative ECG signals. The loss function of the ECG generation model is a composite loss function. Based on adversarial loss and cycle consistency loss, a PAT-based loss term is introduced to constrain the temporal consistency of the generative ECG signal. The peak time point of a real ECG signal is defined as... The peak time point of the generative ECG signal is K-pair matching peaks were established through dynamic time warping. Therefore, the PAT loss function is constructed as follows: ； The PAT loss function measures the degree of shift of the peak of the generative ECG signal relative to the peak of the real ECG signal. The final objective loss function of the GAN network is calculated as follows: ； in and These are the adversarial losses of the time-domain discriminator and the frequency-domain discriminator against the generator. and These are the corresponding coefficients; It is a cycle consistency loss. These are the corresponding coefficients; It is the PAT loss coefficient.

2. The single-channel PPG non-invasive blood pressure monitoring system according to claim 1, characterized in that, The specific preprocessing steps in the data preprocessing module include: (1) Use cubic spline interpolation to synchronize the sampling rates of different datasets; (2) A fourth-order Butterworth filter is used to filter the PPG signal, a bandpass filter is used to filter the ECG signal, and finally a zero-phase filter is used to preserve the timing accuracy of the PPG and ECG signals. (3) Perform z-score normalization on the ECG and PPG signal segments according to the subjects, and then perform min-max normalization on each ECG and PPG signal segment to the [-1,1] interval.

3. The single-channel PPG non-invasive blood pressure monitoring system according to claim 1, characterized in that, ECG generation model is used to generate ECGs from PPGs. E Generator G for generating PPG from ECG P Generator G E A corresponding time-domain discriminant is provided. and frequency domain discriminator Generator G P A corresponding time-domain discriminant is provided. and frequency domain discriminator ; The input to the time-domain discriminator is a time-series signal, while the input to the frequency-domain discriminator is the spectrum obtained by performing a frequency domain transformation on the PPG and ECG signals using short-time Fourier transform.

4. The single-channel PPG non-invasive blood pressure monitoring system according to claim 3, characterized in that, Generator G E and generator G P All models adopt the WaveNet architecture, which includes a starting convolutional module, a core residual module, and an output module. The starting convolutional module consists of a 1×1 pointwise convolutional layer that receives a single-channel one-dimensional temporal input signal and linearly maps it to a 64-channel high-dimensional feature space. The residual module includes several sequentially stacked residual blocks. Each residual block includes a one-dimensional dilated convolutional layer, a non-linear activation layer, and a normalization layer connected in sequence. The dilation factor of the one-dimensional dilated convolutional layer increases exponentially with the model depth. The input and output of each residual block are added element-wise to form residual connections. The output module reprojects the features extracted by the residual module back to the single-channel features through the ReLU activation function and a 1×1 pointwise convolutional layer, ultimately generating the target time-series waveform.

5. The single-channel PPG non-invasive blood pressure monitoring system according to claim 4, characterized in that, definition Given a time series, perform a short-time Fourier transform on the time series. Defined as: ； in Indicates the step size. Indicates frequency parameters, The Hamming window function is represented as follows: , Used to avoid infinite conditions.

6. The single-channel PPG non-invasive blood pressure monitoring system according to claim 3, characterized in that, Time-domain discriminator , The structure includes four downsampling modules and a final classification module. Each downsampling module consists of a one-dimensional convolutional layer with a kernel size of 15 and a stride of 2, an instance normalization layer, and a Leaky-ReLU activation function. The first module does not have a normalization layer. After four downsampling operations, a scalar probability is finally output through the Sigmoid function to achieve a global judgment on the time dependence of the entire waveform. Frequency domain discriminator , The PatchGAN architecture is adopted, which includes four fully convolutional modules. Each fully convolutional module consists of a two-dimensional convolutional layer with a kernel size of 4, a batch normalization layer, and a Leaky-ReLU activation function. The first three fully convolutional modules have a stride of 2 to achieve downsampling, and the first fully convolutional module does not have a normalization layer. The frequency domain discriminator finally outputs a 13×13 two-dimensional confidence map. Each element in the confidence map corresponds to the authenticity score of a local region on the input spectrum.

7. The single-channel PPG non-invasive blood pressure monitoring system according to claim 1, characterized in that, The blood pressure prediction model includes a feature extractor and a mapping regressor. The feature extractor adopts a U-Net-based encoder and decoder architecture, using multi-scale convolutional layers to replace CNN layers to enhance the U-Net structure, and adding a cross-modal attention fusion module to enhance skip connections. The encoder includes a four-layer multi-scale temporal convolutional module. The cross-modal attention fusion module performs weighted calculations on the extracted PPG and ECG signal features, including the following steps: (1) Convert the features output by the multi-scale temporal convolution module into feature tokens through average pooling and depthwise convolution. For each feature token... The query is obtained through depth convolution projection. ; Concatenate the feature tokens of all stages along the channel dimension to generate Then, the key K and value V are obtained through depthwise convolution projection; cross-channel attention (CCA) is obtained by the following formula: ； in For querying, K and V represent global channel information; This represents the total number of channels. For query transpose, for The transpose of the function is used; the output is the weighted channel features; the features weighted by cross-channel attention are obtained through residual connections; (2) Spatial Cross-Attention (SCA) is used to capture the long-distance dependence of multi-scale features in the spatial dimension, and the feature tokens output by CCA are concatenated along the channels to form The query Q and key K are generated through depthwise convolution, while V is obtained from the features output by CCA through depthwise convolution. : ； in , For the number of attention heads; The key K is the transpose; the output is the weighted spatial feature; the spatial attention-weighted feature is obtained through residual connection; (3) The features after cross-modal attention weighting are subjected to deep convolution, layer normalization, and GELU activation, and finally connected to the decoder through upsampling.

8. The single-channel PPG non-invasive blood pressure monitoring system according to claim 7, characterized in that, The decoder of the feature extractor includes four temporal convolutional modules. The input of each temporal convolutional module is a concatenation of the features obtained by skip connections and the upsampled features output by the previous temporal convolutional module. Each temporal convolutional module includes two one-dimensional convolutional layers. The first convolutional layer expands the channel dimension, while the second convolutional layer remains unchanged. The temporal convolutional modules adopt a layer-by-layer channel reduction design.

9. The single-channel PPG non-invasive blood pressure monitoring system according to claim 7, characterized in that, The mapping regressor of the blood pressure prediction model includes a one-dimensional global average pooling layer and a multilayer perceptron head. The output of the feature extractor is first aggregated into a multi-dimensional global feature vector in the temporal dimension through one-dimensional global average pooling, and then input into the multilayer perceptron head for nonlinear transformation, which maps it into a single output node.

Images