Non-invasive blood pressure prediction method of Conformer-LSTM model based on multi-scale feature fusion

Abstract

The invention discloses a noninvasive blood pressure prediction method of a Conformer-LSTM model based on multi-scale feature fusion, and the method comprises the steps: synchronously collecting a PPG signal and an ABP signal of a user, and carrying out the preprocessing of the collected PPG signal; the preprocessed PPG signals and the preprocessed ABP signals are input into a Conformer-LSTM blood pressure prediction model; the input PPG signal is subjected to Conforeemer model extraction and fusion of features of multiple scales, the features are transmitted to the LSTM network, the LSTM network carries out modeling on a time sequence through the extracted features of multiple scales, and finally a predicted ABP waveform is output through a full connection layer; and calculating systolic pressure and diastolic pressure through the predicted ABP waveform. According to the method, the Conformer-LSTM model is adopted, the features of different scales of the PPG signals can be extracted, the features of multiple scales are effectively fused, the key information of the current task can be rapidly extracted from the complex waveform features through the algorithm, the problem that the manually extracted features are unstable due to the diversity of crowds is avoided, and the accuracy of the algorithm is improved. And the applicability of the model to different waveforms is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of blood pressure prediction, and particularly to a non-invasive blood pressure prediction method based on a Conformer-LSTM model with multi-scale feature fusion. Background Art

[0002] Cardiovascular diseases have seriously threatened human life and health. Each year, the number of people who die from cardiovascular diseases globally is as high as 10.4 million. Hypertension is one of the main risk factors for cardiovascular diseases. Therefore, blood pressure monitoring is very necessary for diagnosis and screening, monitoring treatment plans, and indicating the health status of patients. By detecting blood pressure in real time and controlling blood pressure in a timely manner, the risk of cardiovascular complications and death can be greatly reduced.

[0003] Generally, there are two methods for measuring blood pressure: invasive and non-invasive measurements. The invasive method is to invasively insert a cannula into the patient's artery to directly measure blood pressure. It is the gold standard method for continuously and accurately monitoring blood pressure. However, due to its invasiveness, this method is only limited to patients in critical health conditions. The non-invasive method mainly performs non-invasive blood pressure measurement through the Korotkoff sound method and the oscillometric method. However, this method cannot be used for continuous blood pressure monitoring, nor can it provide the arterial blood pressure (ABP) waveform. Moreover, this method will compress the arterial blood vessels and cause discomfort to the patient if measured frequently. Therefore, researchers at home and abroad are committed to finding a blood pressure measurement method that can be continuous, non-invasive, and suitable for daily blood pressure monitoring.

[0004] In recent years, some studies have used photoplethysmography (PPG) for non-invasive blood pressure prediction. Photoplethysmography (PPG) can measure signals related to the time variation of blood volume in blood vessels and tissues. The PPG waveform carries rich information reflecting the cardiovascular function of the subject. Existing studies have proved that its formation is closely related to blood pressure. Moreover, the PPG sensor is small in size and low in manufacturing cost, and can be embedded in wearable devices to achieve daily blood pressure monitoring.

[0005] Many studies have been conducted on PPG signals, using various physiological parameters to continuously and non-invasively estimate blood pressure. Pulse wave velocity (PWV) is one of the important parameters related to blood pressure and arterial stiffness. PWV is the speed at which blood pressure pulses propagate through the circulatory system, and the corresponding PWV is mainly represented by measuring pulse transit time (PTT) and pulse arrival time (PAT). However, measuring PTT or PAT requires the acquisition of two PPG synchronous signals or PPG and electrocardiogram (ECG) synchronous signals, and the blood pressure prediction model established only with PTT or PAT has poor generalization. Some studies have further improved the accuracy of blood pressure prediction by extracting features from PPG signals and using machine learning methods such as linear regression, regression forest, and AdaBoost to establish blood pressure prediction models. However, this method relies on manually extracted features, and the diversity of the population makes the accuracy of feature extraction unstable, which has certain limitations. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a non-invasive blood pressure prediction method based on a Conformer-LSTM model with multi-scale feature fusion, so as to avoid the problem of instability of artificially extracted features due to the diversity of the population and improve the applicability of the model to different waveforms.

[0007] To achieve the above purpose, the technical solution provided by the present invention is:

[0008] A non-invasive blood pressure prediction method based on a Conformer-LSTM model with multi-scale feature fusion, comprising:

[0009] S1. synchronously collect the user's PPG signal and ABP signal, and perform preprocessing operations on the collected PPG signal;

[0010] S2, input the preprocessed PPG signal and ABP signal into the Conformer-LSTM blood pressure prediction model;

[0011] S3, the input PPG signal is extracted and fused with features of multiple scales through the Conforemer model, and the features are passed to the LSTM network; the LSTM network models the time series through the extracted multi-scale features; finally, the predicted ABP waveform is output through the fully connected layer;

[0012] S4. Calculate systolic and diastolic blood pressures using the predicted ABP waveform.

[0013] Furthermore, the preprocessing operation of the collected PPG signal includes:

[0014] S1-1, signal denoising;

[0015] A fourth-order Butterworth band-pass filter with a cut-off frequency of 0.5 - 20 Hz is used to filter out high-frequency noise and baseline drift in the PPG signal. Additionally, a narrow-band low-pass filter with a cut-off frequency of 0.05 Hz is used to extract the information related to the average blood volume contained in the DC component of the PPG signal.

[0016] S1-2, PPG signal segmentation;

[0017] According to the preset sample duration, the denoised PPG signal is segmented to obtain multiple PPG signal samples;

[0018] S1-3, Normalize the amplitudes of the PPG and ABP signals to be between [0, 1].

[0019] Compared with the prior art, the principles and advantages of this solution are as follows:

[0020] Compared with traditional machine learning methods that rely on manually extracted features, this solution uses a Conformer-LSTM model. The two branches of the Conformer model can effectively extract features of multiple scales of the PPG signal. The multi-scale cross-attention module can effectively fuse features of multiple scales, enhance information interaction at different scales of the time series, avoid the problem of unstable features extracted manually due to the diversity of the population, and improve the applicability of the model to different waveforms.

[0021] Compared with other machine learning or deep learning methods that require separately establishing prediction models for systolic blood pressure and diastolic blood pressure, this solution only needs to establish one blood pressure prediction model to achieve continuous blood pressure waveform prediction. Using the PPG signal as the input to predict the blood pressure waveform, blood pressure parameters such as systolic blood pressure and diastolic blood pressure can be calculated from the predicted blood pressure waveform. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the services required for the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is the principle flow chart of a non-invasive blood pressure prediction method based on a Conformer-LSTM model with multi-scale feature fusion of the present invention;

[0024] Figure 2 It is the structural schematic diagram of the Conformer-LSTM blood pressure prediction model of the present invention;

[0025] Figure 3Schematic diagrams of the multi-scale cross-attention module and the MR module in the present invention;

[0026] Figure 4 Schematic diagram of the LSTM network in the present invention;

[0027] Figure 5 Graph for the result analysis of systolic blood pressure and diastolic blood pressure. Specific embodiments

[0028] The present invention will be further described below in conjunction with specific embodiments:

[0029] As Figure 1 shown, a non-invasive blood pressure prediction method based on a Conformer-LSTM model with multi-scale feature fusion described in this embodiment includes:

[0030] S1. Synchronously collect the PPG signal and the ABP signal of the user;

[0031] S2. Perform preprocessing operations on the collected PPG signal. The preprocessing process includes:

[0032] S2-1. Signal denoising;

[0033] Use a fourth-order Butterworth band-pass filter with a cut-off frequency of 0.5 - 20 Hz to filter out high-frequency noise and baseline drift in the PPG signal. Additionally, use a narrow-band low-pass filter with a cut-off frequency of 0.05 Hz to preprocess the PPG signal;

[0034] S2-2. PPG signal segmentation;

[0035] Segment the denoised PPG signal according to the preset sample duration to obtain multiple PPG signal samples;

[0036] S2-3. Normalize the amplitudes of the PPG and ABP signals to between [0, 1], which can make the training process more stable and improve the convergence speed. The normalization operation is as follows:

[0037]

[0038] Where is the normalized signal value, x i is the original signal value, x max and x min are the maximum and minimum values of the data respectively.

[0039] S3. Divide the ABP signal and the preprocessed PPG signal samples into a training set and a test set;

[0040] S4. Construct a Conformer-LSTM blood pressure prediction model based on multi-scale feature fusion, and use the training set data to train the model;

[0041] The PPG signal is a key time series signal that reflects the state of the cardiac activity and the vascular system at any given moment. Real-world time series often exhibit different patterns and fluctuations at different time scales. This variability requires multi-scale modeling for time series prediction to extract temporal features and dependencies at different time intervals.

[0042] Since the cascaded self-attention modules in Transformer can capture the feature dependencies of signals at long distances, but unfortunately they will destroy the local feature details; convolutional operations are good at extracting local features but difficult to capture global representations. The Conformer model is a dual network structure designed to combine CNN-based local features with Transformer-based global features to enhance representation learning.

[0043] Cross-attention is a method to enhance multi-scale feature fusion, which can complement the information differences of features at different scales and enhance the information interaction of features at different scales.

[0044] Since the classical CNN-LSTM model only uses a convolutional neural network to extract the features of the signal in the first stage, convolutional operations are good at extracting the local features of the signal but difficult to capture the global features, while the Conformer network can effectively combine the local features and global features of the signal, so replacing the convolutional neural network that extracts features in the first stage with the Conformer network will be a good research direction.

[0045] To address the above problems, this embodiment improves the Conformer model to make it suitable for the blood pressure prediction task. The specific structure of the Conformer-LSTM blood pressure prediction model provided in this embodiment is as follows Figure 2As shown in (a), first, the preprocessed PPG signal is input into the Conformer model, which consists of two branches, namely the Transformer branch and the convolutional neural network branch. In the Transformer branch, the signal is mapped to a high-dimensional space through the Embedding layer, and the features are extracted by the encoder; the convolutional branch is composed of the SE-MSFE and the self-attention module. Since the cascaded self-attention modules in the Transformer can capture the long-distance feature dependencies of the signal, but unfortunately, they will destroy the local feature details; the convolutional operation is good at extracting local features but difficult to capture the global representation. In order to enable the model to better mine and fuse the information of the local and global modalities of the signal, the features extracted from the two branches are respectively input into two multi-scale cross-attention modules to fuse the features of multiple scales and enhance the information interaction of the time series at different scale features, thereby improving the prediction accuracy of the model. Secondly, since the Transformer branch serially connects multiple encoders, this may cause the model to lose some important information due to the deepening of the number of layers. Therefore, Maxpool-Resblock is designed to avoid overfitting and reduce the learning difficulty of the model by means of skip connections. Finally, the feature information of the five branches is concatenated together, and the adaptive feature fusion module is used to fuse the feature information of the five branches, and the two-layer LSTM layer is used to simulate the features of the time series and output the complete ABP waveform through the linear layer. Next, the principles and functions of each module will be introduced in detail. Encoder:

[0046] Transformer is a model based on the encoder-decoder structure. The encoder consists of N identical sub-layers, and each sub-layer is composed of two modules, namely the multi-head attention (Multi-Head Attention) and the feed-forward network. The data is processed through the residual connection and the layer normalization operation between the two modules. The structural diagram of the encoder is as Figure 2 shown in (c). The attention mechanism is the main component module of the Transformer model. The self-attention mechanism maps the input to the query matrix Q, the key matrix K, and the value matrix V through three linear layers. The self-attention calculation formula can be described as:

[0047]

[0048] where represents the dimension of each row of the Q, K, and V matrices; Q and K T are dot-multiplied to obtain the correlation weight score of the vector, divided by to alleviate the problem of gradient disappearance existing in the softmax function, and then multiplied by V through the softmax function to obtain the attention matrix;

[0049] Multi-head attention projects the input into multiple different subspaces by performing multiple different linear transformations on the Q, K, and V matrices, calculates self-attention separately to obtain multiple output matrices, and concatenates them to get the final output, which can enable the model to associate information from different aspects and improve the model's feature extraction performance;

[0050] The multi-head attention mechanism is as follows:

[0051] MultiHead(Q, K, V) = Concat(head1,..., head h )W o

[0052]

[0053] where, are the linear transformation weight matrices for the i-th head of Q, K, and V respectively, and W o is the linear transformation matrix after concatenation.

[0054] The feed-forward network consists of two fully connected layers, a non-linear activation function, and a Dropout layer, and is described as:

[0055] F(X) = Linear2(Dropout(RELU(Linear1(X))))

[0056] where Linear1 and Linear2 represent two fully connected layers, RELU is the non-linear activation function used, and the Dropout layer randomly discards neurons according to the set ratio.

[0057] A residual connection is used between the attention module and the feed-forward network to alleviate the problem of gradient disappearance and accelerate convergence through layer normalization operation. The process is as follows:

[0058] H′ = LN(MSA(X) + X)

[0059] H = LN(F(H′) + H′)

[0060] where MSA represents the multi-head attention module, LN represents the layer normalization operation, and F represents the feed-forward network.

[0061] SE-MSFE module:

[0062] Single-scale feature extraction can only use a convolutional kernel of a fixed size, which may lead to insufficient feature extraction and inability to extract deep feature information well. The multi-scale feature extraction module actually enables the model to have receptive fields of different scales, thereby mining features of different fineness of the signal. Based on the existing MSFE module, this paper designs the SE-MSFE module, mainly adding the SE module and the ASFF module after the splicing operation. The ASFF module eliminates redundant information between features of different scales through squeezing and dilation operations, thereby improving the prediction accuracy of the model. The channel attention mechanism is a method to improve the model performance by focusing on different channels of the feature map. In a deep learning model, each channel of the feature map can be regarded as a feature detector for capturing different features in the input data. The channel attention mechanism can improve the model's attention to key information by automatically learning the importance of each channel and adjusting the representation of the input data accordingly. Therefore, the SE module can improve the model's performance by explicitly modeling the mutual dependence between convolutional feature channels. The results of the SE-MSFE module designed in this study are as Figure 2 shown in (b).

[0063] The SE-MSFE module consists of three conventional convolutions and one depthwise separable convolution in parallel. The convolutional kernels of the three conventional convolutions are set to the common 3, 5, and 7, and the convolutional kernel size of the depthwise separable convolution is set to 13. This is because large convolutional kernels will increase the computational cost of the model. Therefore, depthwise separable convolution is used instead of conventional convolution in the large convolution branch. The depthwise separable convolution can significantly reduce the number of parameters in the network by performing convolutional operations independently on each channel of the input layer. Subsequently, pointwise convolution is used to adjust the channels to make up for the loss of information exchange between channels caused by the depthwise separable convolution. In addition, after each of the four parallel branches, batch normalization, relu activation function, and max pooling operations are performed. Then, the results of the four branches are spliced together. To retain the original features and make full use of the feature information of each scale, the spliced signal is passed through the SE module and the AFF module.

[0064] Multi-scale cross-attention module:

[0065] To enable the model to better fuse the global and local features of the signal at multiple scales, two multi-scale cross-attention modules (CMA1 and CMA2) are designed to enhance the information interaction between the global and local features. For convenience, the structure of the cross-attention module is shown here taking the global-local as an example, as Figure 3As shown in (a). CMA performs information fusion by calculating the correlation feature matrix between the global-local and local-global subsequences, enabling the model to learn deeper features, improve the prediction accuracy of the model, and accelerate the convergence of the model. Through cross-attention, the correlation between each point in the sequence can be extracted, namely Q and K in the attention mechanism, and the correlation between each point in each segment of the PPG signal is different, which can be used as important information for predicting the ABP waveform.

[0066] In the figure, A is the correlation information of the global feature, and B is the correlation information of the local feature. The feature matrix X A For X B The cross-attention of uses X A The query matrix Q of A Calculate the correlation with X B This operation can strengthen the information interaction between the two feature matrices and better fuse the information between the two features. Define X A For X B The cross-attention output of is:

[0067]

[0068]

[0069] In the formula, And Are respectively the key matrix and value matrix of X B , And Respectively represent X B To K B And V B The linear transformation parameters of. The finally fused output Y A，B Is the stacking of cross-attention and respective attention for the information interaction between global and local features:

[0070] Y A,B = CMA A→B (X A , X B ) + CMA B→A (X B , X A ) + IMHA(X A ) + IMHA(X B )

[0071] Among them, Y A,B Contains strong interaction information of the signal in both global and local features, and is input into the subsequent module for predicting the ABP waveform.

[0072] MR module:

[0073] The residual connection directly passes the input features to the output in the way of shortcut connection, enabling the entire model to only learn the difference between the input and the output. While simplifying the network learning task, it improves the discriminative ability of the network. The learning process is as follows:

[0074] H(x) = F(x) + x

[0075] In the formula, x is the input feature; H(x) represents the fitting target; F(x) represents the residual mapping. Through the shortcut connection, the original input x is directly associated with the output, thus largely avoiding information loss and greatly reducing the learning difficulty.

[0076] For the proposed one-dimensional residual network ResNet, this residual connection mainly includes max-pooling operation, convolution operation, and batch normalization operation. The Maxpool-Resblock proposed in this paper makes the following improvements based on this: Maxpool-Resblock consists of two serial residual blocks, and each residual block sequentially includes convolution, BatchNorm, ReLU activation function, and max-pooling. The max-pooling operation can effectively reduce the feature dimension of the convolution layer output, while reducing network parameters and computational costs, and reducing the overfitting phenomenon. Therefore, Maxpool-Resblock can effectively reduce the learning difficulty of the network and prevent overfitting. The signal output through a linear layer after the two residual blocks can match the dimension requirements of the subsequent splicing operation.

[0077] AFF module:

[0078] The features output by the cross-attention mechanism contain global and local interaction information. To make full use of the feature information output by self-attention and cross-attention, feature fusion is required to be used for the prediction of ABP waveforms. Currently, the commonly used fusion methods include splicing, summation, and global average pooling. However, such fusion methods do not consider the scale relationship between features and it is difficult to obtain the most discriminative information between multiple feature sets. To solve the above problems, an adaptive feature fusion module is used for feature fusion to eliminate the redundant information between different feature sets and improve the prediction accuracy. For the feature matrices output by different modules, first perform feature splicing in the first dimension to obtain the input of the adaptive feature fusion module F i represents the i-th feature matrix, where it is indicated that there are four feature matrices. First, perform a global average pooling operation to obtain the feature vector Z =

[0079] [Z 1 , Z 2 , Z 3 , Z 4, then compress the feature dimension of the feature vector Z by 1 / 4 through a linear transformation, then pass through the ReLU activation function and perform a dilation operation on the feature vector through a linear transformation to restore it to the original dimension, and finally use the Sigmoid activation function to map the eigenvalue to the interval [0,1]. The fused feature result is obtained by multiplying the fusion weight of the learned feature matrix by the original matrix:

[0080] Z cmps = ReLU(FC(Z, W cmps ))

[0081] Z exc = σ(FC(Z cmps , W exc ))

[0082]

[0083] Among them, FC (Fully Connected) represents the fully connected layer, W cmps and W exc represent the linear transformation parameters, ReLU(x) = max(0, x) represents the non-linear activation function, and W cmps is learnable, so the parameters can be dynamically modified during the backpropagation process to reduce the redundant information between each feature set. is the Sigmoid activation function.

[0084] LSTM network:

[0085] After performing feature fusion on the feature information, two layers of LSTM are used to model the features of the time series. The structure of the LSTM cell is as Figure 4 shown. Among them, the input of the first layer of LSTM is the output of the adaptive feature fusion module, and the number of neurons is 64; the input of the second layer of LSTM is the output of the previous layer of LSTM, and the number of neurons is 64. At time interval t, the input of the LSTM cell is y t , and the output is h t . For the LSTM cell, it consists of an input gate g i , a forget gate g f , an output gate g o , a mapping function m t and a memory cell c t .

[0086] Mathematically, each vector in these components of the LSTM cell can be calculated as follows:

[0087] γ = y t ⊕ h t-1

[0088]

[0089] m t = tanh(W m ·γ + b m )

[0090]

[0091] σ(·) and ⊙ are the sigmoid function and element-wise multiplication. W i , W f , W o and W m are weight matrices, b i , b f , b o and b m are bias vectors, and these parameters are learned during the training process. The three gates have their own weights, and then each LSTM cell works like a state machine. Therefore, the LSTM network can handle time series problems well.

[0092] Predictor:

[0093] The predictor of this model is essentially a linear layer, whose main function is to map the features extracted by the previous modules of the model into subsequences of the ABP waveform. Each neuron in the linear layer is connected to all neurons in the previous layer, and it follows the basic method of a feedforward artificial neural network. After obtaining the subsequences of the ABP waveform, the complete ABP waveform can be obtained by concatenating the subsequences.

[0094] S5. Pass the test set data through the trained Conformer-LSTM blood pressure prediction model to predict the ABP waveform;

[0095] S6. Calculate the systolic blood pressure and diastolic blood pressure from the predicted ABP waveform;

[0096] During one cardiac cycle of the ABP waveform, the systolic blood pressure SBP and diastolic blood pressure DBP are the maximum and minimum values of the blood pressure respectively, and the calculation formulas are as follows:

[0097] SBP = Max(ABP)

[0098] DBP = Min(ABP)

[0099] To prove the effectiveness and superiority of the method described in this embodiment, the following experiments are carried out:

[0100] Experimental Results

[0101] Data Source:

[0102] The PPG and ABP signals used are from the UCI_BP dataset. The dataset includes normal, hypotensive, and hypertensive populations. Among them, the PPG signal is obtained by collecting the fingertip PPG signal with a pulse oximeter, and the ABP signal is obtained by means of intra-arterial measurement. The sampling rate of the PPG and ABP signals is 125 Hz.

[0103] Model parameter settings

[0104] The sample duration of the PPG signal is set to 4.992 seconds (624 sampling points). The number of Encoder layers in the encoder is 4, the number of attention heads is 4, the number of network channels in the feed-forward network is 1024, the learning rate is set to 0.001, the MSE loss function is adopted, and the Adam optimizer is used.

[0105] Evaluation metrics

[0106] In this experiment, the mean absolute error (MAE) and standard deviation (STD) are used for evaluation, and their calculation formulas are as follows:

[0107]

[0108] The SBP error (MAE±STD) predicted by the method proposed in this experiment is (3.70±5.62) mmHg, and the DBP error is (2.18±3.72) mmHg.

[0109] Bland-Altman

[0110] The Bland-Altman plot is a common method in medical applications to evaluate the level of consistency between two measurement methods. In this experiment, the systolic and diastolic blood pressures predicted by the Conformer-LSTM model are analyzed for consistency with the blood pressure values obtained by intra-arterial measurement.

[0111] The analysis results of systolic and diastolic blood pressures are as Figure 5 shown. It can be seen that most of the SBP and DBP are included in the 95% confidence interval, indicating that the prediction of the Conformer-LSTM model in the experiment has good consistency with the intra-arterial measurement method.

[0112] BHS standard

[0113] The British Hypertension Society (BHS) provides a grading standard for blood pressure measurement accuracy. Based on the absolute error, three thresholds (5, 10, and 15 mmHg) are set. By calculating the percentages of the predicted values whose absolute errors are lower than each threshold respectively, the measurement accuracy is divided into three grades: A, B, and C. Meeting grade A or B meets the clinical use conditions. As shown in Table 1, the SBP and DBP predicted by the Conformer-LSTM model used in the experiment both reach grade A.

[0114] Table 1 BHS Standard and Experimental Results

[0115]

[0116] AAMI Standard

[0117] The Association for the Advancement of Medical Instrumentation (AAMI) in the United States provides another blood pressure measurement accuracy index, which stipulates that the absolute mean error (MAE) of blood pressure measurement instruments and methods should be less than 5 mmHg, and the standard deviation (STD) should be less than 8 mmHg. It can be found from Table 2 that both the SBP and DBP predicted by the Conformer-LSTM model used in the experiment meet the AAMI standard.

[0118] Table 2 AAMI Standard and Experimental Results

[0119]

[0120] The above-described embodiments are only the preferred embodiments of the present invention, and do not limit the scope of implementation of the present invention. Therefore, any changes made according to the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A non-invasive blood pressure prediction method based on a Conformer-LSTM model with multi-scale feature fusion, characterized in that Including: S1. Synchronously collect the user's PPG signal and ABP signal, and perform preprocessing operations on the collected PPG signal; S2. Input the preprocessed PPG signal and ABP signal into the Conformer-LSTM blood pressure prediction model; S3. The input PPG signal extracts and fuses features of multiple scales through the Conforemer model, transfers the features to the LSTM network. The LSTM network models the time series through the extracted multi-scale features, and finally outputs the predicted ABP waveform through the fully connected layer; S4. Calculate the systolic blood pressure and diastolic blood pressure through the predicted ABP waveform.

2. The non-invasive blood pressure prediction method based on the Conformer-LSTM model with multi-scale feature fusion according to claim 1, wherein, Performing preprocessing operations on the collected PPG signal includes: S2-1. Use a fourth-order Butterworth band-pass filter to filter out high-frequency noise and baseline drift in the PPG signal. Additionally, use a narrow-band low-pass filter to extract the average blood volume-related information contained in the DC component of the PPG signal; S2-2. Segment the denoised PPG signal according to the preset sample duration to obtain multiple PPG signal samples; S2-3. Normalize the amplitudes of the PPG and ABP signals to between [0, 1].

3. An non-invasive blood pressure prediction method based on a Conformer-LSTM model with multi-scale feature fusion according to claim 1, characterized in that, First: S3-1. Input the preprocessed PPG signal into the Conformer model. The model contains two branches, namely the Transformer branch and the convolutional neural network branch; S3-2. In the Transformer branch, the signal is mapped to a high-dimensional space through the Embeding layer, and features are extracted through the encoder. Among them, due to the cascaded self-attention modules in the encoder, it can effectively capture the global features of the PPG signal. The convolutional branch consists of an SE-MSFE and a self-attention module. The SE-MSFE consists of multiple parallel convolutional blocks, and the convolutional kernel sizes of each convolutional block are different, aiming to extract local features of different scales of the PPG signal. Then, fuse the features of multiple scales through the AFF module, and the self-attention module focuses on the important features extracted; S3-3. Input the features extracted by the two branches into two multi-scale cross-attention modules respectively. Complement the information differences of different-scale features through the cross-attention module, fuse the features of multiple scales, and enhance the information interaction of the time series in different-scale features; S3-4. Design a Maxpool-Resblock module to enable the signal to jump to the splicing operation through a skip connection; S3-5. Finally, splice the feature information of the five branches together, use an adaptive feature fusion module to fuse the feature information of the five branches, and simulate the features of the time series through two LSTM layers and output the complete ABP waveform through the linear layer.

4. The non-invasive blood pressure prediction method based on the Conformer-LSTM model with multi-scale feature fusion according to claim 1, characterized in that, In one cardiac cycle of the ABP waveform, the systolic blood pressure SBP and diastolic blood pressure DBP are respectively the highest and lowest values of the blood pressure, and the calculation formulas are as follows: SBP = max(ABP) DBP = min(ABP).

Images