High robustness cuffless continuous blood pressure estimation method based on multi-view joint modeling

Abstract

The application discloses a high-robustness cuffless continuous blood pressure estimation method based on multi-view joint modeling, and comprises the following steps: 1, collecting signals of a subject and corresponding blood pressure labels, and performing pretreatment to obtain blood pressure signal samples; 2, constructing a blood pressure prediction neural network, comprising a channel interaction feature extraction module, a time sequence feature coding module, a multi-scale sparse fusion module and a regression module, and processing the blood pressure signal samples to obtain predicted systolic pressure and diastolic pressure; 3, constructing a double-target adaptive balance loss function to optimize the blood pressure prediction neural network to obtain an optimal blood pressure prediction model. The application utilizes signals, first-order derivatives and second-order derivatives thereof, realizes multi-view information complementation, and utilizes a double-target adaptive balance loss function to relieve regression deviation problems, thereby improving the accuracy and robustness of blood pressure prediction.

Description

Technical Field

[0001] This invention belongs to the field of intelligent healthcare, specifically relating to a highly robust cuffless continuous blood pressure estimation method based on multi-view joint modeling. Background Technology

[0002] Blood pressure is a vital sign reflecting cardiovascular health, and continuous, accurate blood pressure monitoring is crucial for the early prevention and clinical management of cardiovascular diseases such as hypertension. Traditional invasive blood pressure measurement relies on the direct insertion of an arterial catheter into a blood vessel to obtain blood pressure waveforms in real time. While highly accurate, this is an invasive procedure, unsuitable for long-term use, and may cause infection or vascular damage. Therefore, non-invasive, cuffless blood pressure measurement methods have gradually gained widespread attention and become a research hotspot. However, although non-invasive methods offer advantages in convenience and safety, their measurement results are often affected by individual differences, environmental noise, and sensor conditions, and their predictive accuracy still cannot match that of invasive measurements.

[0003] Against this backdrop, the application of machine learning and deep learning in the medical field has significantly improved the accuracy of blood pressure prediction. These methods, by establishing nonlinear mapping relationships, can automatically learn deep features from complex physiological signals, avoiding the limitations of traditional methods that rely excessively on manual feature extraction. Especially driven by deep neural networks, models can capture implicit correlations in the time domain, frequency domain, and even across modalities, thereby improving the reliability of blood pressure estimation. However, deep learning methods inherently rely on large-scale data, therefore, in recent years, multimodal and multidomain features have been widely introduced into blood pressure prediction tasks, such as electrocardiograms (ECG), photoplethysmography (PPG), demographic information, and spectral features of signals.

[0004] While multimodal inputs can enrich the representational capabilities of models, they also bring a series of challenges. On the one hand, multimodal features may have strong redundancy, leading to increased computational overhead during model training and inference. On the other hand, multimodal measurements often require additional acquisition equipment. For example, ECG acquisition usually relies on electrocardiogram monitors, Holter monitors, or multi-lead electrode patches, which are not readily available or suitable for long-term wear outside of clinical settings.

[0005] Therefore, to balance prediction accuracy and ease of practical application, lightweight blood pressure prediction methods based on neural networks for single modalities are gaining increasing attention. In contrast, PPG signal acquisition only requires photoelectric sensors and can be easily achieved through smart bracelets, finger clip devices, etc., making it more suitable for wearable and long-term home monitoring needs.

[0006] However, despite the widespread attention given to PPG-based monomodal blood pressure prediction due to its convenience and wearability, current research still faces the following challenges:

[0007] 1. Limited Single-Modal, Single-Scale Representation: A single modality can provide limited feature information and has relatively insufficient semantic depth, making it difficult to fully reflect the complex physiological mechanisms related to blood pressure. In this case, shallow networks are limited by their representational capabilities and struggle to capture features; while deep networks, although possessing stronger modeling capabilities, are prone to overfitting to noise or learning redundant information under conditions of insufficient feature semantics, thus leading to a decline in the model's generalization performance.

[0008] 2. Feature-enhanced prediction methods have excessive computational overhead: To compensate for the inadequacy of single-modal representation, some studies have introduced graph structure modeling methods such as recursive graphs and Markov random fields to enhance feature interactivity. However, these methods require a large number of matrix operations and graph structure operations during construction and inference, resulting in high computational complexity and making them difficult to adapt to real-time and resource-constrained scenarios.

[0009] 3. Limitations of Multi-Domain and Multi-View Feature Fusion: With the introduction of frequency domain features and one-dimensional and two-dimensional multi-view representations, the model input dimension has been significantly expanded, making efficient fusion of multiple types of features a key issue. However, most existing methods rely on Transformer and its variant architectures to achieve cross-domain and cross-view information interaction. Although they theoretically possess strong modeling capabilities, their parameter scale and computational complexity are quite large, making it difficult to meet the deployment requirements in resource-constrained environments and limiting the practical application value of such methods in long-term wearable blood pressure monitoring.

[0010] 4. Task bias issue not adequately addressed: In blood pressure prediction, the numerical ranges of systolic blood pressure (SBP) and diastolic blood pressure (DBP) differ significantly. This causes the commonly used mean squared error (MSE) loss function to favor the accuracy of SBP during optimization, thus amplifying the prediction error of DBP. If this "task bias" phenomenon is not effectively addressed, it often leads to an imbalance in the predictive performance of the two indicators, reducing the reliability of the model in clinical applications. Summary of the Invention

[0011] The present invention addresses the shortcomings of the existing technology by proposing a highly robust cuffless continuous blood pressure estimation method based on multi-view joint modeling. This method aims to achieve complementary features from multiple perspectives and regression correction, thereby improving the accuracy and robustness of blood pressure prediction methods.

[0012] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0013] The robust cuffless continuous blood pressure estimation method based on multi-view joint modeling of this invention is characterized by the following steps:

[0014] Step 1: Collect blood pressure signals and perform preprocessing to obtain a blood pressure signal sample set. and its systolic blood pressure label and diastolic blood pressure label, among which, Indicates the first A blood pressure signal sample, This represents the number of channels for each blood pressure signal sample, and , This represents the total number of time sampling points for each blood pressure signal sample. Indicates the number of blood pressure signal samples; let express The systolic pressure label makes express The diastolic blood pressure label;

[0015] Step 2: Construct a blood pressure prediction neural network, including: a channel interaction feature extraction module, a temporal feature encoding module, a multi-scale sparse fusion module, and a regression module, and then... The data is processed to obtain a set of predicted blood pressure values. ,in, express Predicted systolic blood pressure express Predicted diastolic blood pressure;

[0016] Step 3: Based on and as well as and Construct a dual-objective adaptive balance loss function This is used to train the blood pressure prediction neural network and minimize... The network parameters are updated to obtain the optimal blood pressure prediction model, which is then used to predict blood pressure. Represents the smoothing factor. This represents the stabilizing factor.

[0017] The robust cuffless continuous blood pressure estimation method based on multi-view joint modeling described in this invention is also characterized in that step 1 includes:

[0018] Step 1.1: Obtain any subject of The raw signal is processed and baseline and sample segmentation is performed to obtain the subject's data. Pre-processed The set of signal samples, denoted as ,in, Indicates the first Preprocessed Signal samples, Indicates the preprocessed The number of channels in the signal sample, and ;

[0019] Step 1.2: Solve separately The first and second derivatives, respectively, yield the... indivual signal samples and the indivual signal samples Thus obtain signal sample set and signal sample set ;

[0020] In terms of channel dimension , , By splicing the data, a set of blood pressure signal samples can be obtained. ;

[0021] Step 1.3: Solve for and respectively Time-aligned The mean peak value and mean trough value of the signal segment, and accordingly used as... systolic pressure label And diastolic pressure label .

[0022] Furthermore, step 2 includes:

[0023] Step 2.1: The channel interaction feature extraction module consists of... It is composed of stacked inter-channel convolutional blocks, and is used for... The process yields a set of channel interaction feature sequences. ,in, express The Each channel interaction feature sequence express The number of features in each channel, express The number of time points in each channel; This represents the total number of channel interaction feature sequences;

[0024] Step 2.2: The time-series feature encoding module utilizes... One gated loop unit To each Each channel interaction feature sequence is processed to obtain an enhanced set of time-series feature sequences. ,in, express The A time-series feature sequence, Indicates the first One gated loop unit The number of hidden layers;

[0025] Step 2.3: Utilize the multi-scale sparse fusion module to... The process is performed to obtain a multi-scale blood pressure feature vector. ,in, The dimension representing the multi-scale blood pressure feature vector;

[0026] Step 2.4: The regression module utilizes two fully connected layers to respectively... Processing is performed to obtain Predicted systolic blood pressure and predicted diastolic blood pressure .

[0027] Furthermore, the channel-interactive convolutional block in step 2.1 is performed according to the following steps:

[0028] Step 2.1.1: When At that time, with For the first Channel interaction feature sequences output by each channel interactive convolutional block ,in, express The number of time points in each channel express The number of features in each channel;

[0029] Step 2.1.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation Enter the first In each channel-interactive convolutional block, thus providing a temporal dimension for... Perform downsampling to obtain the first... A downsampled time series ;

[0030] Step 2.1.3: Using equation (1) to analyze the time dimension Feature extraction is performed to obtain the first Time-domain modeling time series :

[0031] (1)

[0032] In equation (1), Indicates batch normalization, Represents the linear rectified function. Indicates grouping as The number of input features is The number of output features is One-dimensional convolution;

[0033] Step 2.1.4: Use equation (2) to obtain the first... Channel interaction feature sequence :

[0034] (2)

[0035] In equation (2), It is element-wise addition. It is grouped as Both the input and output feature numbers are One-dimensional convolution;

[0036] Step 2.1.5: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation Assign to Then, return to step 2.1.2 and execute sequentially until... until.

[0037] Furthermore, step 2.3 is performed as follows:

[0038] Step 2.3.1: Obtain the feature weight sequence set using equation (3). ,in, Indicates the first A sequence of feature weights, and , express The Middle Feature weights at each time point:

[0039] (3)

[0040] In equation (3), Indicates from arrive A linear mapping on the feature dimension;

[0041] Step 2.3.2: Preset retention coefficient as follows , build Weighted Retention Sequence ,in, express The weighted number of retained, and , Indicates rounding up;

[0042] right After sorting the feature weights in descending order, select the first... Feature weights at each time point As The sparsity threshold, denoted as Thus, a sparse threshold set is obtained. ;

[0043] Step 2.3.3: Construct the first step using equation (4) Sparse weights at each time point Thus, the first A sparse feature weight sequence This leads to the sparse feature weight sequence set. :

[0044] (4)

[0045] Step 2.3.4: For Along the time dimension After normalization, we obtain the first... A normalized sparse weight sequence This yields the normalized sparse feature weight sequence set. ;

[0046] Step 2.3.5: Use equation (5) to obtain the sparse weighted feature vector set. ,in, Indicates the first Sparse weighted feature vectors:

[0047] (5)

[0048] In equation (5), This indicates summation over the time dimension. This indicates element-wise multiplication over time.

[0049] Step 2.3.6: After concatenating the sparse weighted feature vectors within the range, we obtain... .

[0050] Furthermore, the bi-objective adaptive balance loss function in step 3 It is constructed according to the following steps:

[0051] Step 3.1: Use equation (6) to obtain the first... Mean square error of systolic pressure in the next iteration and the Mean square error of diastolic blood pressure in the next iteration :

[0052] (6)

[0053] In equation (6), Indicates the first Predicted systolic blood pressure value under the next iteration Indicates the first Predicted diastolic blood pressure values ​​under the next iteration;

[0054] Step 3.2: Construct the first step using equation (7) The balancing weighting factor between systolic and diastolic blood pressure in the next iteration :

[0055] (7)

[0056] Step 3.3: Construct the first step using equation (8) Sliding exponential weighted balance factor in the next iteration :

[0057] (8)

[0058] In equation (8), Indicates the first Sliding exponential weighted balance factor in the next iteration;

[0059] Step 3.4: Construct the first step using equation (9) Loss value in the next iteration :

[0060] (9).

[0061] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing the highly robust cuffless continuous blood pressure estimation method based on multi-view joint modeling, and the processor is configured to execute the program stored in the memory.

[0062] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of the described robust cuffless continuous blood pressure estimation method based on multi-view joint modeling.

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0064] 1. This invention addresses the limitation of single-modality, single-scale representation by proposing a multi-scale temporal modeling and fusion method. By introducing GRUs to features at different scales to capture multi-granularity dynamic dependencies, and eliminating redundancy and strengthening complementary information during the fusion process, the diversity and robustness of feature representation are improved, effectively enhancing the model's ability to characterize complex physiological mechanisms related to blood pressure, and improving generalization performance and prediction accuracy;

[0065] 2. This invention addresses the problem of excessive computational overhead in feature-enhanced prediction methods by abandoning complex graph structures and utilizing PPG and its first derivative (APG) and second derivative (VPG) for joint modeling. Obtaining the first and second derivatives involves only simple difference operations, resulting in extremely low computational overhead, making it suitable for real-time and resource-constrained scenarios. Specifically, APG reflects changes in blood flow velocity and vascular compliance, while VPG reveals pulse wave acceleration characteristics and vascular elasticity and peripheral resistance. Through the complementary representation of these three components, this method effectively enhances the model's ability to characterize multi-level physiological mechanisms related to blood pressure and improves prediction accuracy while maintaining computational efficiency.

[0066] 3. This invention addresses the limitations of multi-domain, multi-view feature fusion by introducing a lightweight attention mechanism in the multi-scale GRU feature fusion process: feature scoring is achieved through linear mapping, combined with sparse filtering and softmax weight allocation, effectively suppressing noise interference while highlighting key information. Compared to Transformer-based self-attention mechanisms, this method significantly reduces parameter scale and computational complexity, making it more suitable for the resource constraints of edge computing and wearable devices. Furthermore, compared to traditional average pooling strategies, this method achieves a balance between information selectivity and robustness, thereby effectively improving the generalization performance and clinical applicability of blood pressure prediction models.

[0067] 4. This invention addresses the issue of insufficient attention being paid to task bias by proposing a dual-objective adaptive balancing loss function: balancing weights are dynamically calculated based on the mean squared errors of systolic blood pressure (SBP) and diastolic blood pressure (DBP) in each batch, and smoothed using exponential moving averages, thereby adaptively adjusting the optimization ratio of the two during training. This method avoids the bias towards SBP inherent in traditional loss functions and effectively suppresses batch-to-batch fluctuations, ultimately achieving balanced predictions of SBP and DBP, thus improving the model's robustness and reliability for clinical applications. Attached Figure Description

[0068] Figure 1 This is a flowchart of the process framework of the present invention;

[0069] Figure 2 This is a flowchart of the multi-scale sparse fusion module of the present invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. In order to make the objectives, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] This embodiment uses a robust cuffless continuous blood pressure estimation method based on multi-view joint modeling, such as... Figure 1 As shown, the procedure is as follows:

[0072] Step 1: Collect blood pressure signals and perform preprocessing to obtain a blood pressure signal sample set. and its systolic blood pressure label and diastolic blood pressure label, among which, Indicates the first A blood pressure signal sample, This represents the number of channels for each blood pressure signal sample, and , This represents the total number of time sampling points for each blood pressure signal sample. Indicates the number of blood pressure signal samples; let express The systolic pressure label makes express The diastolic blood pressure label; in this embodiment, each signal segment contains 1248 time sampling points, for a total of 415,000 samples.

[0073] Step 1.1: Obtain any subject of The raw signal is processed and baseline and sample segmentation is performed to obtain the subject's data. Pre-processed The set of signal samples, denoted as ,in, Indicates the first Preprocessed Signal samples, Indicates the preprocessed The number of channels in the signal sample, and In this embodiment, a fourth-order Butterworth filter is used for filtering, employing a passband of 0.5Hz to 8Hz. The signal is processed.

[0074] Step 1.2: Solve separately The first and second derivatives are solved using a simple finite difference method in this embodiment; the corresponding first and second derivatives are obtained. indivual signal samples and the indivual signal samples Thus obtain signal sample set and signal sample set ;

[0075] In terms of channel dimension , , By splicing the data, a set of blood pressure signal samples can be obtained. .

[0076] Step 1.3: Solve for and respectively Time-aligned The mean peak value and mean trough value of the signal segment, and accordingly used as... systolic pressure label And diastolic pressure label .

[0077] Step 2: Construct a blood pressure prediction neural network, including: a channel interaction feature extraction module, a temporal feature encoding module, a multi-scale sparse fusion module, and a regression module, and then... The data is processed to obtain a set of predicted blood pressure values. ,in, express Predicted systolic blood pressure express Predicted diastolic blood pressure;

[0078] Step 2.1: The channel interaction feature extraction module consists of... It is composed of stacked inter-channel convolutional blocks, and is used for... The process yields a set of channel interaction feature sequences. ,in, express The Each channel interaction feature sequence express The number of features in each channel, express The number of time points in each channel; This represents the total number of channel interaction feature sequences; in this embodiment, The value is 5. A value of 3 indicates that only the outputs of the last 3 channels of the interactive convolutional block are selected for subsequent processing;

[0079] Step 2.1.1: When At that time, with For the first Channel interaction feature sequences output by each channel interactive convolutional block ,in, express The number of time points in each channel. express The number of features in each channel; in this embodiment, the interactive convolutional block of each channel is downsampled by a factor of two;

[0080] Step 2.1.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation Enter the first In each channel-interactive convolutional block, thus providing a temporal dimension for... Perform downsampling to obtain the first... A downsampled time series ;

[0081] Step 2.1.3: Using equation (1) to analyze the time dimension Feature extraction is performed to obtain the first Time-domain modeling time series :

[0082] (1)

[0083] In equation (1), Indicates batch normalization, Represents the linear rectified function. Indicates grouping as The number of input features is The number of output features is One-dimensional convolution; each channel's interactive convolutional block satisfies: This means doubling the number of channels; the purpose of using grouped convolution is to ensure that different channels are modeled independently and do not interfere with each other, while also processing multi-channel information in parallel to improve the computing speed; the convolution kernel size is 7, the stride is 1, and the padding is 3; the number of features contained in each channel is 16, 32, 48, 72, and 128 respectively.

[0084] Step 2.1.4: Use equation (2) to obtain the first... Channel interaction feature sequence :

[0085] (2)

[0086] In equation (2), It is element-wise addition. It is grouped as Both the input and output feature numbers are One-dimensional convolution; the purpose of this step is to allow interaction between different channels, i.e. , , The three channels interact with each other, and the kernel size is 1.

[0087] Step 2.1.5: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation Assign to Then, return to step 2.1.2 and execute sequentially until... until.

[0088] Step 2.2: The time-series feature encoding module utilizes... One gated loop unit To each Each channel interaction feature sequence is processed to obtain an enhanced set of time-series feature sequences. ,in, express The A time-series feature sequence, Indicates the first One gated loop unit The number of hidden layers; All are single-layer bidirectional, with the number of hidden layers being the same as the number of features contained in each channel of the corresponding channel interactive convolutional block. In this embodiment, the number of features output by the last three channel interactive convolutional blocks from shallow to deep are 144, 216, and 384, respectively, while the corresponding number of hidden layers are 48, 72, and 128.

[0089] Step 2.3: Utilize the multi-scale sparse fusion module to... The process is performed to obtain a multi-scale blood pressure feature vector. ,in, This represents the dimension of the multi-scale blood pressure feature vector; in this embodiment, the workflow of the multi-scale sparse fusion module is as follows: Figure 2 As shown.

[0090] Step 2.3.1: Obtain the feature weight sequence set using equation (3). ,in, Indicates the first A sequence of feature weights, and , express The Middle Feature weights at each time point:

[0091] (3)

[0092] In equation (3), Indicates from arrive A linear mapping on the feature dimension;

[0093] Step 2.3.2: Preset retention coefficient as follows , build Weighted Retention Sequence ,in, express The weighted number of retained, and , This indicates rounding up; in this embodiment, The value is 0.6;

[0094] right After sorting the feature weights in descending order, select the first... Feature weights at each time point As The sparsity threshold, denoted as Thus, a sparse threshold set is obtained. ;

[0095] Step 2.3.3: Construct the first step using equation (4) Sparse weights at each time point Thus, the first A sparse feature weight sequence This leads to the sparse feature weight sequence set. :

[0096] (4)

[0097] Step 2.3.4: For Along the time dimension After normalization, we obtain the first... A normalized sparse weight sequence This yields the normalized sparse feature weight sequence set. ;

[0098] Step 2.3.5: Use equation (5) to obtain the sparse weighted feature sequence set. ,in, Indicates the first A sparse weighted feature sequence:

[0099] (5)

[0100] In equation (5), This indicates summation over the time dimension. This indicates element-wise multiplication over time.

[0101] Step 2.3.6: After concatenating the sparse weighted feature sequences within, we obtain... .

[0102] Step 2.4: The regression module utilizes two single-layer fully connected layers to respectively... Processing is performed to obtain Predicted systolic blood pressure and predicted diastolic blood pressure .

[0103] Step 3: Based on and as well as and Construct a dual-objective adaptive balance loss function This is used to train the blood pressure prediction neural network and minimize... This updates the network parameters to obtain the optimal blood pressure prediction model, which is then used to predict blood pressure. Represents the smoothing factor. This represents the stabilizing factor.

[0104] Step 3.1: Use equation (6) to obtain the first... Mean square error of systolic pressure in the next iteration and the Mean square error of diastolic blood pressure in the next iteration :

[0105] (6)

[0106] In equation (6), Indicates the first Predicted systolic blood pressure value under the next iteration Indicates the first Predicted diastolic blood pressure values ​​under the next iteration;

[0107] Step 3.2: Construct the first step using equation (7) The balancing weighting factor between systolic and diastolic blood pressure in the next iteration :

[0108] (7)

[0109] Step 3.3: Construct the first step using equation (8) Sliding exponential weighted balance factor in the next iteration :

[0110] (8)

[0111] In equation (8), Indicates the first Sliding exponential weighted balance factor in the next iteration;

[0112] Step 3.4: Construct the first step using equation (9) Loss value in the next iteration :

[0113] (9).

[0114] In this embodiment, using Optimizer in Training was performed on the validation set with a learning rate of 0.001, weight decay of 0.05, and 180 batches. Diastolic or systolic blood pressure was retained on the validation set. Minimum model weights.

[0115] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0116] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A robust cuffless continuous blood pressure estimation method based on multi-view joint modeling, characterized in that, The procedure is as follows: Step 1: Collect blood pressure signals and perform preprocessing to obtain a blood pressure signal sample set. and its systolic blood pressure label and diastolic blood pressure label, among which, Indicates the first A blood pressure signal sample, This represents the number of channels for each blood pressure signal sample, and , This represents the total number of time sampling points for each blood pressure signal sample. Indicates the number of blood pressure signal samples; let express The systolic pressure label makes express The diastolic blood pressure label; Step 2: Construct a blood pressure prediction neural network, including: a channel interaction feature extraction module, a temporal feature encoding module, a multi-scale sparse fusion module, and a regression module, and then... The data is processed to obtain a set of predicted blood pressure values. ,in, express Predicted systolic blood pressure express Predicted diastolic blood pressure; Step 2.1: The channel interaction feature extraction module consists of... It is composed of stacked inter-channel convolutional blocks, and is used for... The process yields a set of channel interaction feature sequences. ,in, express The Each channel interaction feature sequence, express The number of features in each channel, express The number of time points in each channel; This represents the total number of channel interaction feature sequences; Step 2.2: The time-series feature encoding module utilizes... One gated loop unit To each Each channel interaction feature sequence is processed to obtain an enhanced set of time-series feature sequences. ,in, express The A time-series feature sequence, Indicates the first One gated loop unit The number of hidden layers; Step 2.3: Utilize the multi-scale sparse fusion module to... The process is performed to obtain a multi-scale blood pressure feature vector. ,in, The dimension representing the multi-scale blood pressure feature vector; Step 2.4: The regression module utilizes two fully connected layers to respectively... Processing is performed to obtain Predicted systolic blood pressure and predicted diastolic blood pressure ; Step 3: Based on and as well as and Construct a dual-objective adaptive balance loss function This is used to train the blood pressure prediction neural network and minimize... The network parameters are updated to obtain the optimal blood pressure prediction model, which is then used to predict blood pressure. Represents the smoothing factor. Indicates the stabilizing factor; Step 3.1: Use equation (6) to obtain the first... Mean square error of systolic pressure in the next iteration and the Mean square error of diastolic blood pressure in the next iteration : (6) In equation (6), Indicates the first Predicted systolic blood pressure value under the next iteration Indicates the first Predicted diastolic blood pressure values ​​under the next iteration; Step 3.2: Construct the first step using equation (7) The balancing weighting factor between systolic and diastolic blood pressure in the next iteration : (7) Step 3.3: Construct the first step using equation (8) Sliding exponential weighted balance factor in the next iteration : (8) In equation (8), Indicates the first Sliding exponential weighted balance factor in the next iteration; Step 3.4: Construct the first step using equation (9) Loss value in the next iteration : (9)。 2. The robust cuffless continuous blood pressure estimation method based on multi-view joint modeling according to claim 1, characterized in that, Step 1 includes: Step 1.1: Obtain any subject of The raw signal is processed and baseline and sample segmentation is performed to obtain the subject's data. Pre-processed The set of signal samples, denoted as ,in, Indicates the first Preprocessed Signal samples, Indicates the preprocessed The number of channels in the signal sample, and ; Step 1.2: Solve separately The first and second derivatives, respectively, yield the... indivual signal samples and the indivual signal samples Thus obtain signal sample set and signal sample set ; In terms of channel dimension , , By splicing the data, a set of blood pressure signal samples can be obtained. ; Step 1.3: Solve for and respectively Time-aligned The mean peak value and mean trough value of the signal segment, and accordingly used as... systolic pressure label And diastolic pressure label .

3. The robust cuffless continuous blood pressure estimation method based on multi-view joint modeling according to claim 2, characterized in that, The channel-interactive convolutional block in step 2.1 is performed as follows: Step 2.1.1: When At that time, with For the first Channel interaction feature sequences output by each channel interactive convolutional block ,in, express The number of time points in each channel express The number of features in each channel; Step 2.1.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Enter the first In each channel-interactive convolutional block, thus providing a temporal dimension for... Perform downsampling to obtain the first... A downsampled time series ; Step 2.1.3: Using equation (1) to analyze the time dimension Feature extraction is performed to obtain the first Time-domain modeling time series : (1) In equation (1), Indicates batch normalization, Represents the linear rectified function. Indicates grouping as The number of input features is The number of output features is One-dimensional convolution; Step 2.1.4: Use equation (2) to obtain the first... Channel interaction feature sequence : (2) In equation (2), It is element-wise addition. It is grouped as Both the input and output feature numbers are One-dimensional convolution; Step 2.1.5: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation Assign to Then, return to step 2.1.2 and execute sequentially until... until.

4. The robust cuffless continuous blood pressure estimation method based on multi-view joint modeling according to claim 3, characterized in that, Step 2.3 shall be performed as follows: Step 2.3.1: Obtain the feature weight sequence set using equation (3). ,in, Indicates the first A sequence of feature weights, and , express The Middle Feature weights at each time point: (3) In equation (3), Indicates from arrive A linear mapping on the feature dimension; Step 2.3.2: Preset retention coefficient as follows , build Weighted Retention Sequence ,in, express The weighted number of retained, and , Indicates rounding up; right After sorting the feature weights in descending order, select the first... Feature weights at each time point As The sparsity threshold, denoted as Thus, a sparse threshold set is obtained. ; Step 2.3.3: Construct the first step using equation (4) Sparse weights at each time point Thus, the first A sparse feature weight sequence This leads to the sparse feature weight sequence set. : (4) Step 2.3.4: For Along the time dimension After normalization, we obtain the first... A normalized sparse weight sequence This yields the normalized sparse feature weight sequence set. ; Step 2.3.5: Use equation (5) to obtain the sparse weighted feature vector set. ,in, Indicates the first Sparse weighted feature vectors: (5) In equation (5), This indicates summation over the time dimension. This indicates element-wise multiplication over time. Step 2.3.6: After concatenating the sparse weighted feature vectors within the range, we obtain... .

5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the highly robust cuffless continuous blood pressure estimation method based on multi-view joint modeling as described in any one of claims 1-4, the processor being configured to execute the program stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the highly robust cuffless continuous blood pressure estimation method based on multi-view joint modeling as described in any one of claims 1-4.

Images