A central arterial pressure reconstruction method based on mamba attention gate recurrent network

Abstract

The application discloses a kind of based on Mamba's attention gate recurrent network central arterial pressure reconstruction method, it is related to cardiovascular health monitoring technical field.The R2A-MAGNet model presented in the application can more accurately reconstruct central arterial pressure (CAP) waveform from radial arterial pressure (RAP) waveform by combining multi-scale CNN, BiGRU, Mamba, self-attention and cross-attention mechanism.The experimental results show that the model realizes the minimum mean absolute error (MAE) of CAP waveform 1.93 mmHg, and the minimum root mean square error (RMSE) of central arterial systolic pressure (CASP) and diastolic pressure (CADP) is 2.77 mmHg and 1.48 mmHg respectively, which is significantly better than the existing model, with stronger robustness and generalization ability: through data enhancement and normalization processing, as well as multi-scale feature extraction and the application of attention mechanism, R2A-MAGNet model shows stronger robustness and generalization ability under different conditions of data of different individuals.

Description

Technical Field

[0001] This invention relates to the field of cardiovascular health monitoring technology, specifically to a method for reconstructing central arterial pressure using an attention-gated circulatory network based on Mamba. Background Technology

[0002] Cardiovascular disease remains a leading cause of morbidity and mortality worldwide, making effective blood pressure monitoring crucial for reducing the incidence and severity of cardiovascular events. Among various blood pressure parameters, central arterial pressure (CAP) is particularly important. CAP, measured at the root of the ascending aorta, is closely related to left ventricular load and provides valuable physiological insights, including systolic and diastolic blood pressure (CASP) and diastolic blood pressure (CADP), enhancement index, and ejection duration. This makes CAP a more reliable predictor of cardiovascular outcomes than peripheral arterial pressure (PAP) measured in the brachial or radial artery.

[0003] Currently, methods for measuring CAP are either invasive or non-invasive. Invasive methods, such as ductus arteriosus insertion, remain the gold standard due to their accuracy and real-time data transmission, but they are limited by high cost, procedural risks, and patient discomfort. Therefore, non-invasive alternatives have been developed. These methods include direct substitution, generalized transfer function (GTF), and N-point moving average (NPMA) techniques, which estimate CAP by transforming the PAP waveform through mathematical models. While safer and easier to use, non-invasive methods often face challenges in terms of accuracy, waveform fidelity, and adaptation to individual physiological differences.

[0004] The limitations of traditional methods, primarily due to the complexity of blood pressure waveforms, have led to the adoption of deep learning techniques for CAP estimation. Deep learning effectively models complex, nonlinear relationships in high-dimensional data. Architectures such as Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, and self-attention mechanisms are particularly well-suited for this task. CNNs extract spatial features from PAP waveforms, identifying subtle patterns without manual feature engineering. BiLSTM networks capture temporal dependencies by processing waveforms in both forward and backward directions. Self-attention mechanisms improve the accuracy of CAP reconstruction by focusing on the most relevant parts of the input sequence. These methods achieve end-to-end learning, directly mapping PAP measurements to CAP waveforms. Further advances include using attention mechanisms to combine photoplethysmography (PPG) and demographic data to better understand mean arterial blood pressure morphology, and combining BiGRU, CNNs, and attention mechanisms to integrate information from BCG, PPG, and ECG to predict central pressure. Hsu et al. also investigated a general PPG-based method, while Xiao et al. advanced CAP waveform reconstruction using CNN-BiLSTM and CBi-SAN models.

[0005] However, traditional methods like GTF and NPMA, as well as earlier deep learning models (such as CNN-BiLSTM), often fail to capture short-term fluctuations and long-term trends. This deficiency not only impairs CAP reconstruction but also stems from insufficient modeling of time dependencies and inadequate focus on key waveform segments. Consequently, these models perform poorly in reconstructing CAP waveforms.

[0006] Therefore, a new solution is needed to address the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method for reconstructing central arterial pressure in attention-gated recurrent networks based on Mamba, so as to solve the technical problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for reconstructing central arterial pressure in an attention-gated recurrent network based on Mamba, comprising at least the following steps:

[0009] S1: Data collection and processing: Obtain a certain amount of CAP waveform data and RAP waveform data from the existing dataset, and perform data preprocessing, data augmentation and data normalization.

[0010] S2: Construct R2A-MAGNet, which consists of a local feature capture stage and a global feature extraction stage;

[0011] S3: The local feature capture stage uses multi-scale CNN for feature extraction, that is, it uses convolution kernels of different sizes to extract the short-term and long-term trends of RAP waveforms, thereby enhancing the model's ability to capture local temporal features.

[0012] S4: The global feature extraction stage sequentially employs stacked BiGRU (bidirectional gated recurrent unit) layers, Mamba layers, self-attention layers, and cross-attention layers to extract features, capturing temporal dependencies and enhancing information interaction.

[0013] S5: The final output is obtained through the output layer of the R2A-MAGNet model.

[0014] Furthermore, in S1, the signal preprocessing uses a high-fidelity pressure measurement system to acquire data. The frequency response of the high-fidelity pressure measurement system is >20Hz, each record lasts for at least 20 seconds, and after sampling at a sampling rate of 200Hz, it is downsampled to 128Hz through linear interpolation.

[0015] The data augmentation method involves using a sliding window approach to divide continuous CAP and RAP waveforms into fixed-length windows, each containing 500 data points with a step size of 50 data points, in order to increase the amount of training data and capture subtle temporal changes in physiological signals.

[0016] The data normalization involves using Z-score normalization to rescale the data points of each segment to a standardized range. The specific formula is as follows:

[0017]

[0018] Where E represents the original data point, x′ is the normalized value, μ is the mean of the segment, and σ is the standard deviation of the segment.

[0019] Furthermore, the feature extraction in the local feature capture stage includes at least the following steps:

[0020] The convolutional layers are used in parallel one-dimensional convolutional layers with kernel sizes of 16 and 25, respectively. Each convolutional layer is followed by a batch normalization layer and a GELU activation function.

[0021] The output connections form a comprehensive feature map, as shown in the following formula:

[0022] conv1=GELU(BN(W1*x+b1))∈R d×l

[0023] conv2=GELU(BN(W2*x+b2))∈R d×l

[0024] f local =Concat(Conv1,Conv2)∈R 2d×l

[0025] Where, x∈R 1×l The input represents the RAP waveform; W1 and W2 ∈ R d×k Denotes a convolution filter; b1 and b2 ∈ R d d represents the bias term; l represents the dimension of the hidden layer; k represents the length of the input waveform; and k represents the kernel size, which is set to 16 and 25 respectively.

[0026] Furthermore, the feature extraction in the global feature extraction stage includes at least the following steps:

[0027] By capturing the temporal dependencies of the input feature map through BiGRU layers, the model can learn long-range dependencies and complex patterns in RAP waveforms. Specifically, a BiGRU block consists of stacked BiGRU layers, processing the sequence in both forward and backward directions. The output at each time step is:

[0028] ht =BiGRU(f local,h ,h t-1 )∈R 2d

[0029] Among them, h t The hidden state at time t is the rich context representation h generated by the BiGRU block. GRU ∈R 2d ×l It encodes the time dynamics of the RAP waveform;

[0030] To further capture complex relationships, a Mamba layer is used. This Mamba block utilizes its selective state space to simulate the interactions between different segments of the input, processing h... GRU To generate h Mamba ∈R 2d×l Selectively focus on the most informative segments while suppressing irrelevant information;

[0031] For h GRU Another perspective is to apply a self-attention layer to capture the relationships between segments, generating h SA ∈R 2d×l This mechanism is represented as:

[0032]

[0033] Q = W Q h GRU K = W K h GRU V = W V h GRU

[0034] Where Q, K, and V are learned weight matrices W Q W K and W V From h GRU Calculated;

[0035] Subsequently, a cross-attention layer was used to enhance the interaction between the feature maps processed by self-attention and those processed by Mamba, i.e., h SA and h Mamba The interaction between the two feature maps is achieved by the cross-attention layer, which determines the attention weights by calculating the similarity between the two feature maps.

[0036]

[0037] Q = W Q h SA K = W K h Mamba V = W Vh Mamba

[0038] Where Q, K, and V are learned matrices from h SA and h Mamba The calculated result shows that the operation of the poor attention layer produces a weighted sum that highlights the most relevant segments in the two feature maps.

[0039] Furthermore, the output layer of the R2A-MAGNet model is designed to extract features from the processed feature map h. CA ∈R 2d×l Reconstruct the CAP waveform;

[0040] The output layer consists of a fully connected layer, which converts the final feature map h... CA Mapped to the reconstructed CAP waveform The output layer is represented as:

[0041]

[0042] Among them, W out ∈R 1×2d This is the weight matrix of the fully connected layer; b out ∈R is the bias term; output This represents the reconstructed CAP waveform.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] 1. Higher Reconstruction Accuracy: The R2A-MAGNet model proposed in this invention, by combining multi-scale CNN, BiGRU, Mamba, self-attention, and cross-attention mechanisms, can more accurately reconstruct the central arterial pressure (CAP) waveform from the radial artery pressure (RAP) waveform. Experimental results show that this model achieves the lowest mean absolute error (MAE) of 1.93 mmHg for the CAP waveform, and the lowest root mean square error (RMSE) of 2.77 mmHg and 1.48 mmHg for the central arterial systolic pressure (CASP) and diastolic pressure (CADP), respectively, which are significantly better than existing models.

[0045] 2. Enhanced robustness and generalization ability: Through data augmentation and normalization, as well as the application of multi-scale feature extraction and attention mechanisms, the R2A-MAGNet model exhibits stronger robustness and generalization ability on data from different individuals and under different conditions. This enables the model to maintain consistent high performance on diverse datasets and adapt to the physiological differences among individuals.

[0046] 3. More Comprehensive Feature Capture: Compared to traditional methods and existing deep learning models, R2A-MAGNet can capture the spatial and temporal features of CAP waveforms more comprehensively. Multi-scale CNNs can extract short-term and long-term spatial information, while BiGRU and Mamba layers can effectively model temporal dependencies and long-distance dependencies. Self-attention and cross-attention mechanisms further enhance feature focusing and information interaction, improving the model's attention to key waveform segments.

[0047] 4. Enhanced computational performance: The introduction of the selective state-space model (Mamba) enables R2A-MAGNet to efficiently handle long-range dependencies while avoiding the computational overhead of traditional recursive units. This not only improves the model's computational efficiency but also reduces its complexity, making it more suitable for applications in real-time monitoring and wearable devices.

[0048] 5. Ability to adapt to complex waveforms: The R2A-MAGNet model performs exceptionally well when handling waveforms with high-frequency components or rapid pressure changes, and can more accurately capture the complex dynamic changes of CAP waveforms. This enables the model to provide more reliable CAP estimates when faced with complex physiological signals, further enhancing its value in practical applications. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a diagram of the R2A-MAGNet model architecture proposed in this invention;

[0051] Figure 2 This is a visualization of the CAP waveform reconstructed by the R2A-MAGNet model of this invention; Figure 2 (a) is a normal CAP, Figure 2 (b) is a high-frequency CAP, Figure 2 (c) represents a low-frequency CAP;

[0052] Figure 3 This invention provides a scatter plot comparing the predicted and actual values ​​of CASP and CADP.

[0053] Figure 4 This invention uses Bland-Altman plots to compare reconstructed data with real data. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0055] The R2A-MAGNet model constructed in this invention achieves efficient reconstruction from RAP waveforms to CAP waveforms by combining CNN, BiGRU, Mamba, self-attention, and cross-attention mechanisms. This model excels in feature extraction and information interaction, significantly improving the accuracy and robustness of CAP waveform reconstruction. Specifically, it forms a Mamba-based attention-gated recurrent network method for central arterial pressure reconstruction, comprising at least the following steps:

[0056] S1: Data collection and processing: Obtain a certain amount of CAP waveform data and RAP waveform data from the existing dataset, and perform data preprocessing, data augmentation and data normalization.

[0057] In this embodiment, the central arterial pressure (CAP) and radial arterial pressure (RAP) waveform data were obtained from a dataset provided by Pauca et al., containing CAP and RAP waveforms from 62 subjects, 58 of whom were hypertensive patients. Each subject underwent measurement once before and once after the vascular interventional procedure, resulting in a total of 124 waveform records. Different datasets may be used for data acquisition in different embodiments.

[0058] S2: Building R2A-MAGNet (see...) Figure 1 R2A-MAGNet consists of a local feature capture stage and a global feature extraction stage.

[0059] S3: In the local feature capture stage, multi-scale CNN is used for feature extraction, that is, convolution kernels of different sizes are used to extract the short-term and long-term trends of RAP waveforms, thereby enhancing the model's ability to capture local temporal features.

[0060] S4: In the global feature extraction stage, stacked BiGRU (bidirectional gated recurrent unit) layers, Mamba layers, self-attention layers, and cross-attention layers are used sequentially to extract features, capture temporal dependencies, and enhance information interaction.

[0061] S5: The final output is obtained through the output layer of the R2A-MAGNet model.

[0062] In S1, signal preprocessing uses a high-fidelity pressure measurement system to acquire data. The frequency response of the high-fidelity pressure measurement system is >20Hz, each record lasts for at least 20 seconds, and after sampling at a sampling rate of 200Hz, it is downsampled to 128Hz through linear interpolation.

[0063] Data augmentation involves using a sliding window method to divide continuous CAP and RAP waveforms into fixed-length windows, each containing 500 data points with a step size of 50 data points, in order to increase the amount of training data and capture subtle temporal changes in physiological signals.

[0064] Data normalization involves using Z-score normalization to rescale the data points in each segment to a standardized range. The specific formula is as follows:

[0065]

[0066] Where E represents the original data point, x′ is the normalized value, μ is the mean of the segment, and σ is the standard deviation of the segment.

[0067] Feature extraction in the local feature capture stage includes at least the following steps:

[0068] The convolutional layers are used in parallel one-dimensional convolutional layers with kernel sizes of 16 and 25, respectively. Each convolutional layer is followed by a batch normalization layer and a GELU activation function.

[0069] The output connections form a comprehensive feature map, as shown in the following formula:

[0070] conv1=GELU(BN(W1*x+b1))∈R d×l

[0071] conv2=GELU(BN(W2*x+b2))∈R d×l

[0072] f local =Concat(Conv1,Conv2)∈R 2d×l

[0073] Where, x∈R 1×l The input represents the RAP waveform; W1 and W2 ∈ R d×k Denotes a convolution filter; b1 and b2 ∈ R d d represents the bias term; l represents the dimension of the hidden layer; k represents the length of the input waveform; and k represents the kernel size, which is set to 16 and 25 respectively.

[0074] Feature extraction in the global feature extraction stage includes at least the following steps:

[0075] By capturing the temporal dependencies of the input feature map through BiGRU layers, the model can learn long-range dependencies and complex patterns in RAP waveforms. Specifically, a BiGRU block consists of stacked BiGRU layers, processing the sequence in both forward and backward directions. The output at each time step is:

[0076] ht =BiGRU(f local,h ,h t-1 )∈R 2d

[0077] Among them, h t The hidden state at time t is the rich context representation h generated by the BiGRU block. GRU ∈R 2d ×l It encodes the time dynamics of the RAP waveform;

[0078] BiGRU: By processing RAP sequences in both forward and backward directions, it captures dynamic relationships and enhances the model's ability to understand and model time series data.

[0079] To further capture complex relationships, a Mamba layer is used. Mamba blocks utilize their selective state spaces to simulate the interactions between different pieces of the input, processing h... GRU To generate h Mamba ∈R 2d×l Selectively focus on the most informative segments while suppressing irrelevant information;

[0080] For h GRU Another perspective is to apply a self-attention layer to capture the relationships between segments, generating h SA ∈R 2d×l This mechanism is represented as:

[0081]

[0082] Q = W Q h GRU K = W K h GRU V = W V h GRU

[0083] Where Q, K, and V are learned weight matrices W Q W K and W V From h GRU Calculated;

[0084] Subsequently, a cross-attention layer was used to enhance the interaction between the feature maps processed by self-attention and those processed by Mamba, i.e., h SA and h Mamba The interaction between the two feature maps is achieved by the cross-attention layer, which determines the attention weights by calculating the similarity between the two feature maps.

[0085]

[0086] Q = WQ h SA K = W K h Mamba V = W V h Mamba

[0087] Where Q, K, and V are learned matrices from h SA and h Mamba The calculated result shows that the operation of the poor attention layer produces a weighted sum that highlights the most relevant segments in the two feature maps.

[0088] The output layer of the R2A-MAGNet model aims to extract from the processed feature map h CA ∈R 2d×l Reconstruct the CAP waveform;

[0089] The output layer consists of a fully connected layer, which converts the final feature map h... CA Mapped to the reconstructed CAP waveform The output layer is represented as:

[0090]

[0091] Among them, W out ∈R 1×2d This is the weight matrix of the fully connected layer; b out ∈R is the bias term; output This represents the reconstructed CAP waveform.

[0092] Experimental results

[0093] 1. Performance Evaluation

[0094] To verify the effectiveness of the proposed R2A-MAGNet model, its performance was compared with several existing deep learning models used for CAP estimation, including LSTM, BiLSTM, LSTM-BiLSTM, CNN-LSTM, CNN-BiLSTM, and CBi-SAN. Table 1 summarizes the performance comparison results. The R2A-MAGNet model achieved the lowest MAECAP (1.93 mmHg), the lowest RMSECASP (2.77 mmHg), and the lowest RMSECADP (1.48 mmHg), outperforming all other models. These results demonstrate the superior accuracy, robustness, and generalization ability of the R2A-MAGNet model in reconstructing CAP waveforms from RAP waveforms.

[0095] Table 1: Performance Comparison of Different Models

[0096]

[0097] The results for LSTM, BiLSTM, LSTM-BiLSTM, and CNN-LSTM come from H. Xiao’s previous work, and these models were implemented on the same dataset.

[0098] 2. Visualization of reconstructed CAP waveforms

[0099] To qualitatively evaluate the performance of the R2A-MAGNet model, this invention visualizes the reconstructed CAP waveforms under three conditions: normal CAP, high-frequency CAP, and low-frequency CAP. These reconstructed waveforms are compared with ground truth values ​​to assess the model's ability to capture complex patterns and dynamics of central arterial pressure. Figure 2 The reconstructed waveform matches the actual ground value very well, proving the accuracy and fidelity of the model.

[0100] By presenting scatter plots of the predicted and actual values ​​of CASP and CADP (e.g.) Figure 3 As shown, Figure 3 The x-axis represents the predicted value, the y-axis represents the true value, and the red dashed diagonal line represents a perfect match; R2A-MAGNet was further evaluated. The red dashed diagonal line represents a perfect match. Among them, CASP and CADP's R... 2 The values ​​were 0.97 and 0.99, respectively, with relative RMSE values ​​of 6.46% and 4.97%, indicating that the model can accurately predict these two indicators, with CADP showing slightly higher accuracy.

[0101] The trend line for CASP is y = 0.97x + 4.23, and the trend line for CADP is y = 1.03x - 2.21, indicating a slight systematic error. The distribution of data points further supports the consistency of the model, especially the low variability of CADP.

[0102] Figure 4 The Bland-Altman plots shown further evaluate the consistency between the reconstructed and true CAP waveforms, where the red dashed line represents the average difference and the green dashed line marks the consistency limit. These plots show the differences against their mean, with the red dashed line representing the average difference and the green dashed line representing the consistency limit (±1.96 SD). The differences are symmetrically distributed around the mean, and most points fall within the consistency limit (<95%), indicating that R2A-MAGNet provides CAP estimates with consistency and reliability, and with minimal bias.

[0103] Data collection

[0104] Central arterial pressure (CAP) and radial arterial pressure (RAP) waveform data were obtained from a dataset provided by Pauca et al., which included 62 subjects (mean age: 61 ± 11 years; 45 males and 17 females), of whom 58 were diagnosed with hypertension.

[0105] Each subject underwent two measurements: one at baseline before the interventional procedure and another after administration of antihypertensive medication during the procedure, resulting in a total of 124 waveform recordings. A high-fidelity pressure measurement system (frequency response >20Hz) was used in one study to ensure accurate acquisition of CAP and RAP waveforms. Each recording lasted at least 20 seconds, capturing at least 10 cardiac cycles. The acquired signals were initially sampled at a rate of 200Hz and then downsampled to 128Hz via linear interpolation to reduce redundancy and preserve key waveform features.

[0106] Evaluation indicators

[0107] The performance of the R2A-MAGNet model in reconstructing CAP waveforms from RAP waveforms is evaluated using three key metrics: the mean absolute error (MAE) of CAP, denoted as MAE. CAP The root mean square error (RMSE) of systolic blood pressure (CASP) and diastolic blood pressure (CADP), denoted as RMSE respectively. CASP and RMSE CADP Specifically, MAE CAP The overall accuracy of the reconstruction is quantified by calculating the mean absolute error between the predicted and actual CAP waveforms. In contrast, RMSE... CASP and RMSE CADP These are used to evaluate the model's accuracy in predicting CASP and CADP values. RMSE assigns more weight to larger errors by squaring each error, then averaging and taking the square root, making it particularly sensitive to outliers and significant biases in the predictions. These metrics are defined as follows:

[0108]

[0109] Where N is the total number of data points, y i and These are the true central arterial pressure (CAP) value and the predicted CAP value at time step i, respectively, while y CASP,i , y CADP,i and These are the actual systolic blood pressure (CASP) value, the predicted CASP value, the actual diastolic blood pressure (CADP) value, and the predicted CADP value at time step i, respectively.

[0110] in conclusion

[0111] This invention proposes a non-invasive central arterial pressure (CAP) waveform reconstruction method based on deep learning. By combining convolutional neural networks (CNN), bidirectional gated recurrent units (BiGRU), a selective state-space model (Mamba), self-attention, and cross-attention mechanisms, the CAP waveform is accurately reconstructed from the radial artery pressure (RAP) waveform. This method aims to address the high cost and procedural risks of existing invasive methods, as well as the shortcomings of non-invasive methods in terms of accuracy, waveform fidelity, and individual adaptability.

[0112] The core of this invention lies in the R2A-MAGNet model architecture, which extracts local spatial features through multi-scale CNNs and utilizes BiGRU to capture temporal dependencies, thereby effectively handling short-term fluctuations and long-term trends. The Mamba layer further captures complex relationships and suppresses irrelevant information through selective state-space modeling. Self-attention and cross-attention mechanisms enhance feature focusing and information interaction, significantly improving the model's accuracy and robustness.

[0113] Extensive evaluation on real-world datasets demonstrates that the R2A-MAGNet model outperforms existing models in CAP waveform reconstruction, achieving the lowest mean absolute error (MAE) and root mean square error (RMSE), reaching 1.93 mmHg, 2.77 mmHg (CASP), and 1.48 mmHg (CADP), respectively. Visualization and statistical analysis further validate the model's accuracy and reliability. Bland-Altman analysis shows good consistency between the reconstructed waveform and the true waveform with minimal deviation.

[0114] In summary:

[0115] This invention integrates multi-scale convolutional neural networks (CNNs), bidirectional gated recurrent units (BiGRUs), selective state-space models (Mamba), self-attention, and cross-attention mechanisms to improve the accuracy and robustness of reconstructing CAP waveforms from radial artery pressure (RAP) waveforms. The process includes data preprocessing, model building, feature extraction, application of attention mechanisms, and an output layer. The data preprocessing stage primarily denoises, segments, and normalizes the original RAP signal to generate feature representations suitable for model input. The model building stage proposes the R2A-MAGNet model, which consists of a local feature capture stage and a global feature extraction stage, achieving efficient feature extraction and adaptive focusing. The feature extraction stage efficiently extracts spatiotemporal features using multi-scale CNNs and BiGRUs, reducing computational complexity. The output layer maps the processed features to the CAP waveform space, achieving accurate reconstruction of the CAP waveform. Real-world dataset tests demonstrate that R2A-MAGNet outperforms existing models, achieving the lowest mean absolute error (MAE) for CAP (1.93 mmHg), the lowest root mean square error (RMSE) for central systolic blood pressure (2.77 mmHg), and the lowest RMS error for central diastolic blood pressure (1.48 mmHg). This invention innovatively addresses the shortcomings of traditional CAP reconstruction models in handling complex signals by combining multi-scale convolution, BiGRU, Mamba, and multi-attention mechanisms, significantly improving reconstruction accuracy and robustness, especially in dynamic and complex blood pressure signal scenarios.

[0116] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A Mamba-based attention-gated recurrent network central arterial pressure reconstruction method, characterized in that: At least the following steps are included: S1: Data collection and processing: Obtain a certain amount of CAP waveform data and RAP waveform data from the existing dataset, and perform data preprocessing, data augmentation and data normalization. S2: Construct R2A-MAGNet, which consists of a local feature capture stage and a global feature extraction stage; S3: The local feature capture stage uses multi-scale CNN for feature extraction, that is, it uses convolution kernels of different sizes to extract the short-term and long-term trends of RAP waveforms, thereby enhancing the model's ability to capture local temporal features. S4: The global feature extraction stage sequentially employs stacked BiGRU layers, Mamba layers, self-attention layers, and cross-attention layers to extract features, capturing temporal dependencies and enhancing information interaction. The feature extraction in the global feature extraction stage includes at least the following steps: By capturing the temporal dependencies of the input feature map through BiGRU layers, the model can learn long-range dependencies and complex patterns in RAP waveforms. Specifically, a BiGRU block consists of stacked BiGRU layers, processing the sequence in both forward and backward directions. The output at each time step is: wherein, is the hidden state at time t, the BiGRU block produces a rich contextual representation that encodes the temporal dynamics of the RAP waveform; To further capture complex relationships, Mamba layers are used. Mamba blocks utilize their selective state spaces to simulate the interactions between different pieces of the input, processing... To generate Selectively focus on the most informative segments while suppressing irrelevant information; For another perspective on , the self-attention layer is applied to capture the inter-segment relationships, resulting in , which is represented as: ， ， wherein , and are computed from , and by means of learnable weight matrices ; Subsequently, a cross-attention layer is used to enhance the interaction between the self-attention processed feature map and the Mamba processed feature map, i.e. and The cross-attention layer determines the attention weights by computing the similarity between the two feature maps: ， ， where, , and are computed from and ; the operation of the poor attention layer produces a weighted sum, emphasizing the most relevant patches in both feature maps; S5: The final output is obtained through the output layer of the R2A-MAGNet model.

2. The central arterial pressure reconstruction method based on Mamba attention-gated recurrent networks according to claim 1, characterized in that: In S1, the signal preprocessing uses a high-fidelity pressure measurement system to acquire data. The high-fidelity pressure measurement system has a frequency response >20Hz, each record lasts for at least 20 seconds, and after sampling at a sampling rate of 200Hz, it is downsampled to 128Hz through linear interpolation. The data augmentation method involves using a sliding window approach to divide continuous CAP and RAP waveforms into fixed-length windows, each containing 500 data points with a step size of 50 data points, in order to increase the amount of training data and capture subtle temporal changes in physiological signals. The data normalization involves using Z-score normalization to rescale the data points of each segment to a standardized range. The specific formula is as follows: wherein, denotes the original data point, is the normalized value, μ is the mean value of the segment, and σ is the standard deviation of the segment.

3. The Mamba-based attention-gated recurrent network central aortic pressure reconstruction method of claim 1, wherein: The feature extraction in the local feature capture stage includes at least the following steps: The convolutional layers are used in parallel one-dimensional convolutional layers with kernel sizes of 16 and 25, respectively. Each convolutional layer is followed by a batch normalization layer and a GELU activation function. The output connections form a comprehensive feature map, as shown in the following formula: in, The input represents the RAP waveform; and Represents a convolution filter; and Indicates the bias term; Indicates the dimension of the hidden layer; Indicates the length of the input waveform; This indicates the kernel size, set to 16 and 25 respectively.

4. The Mamba-based attention-gated recurrent network central aortic pressure reconstruction method of claim 1, wherein: The output layer of the R2A-MAGNet model aims to output a probability map from the processed feature maps reconstructing the CAP waveform; The output layer consists of a fully connected layer that maps the final feature maps to the reconstructed CAP waveform ŷ, and the output layer is represented as: wherein, is a weight matrix of the fully connected layer; is a bias term; output denotes the reconstructed CAP waveform.

Images