Arterial blood pressure wave measuring method, device, equipment and medium

Abstract

The invention relates to the technical field of blood pressure waveform prediction, in particular to an arterial blood pressure wave measurement method, device and equipment and a medium, and the method comprises the steps: obtaining a dynamic pulse image of a target area and a corresponding arterial blood pressure waveform label; performing preprocessing and feature point extraction on the dynamic pulse image, and constructing a multi-dimensional pulse wave time sequence feature data set; and establishing a blood pressure prediction model, and inputting the multi-dimensional pulse wave time sequence characteristics into the trained blood pressure prediction model to output a predicted continuous arterial blood pressure waveform. According to the method, the industrial problem that a single model is difficult to capture the instantaneous characteristics and the long-range trend of the pulse waves at the same time is effectively solved, the arterial blood pressure waveform can be noninvasively generated, and the blood pressure prediction precision is improved.

Description

Technical Field

[0001] This invention belongs to the field of blood pressure waveform prediction technology, specifically relating to an arterial blood pressure wave measurement method, device, equipment, and medium. Background Technology

[0002] Cardiovascular disease has become the leading cause of death and disability worldwide. Arterial blood pressure (ABP), as a core physiological indicator of cardiovascular function, is crucial for the accurate and continuous monitoring of diseases such as hypertension and coronary heart disease, enabling early diagnosis, risk assessment, and treatment intervention. While traditional invasive blood pressure monitoring (such as arterial cannulation) provides high-precision data, its invasive procedures pose infection risks and cannot meet the needs of long-term dynamic monitoring, limiting its application to intensive care settings. Therefore, developing non-invasive, high-precision, wearable non-invasive blood pressure monitoring technologies has become a research hotspot in the field of biomedical engineering.

[0003] Among non-invasive blood pressure monitoring methods, indirect measurement techniques based on pulse wave signals (such as photoplethysmography (PPG) and arterial pulse wave (APG)) have attracted widespread attention due to their ease of operation and non-invasiveness. Their core principle is to utilize the physiological correlation between pulse wave propagation characteristics (such as pulse wave velocity (PWV)) and blood pressure, and to estimate blood pressure through feature extraction and modeling. With the rapid development of artificial intelligence technology, machine learning and deep learning methods have gradually become the mainstream solutions for non-invasive blood pressure monitoring, and have undergone a technological evolution from manual feature dependence to end-to-end modeling.

[0004] Traditional machine learning methods have shown new potential in multimodal feature fusion. For example, a dual-modal camera detection system simultaneously acquires facial PPG and chest cardiac oscillation (SCG) signals using a defocused camera, and uses a random forest regression model to estimate blood pressure. The mean absolute error (MAE) for systolic and diastolic blood pressure reached (3.99±5.26) mmHg and (3.70±4.79) mmHg, respectively, demonstrating the application value of non-contact multimodal signals in blood pressure monitoring. In the field of hypertension risk prediction, the XGBoost model exhibits performance comparable to traditional algorithms. For instance, in a cohort study of adolescents, the model achieved a macroscopic AUROC of 0.92 for predicting blood pressure status one year later, significantly higher than the 0.70 of multivariate logistic regression, without relying on historical blood pressure measurement data. While these methods improve feature adaptability, they still rely on manually designed features such as pulse wave transit time (PTT), making it difficult to capture dynamic physiological patterns in the signals.

[0005] The development of deep learning technology has driven breakthroughs in end-to-end modeling. The Transformer architecture and its powerful capabilities in visual temporal tasks provide a new paradigm for extracting physiological information from non-contact signals (such as radar and video). Attention-based models have shown significant advantages in various physiological signal analyses. For long-period waveform prediction, the combination of the Autoformer model and the ICEEMDAN denoising algorithm requires only a small amount of multi-dimensional pulse wave data to predict blood pressure waveforms for eight pulse types, with a mean square error of less than 0.004, overcoming the efficiency bottleneck of Long Short-Term Memory (LSTM) network models in long-term time-dependent modeling. The introduction of Generative Adversarial Networks (GANs) has broken through the limitations of traditional regression models. Blood pressure waveform estimation methods based on conditional GANs generate continuous arterial blood pressure waveforms through adversarial learning, with the standard deviation error of systolic blood pressure controlled within 6.53 mmHg, meeting the clinical standards of the American Association for the Advancement of Medical Instruments (AAMI).

[0006] The fusion of graph neural networks and causal inference provides a new paradigm for blood pressure monitoring. The CiGNN framework identifies key features related to blood pressure changes through causal graphs and processes time-series signals using a spatiotemporal graph neural network (STGNN). In a multi-center dataset, it achieves an accuracy of 3.77 mmHg for systolic blood pressure MAE and 2.52 mmHg for diastolic blood pressure, outperforming traditional PTT-based methods.

[0007] In summary, while existing research has made some progress, it still has several limitations: for example, although traditional machine learning has expanded multimodal applications, its reliance on handcrafted features limits its generalization ability; secondly, single deep learning models (such as Transformer and GAN) struggle to simultaneously capture both the instantaneous characteristics and long-term trends of pulse waves; and thirdly, most methods rely on a single signal source (such as PPG or radar), failing to fully exploit the complementarity of multidimensional physiological information. Furthermore, while the Mamba series models outperform Transformer in temporal processing efficiency, their global critical modeling capabilities are insufficient, and the fusion of their advantages has not yet been achieved in blood pressure monitoring. Summary of the Invention

[0008] In view of the shortcomings of the prior art, the purpose of this invention is to provide an arterial blood pressure wave prediction method, device, equipment and medium.

[0009] According to one aspect of this application, a method for predicting arterial blood pressure waves is disclosed, the method comprising: Acquire dynamic pulse images of the target region and corresponding arterial blood pressure waveform labels; The dynamic pulse images are preprocessed and feature points are extracted to construct a multi-dimensional pulse wave time-series feature dataset; The multi-dimensional pulse wave temporal features are input into the constructed and trained blood pressure prediction model so that the blood pressure prediction model outputs continuous arterial blood pressure waveforms. The blood pressure prediction model adopts an encoder-decoder architecture, and its encoder core includes a hybrid attention mechanism module. This hybrid attention mechanism module captures the local detail features and long-range dependencies of the pulse wave signal through a parallel local self-attention branch and a global feature extraction branch based on the Mamba2 model, and fuses the outputs of the two branches.

[0010] In some embodiments, acquiring the dynamic pulse image of the target region and the corresponding arterial blood pressure waveform label includes: A binocular stereo vision acquisition system and a high-precision pressure transmitter are used to acquire data from a target area, which is a simulated radial artery on a simulated radial artery platform. A thin-film probe is placed on the target area to sense and visualize the pulsation of the radial artery in the target area. After the binocular stereo vision acquisition system and the high-precision pressure transmitter are triggered synchronously, the binocular stereo vision acquisition system records a sequence of dynamic pulsation images of the target area, and the pressure transmitter synchronously records the radial artery blood pressure waveform in the target area to establish a one-to-one correspondence between each frame of the image and the corresponding blood pressure value at that moment.

[0011] In some embodiments, the preprocessing and feature point extraction of the dynamic pulse image to construct a multi-dimensional pulse wave time-series feature dataset includes: The dynamic pulse image is subjected to median filtering to suppress illumination differences and texture noise; The preprocessed image is sequentially divided into blocks for binarization and color inversion, and abnormal regions are removed based on the area threshold of connected regions. Standardized connectivity analysis is performed on the filtered regions to extract the centroids of the target dimension as feature markers. By combining the calibrated camera parameters, the feature markers in the left and right camera images are matched, and their three-dimensional spatial coordinates are solved by a three-dimensional reconstruction algorithm. Thus, the coordinate changes of each feature marker in time series are constructed as a one-dimensional feature, generating a multi-dimensional pulse wave time series feature dataset of the target dimension.

[0012] In some embodiments, before inputting the multidimensional pulse wave time-series features into the constructed and trained blood pressure prediction model, the method further includes: The pulse wave time series features were processed using a local weighted regression scatter smoothing method, and the synchronously acquired arterial blood pressure waveform labels were subjected to low-pass filtering for noise reduction. At the same time, Butterworth low-pass filters were used to remove baseline drift from both to obtain the initial time series. A normalization method based on the fitting parameters of the training set is adopted to normalize the pulse wave time sequence features and blood pressure wave label amplitudes after the signal purification to the [0,1] interval; and the pulse waves and blood pressure waves of the same type are processed into equal-length sequences by interpolation to establish a one-to-one mapping relationship in order to achieve the standardization of the initial time series. The standardized data of the sequence is organized into feature-label pairs, with the first 16 columns of pulse wave features used as model input features and the last column of blood pressure values ​​used as target labels. Continuous input-output sample pairs are generated using the sliding window method, and the total dataset is divided into training set (70%), validation set (10%) and test set (20%) in chronological order.

[0013] In some embodiments, the multidimensional pulse wave temporal features are input into a constructed and trained blood pressure prediction model so that the blood pressure prediction model outputs continuous arterial blood pressure waveforms, including: The multi-dimensional pulse wave temporal features are input into the constructed and trained blood pressure prediction model, so that the blood pressure prediction model sequentially receives the temporal feature matrix processed by the sliding window, performs multi-scale extraction and adaptive filtering on the temporal feature matrix, maps the preprocessed features to a high-dimensional space through linear transformation, extracts and fuses the deep features of the pulse wave signal, generates a prediction sequence based on the encoder output and its own contextual information, and maps the high-dimensional features output by the decoder to the corresponding time step of the arterial blood pressure prediction value output.

[0014] In some embodiments, the local self-attention branch is configured to employ a multi-head self-attention mechanism, where the calculation range of its attention weights is limited to 10 consecutive time steps, for capturing local dynamic features of the pulse wave signal. The local self-attention branch maps the input features to a query matrix Q, a key matrix K, and a value matrix V, respectively, and calculates the output according to the attention weight formula. ; Where Q is the query matrix obtained from the input feature mapping, K is the key matrix obtained from the input feature mapping, and V is the value matrix obtained from the input feature mapping. It is the query vector at the i-th time step in the query matrix. It is the key vector at the j-th time step of the key matrix. It is its transpose; It is the dimension of the key matrix K. It is the intermediate value of the similarity calculation between the query vector at time step i and the key vector at time step j; Yes The attention weights obtained after Softmax normalization are used to measure the importance of the features at time step j to time step i. It is the value vector of the value matrix at the j-th time step; It is the attention output feature obtained by weighting and aggregating the value vector by attention weights.

[0015] In some embodiments, the method further includes training a blood pressure prediction model, including: The preprocessed pulse wave time series feature dataset is divided into training set, validation set and test set in chronological order, and continuous input-output sample pairs are generated using the sliding window method. A batch data loader is used to load the training set, validation set and test set in batches of 16. The training set data is randomly shuffled at the beginning of each training cycle, while the validation set and test set data maintain their original temporal order. Set the loss function to mean squared error, the optimizer to Adam optimizer, and the learning rate to 0.001; During the training phase, the model calculates the predicted values ​​through forward propagation, calculates the error through the loss function, and updates the model parameters through backpropagation. During the validation phase, the model is switched to evaluation mode, gradient calculation is stopped, and its generalization performance is evaluated. Throughout the training process, the model parameters with the minimum loss on the validation set are saved as the parameters of the blood pressure prediction model.

[0016] According to another aspect of this application, an arterial blood pressure wave prediction device is also disclosed, the device comprising: The information acquisition module is used to acquire dynamic pulse images of the target area and the corresponding arterial blood pressure waveform labels; The dataset construction module is used to preprocess and extract feature points from the dynamic pulse images to construct a multi-dimensional pulse wave time-series feature dataset. The arterial blood pressure waveform prediction module is used to input the multi-dimensional pulse wave temporal features into the constructed and trained blood pressure prediction model so that the blood pressure prediction model outputs a continuous arterial blood pressure waveform. The blood pressure prediction model adopts an encoder-decoder architecture, and its encoder core includes a hybrid attention mechanism module. This hybrid attention mechanism module captures the local detail features and long-range dependencies of the pulse wave signal through a parallel local self-attention branch and a global feature extraction branch based on the Mamba2 model, and fuses the outputs of the two branches.

[0017] According to another aspect of this application, an electronic device is also disclosed, the electronic device including a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform various steps of the arterial blood pressure wave prediction method as described in any of the preceding claims.

[0018] According to another aspect of this application, a computer-readable storage medium is also disclosed, wherein instructions are stored on the computer-readable storage medium, characterized in that, when executed by a processor, the instructions implement the various steps of the arterial blood pressure wave prediction method as described in any of the preceding claims.

[0019] The present invention includes, but is not limited to, the following beneficial effects: (1) By synchronously associating dynamic pulse images with arterial blood pressure waveforms, this scheme provides richer spatiotemporal information in the image sequence provided by binocular vision compared to traditional methods that only use one-dimensional signals such as photoelectric volume recording, providing a data foundation for the model to learn a more accurate blood pressure mapping relationship; (2) Through preprocessing and feature point extraction, this scheme constructs a multi-dimensional pulse wave temporal feature dataset from the original image. This step essentially quantifies the complex visual pulsation pattern into a series of calculable and analyzable spatial motion trajectory temporal signals. This feature construction method can more comprehensively and essentially reflect the pulsation state of the artery than relying on a single feature designed manually; (3) This scheme extracts the centroid as a feature point through standardized connected region analysis, so that the feature is no longer limited to the apparent motion of the image plane, but directly reflects the real three-dimensional spatial motion trajectory of the arterial wall or the marked points above it. This includes depth The features of information (z-axis) are more fundamental to reflecting the mechanical changes of arterial pulsation than two-dimensional features, and are more directly related to the physical mechanism of blood pressure formation; (4) The pulse wave and blood pressure wave of the same pulse pattern are processed into equal-length sequences by interpolation and a one-to-one mapping relationship is established, which solves the problem of feature-label misalignment caused by slight asynchronous or different lengths of the original collected data, and ensures that the mapping learned by the model at each time step is accurate; (5) The data is organized into feature-label pairs, and the first 16 columns of the dataset in .CSV format are clearly organized as pulse time sequence features and the last column is blood pressure label, forming a standard supervised learning format that the model can directly recognize, improving the recognition accuracy, and further using the sliding window method to generate continuous sample pairs, which not only makes full use of the data, but more importantly, transforms long time series data into short sequence samples suitable for model processing, while preserving the continuity information of the data, and improving the efficiency and effect of model training. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0021] Figure 1 This is a flowchart of the arterial blood pressure wave prediction method according to an embodiment of this application; Figure 2 This is a simulated radial artery intravascular pressure waveform diagram, i.e., a simulated blood pressure waveform diagram, according to an embodiment of this application. Figure 3 This is a simulated radial artery pulse wave extraction process according to an embodiment of this application; Figure 4 This is an overall architecture diagram of the blood pressure prediction model according to an embodiment of this application; Figure 5 This is the Mamba-2 network architecture of an embodiment of this application; Figure 6 This is a comparison chart of the blood pressure waveform prediction results and the actual values ​​in an embodiment of this application; Figure 7 This is a performance evaluation chart of blood pressure prediction under different pulse patterns according to an embodiment of this application; Figure 8 This is a comparison chart of R² performance in ablation experiments based on thermograms, according to embodiments of this application. Figure 9 These are evaluation index diagrams of different algorithm models in this application embodiment on pulse promotion; Figure 10 This is a structural block diagram of the arterial blood pressure wave prediction device according to an embodiment of this application; Figure 11 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] For ease of understanding, the specific process of the embodiments of the present invention will be described below. Figure 1 This is a flowchart of the arterial blood pressure wave prediction method according to an embodiment of this application. (See attached diagram.) Figure 1 It includes the following steps: S100: Acquire dynamic pulse images of the target area and corresponding arterial blood pressure waveform labels.

[0024] For example, a binocular stereo vision acquisition system and a high-precision pressure transmitter can be combined to acquire data on a target area. The target area is a simulated radial artery on a simulated radial artery platform. A thin-film probe is placed on the target area to sense and visualize the pulsation of the radial artery in the target area. After the binocular stereo vision acquisition system and the high-precision pressure transmitter are triggered synchronously, the binocular stereo vision acquisition system records a sequence of dynamic pulsation images of the target area, and the pressure transmitter synchronously records the radial artery blood pressure waveform in the target area to establish a one-to-one correspondence between each frame of image and the blood pressure value at the corresponding moment.

[0025] Specifically, in this example, a binocular stereo vision pulse acquisition system was used to construct the experimental platform. This technology is based on the imaging principles of the human eye, simultaneously acquiring target images through two cameras at fixed positions and calculating the target distance using parallax. Its core advantage lies in the controllability of hardware parameters, ensuring high accuracy in data reconstruction and laying a solid foundation for providing abundant radial artery pulsation signals.

[0026] The data acquisition process mainly includes dynamic pulse image acquisition, camera calibration, and feature point spatial coordinate reconstruction. Among them, the camera calibration adopts the Zhang Zhengyou calibration method to quickly obtain the camera's intrinsic parameters (such as focal length and principal point), extrinsic parameters (such as rotation matrix and translation vector), and distortion parameters, providing a reliable basis for subsequent 3D reconstruction.

[0027] The experimental data came from a radial artery simulation platform, which can simulate various clinically representative radial artery pressure waveforms, including rapid pulse, fast pulse, slippery pulse, normal pulse, soft pulse, and choppy pulse.

[0028] In the experiment, a binocular vision system was used to acquire images of the simulated radial artery pulsation, which contained three-dimensional spatial information of the arterial pulsation. Simultaneously, a high-precision pressure transmitter was used to synchronously acquire pressure waveforms within the simulated radial artery, such as... Figure 2 As shown, these serve as labels for arterial blood pressure values. Ultimately, in this example, the pulse waveforms of the six typical pulse types and their corresponding blood pressure waveforms were successfully obtained, forming the dataset for subsequent model training and validation.

[0029] S102. Preprocess and extract feature points from dynamic pulse images to construct a multi-dimensional pulse wave time-series feature dataset.

[0030] In some embodiments, the extraction process may include the following steps: Step 1: Perform median filtering on the dynamic pulse image to suppress illumination differences and texture noise.

[0031] Step 2: Perform block binarization and color inversion on the preprocessed image in sequence, and remove abnormal regions based on the area threshold of connected regions.

[0032] Step 3: Perform standardized connectivity analysis on the filtered regions and extract the centroids of the target dimension as feature markers.

[0033] Step 4: Combine the calibrated camera parameters to match the feature markers in the left and right camera images, and solve their three-dimensional spatial coordinates through a three-dimensional reconstruction algorithm. This will construct the coordinate changes of each feature marker in time as a one-dimensional feature, generating a multi-dimensional pulse wave time-series feature dataset of the target dimension.

[0034] Specifically, during the experiment, the thin-film probe was placed at the simulated radial artery. After powering on, the acquisition software was started, and the dual-camera field of view, working distance, and probe pressure were adjusted to the optimal state to acquire clear dynamic pulse images. At the same time, a high-precision pressure transmitter was used to synchronously acquire simulated radial artery blood pressure data to ensure the integrity of the analysis data.

[0035] The processing flow of binocular visual pulse images is as follows: Figure 3 As shown: First, median filtering is used to suppress illumination differences and texture noise; then, block binarization, color inversion conversion, and area thresholding are used to remove abnormal regions; next, 16 typical centroids are extracted as feature markers through standardized connected component analysis; finally, combined with camera calibration parameters, feature matching and 3D reconstruction algorithms are used to solve the spatial location of feature points and construct a 16-dimensional pulse waveform feature dataset.

[0036] S104. Establish a blood pressure prediction model and input the multi-dimensional pulse wave time sequence features into the trained blood pressure prediction model so that the blood pressure prediction model outputs continuous arterial blood pressure waveforms.

[0037] Specifically, the blood pressure prediction model adopts an encoder-decoder architecture. Its encoder core contains a hybrid attention mechanism module, which captures the local details and long-range dependencies of the pulse wave signal through a parallel local self-attention branch and a global feature extraction branch based on the Mamba2 model, and then fuses the outputs of the two branches.

[0038] Understandably, before inputting the multi-dimensional pulse wave time-series features into the blood pressure prediction model, the following steps can be taken to eliminate noise such as baseline drift and spikes caused by platform jitter: 1) The pulse wave time series characteristics were processed using the local weighted regression scatter smoothing method, and the synchronously acquired arterial blood pressure waveform labels were subjected to low-pass filtering for noise reduction. At the same time, the Butterworth low-pass filter was used to remove baseline drift from both to obtain the initial time series.

[0039] 2) A normalization method based on the fitting parameters of the training set is adopted to normalize the pulse wave time series features and the amplitude of the blood pressure wave label after signal purification to the [0,1] interval; and the pulse waves and blood pressure waves of the same type are processed into equal-length sequences by interpolation to establish a one-to-one mapping relationship, so as to achieve the standardization of the initial time series.

[0040] 3) The standardized data were organized into feature-label pairs, with the first 16 columns of pulse wave time-series features used as model input features and the last column of blood pressure values ​​used as target labels. A sliding window method was used to generate continuous input-output sample pairs, and the total dataset was divided into training set, validation set and test set in chronological order.

[0041] That is, in order to eliminate noise such as baseline drift and spikes caused by platform jitter, the signal can be purified first: the pulse wave is processed by local weighted regression scatter smoothing method, combined with low-pass filtering to reduce blood pressure wave noise, and the baseline drift of both is removed by Butterworth low-pass filter to obtain a stable time series.

[0042] Then, sequence standardization was performed: MinMaxScaler was used to normalize the amplitude of the pulse wave and blood pressure wave (based only on the parameters fitted to the training set, mapped to the [0,1] interval, the formula is shown in equation (1), where h is the amplitude of the pulse wave or blood pressure wave after normalization by MinMaxScaler (mapped to the [0,1] interval). The amplitude of the original pulse wave or blood pressure wave; The minimum amplitude of the pulse wave or blood pressure wave obtained by fitting based on the training set; The maximum amplitude of the pulse wave or blood pressure wave is obtained by fitting the training set; the length of pulse waves of different periods is unified by interpolation, and the pulse waves and blood pressure waves of the same type are interpolated into equal-length sequences to establish a one-to-one correspondence.

[0043] ; The experimental data was extracted and organized into a 17-column XLS text file. The first 16 columns were selected as pulse wave time series features, and the last column was selected as the target arterial blood pressure value. Input-output pairs were generated using the sliding window method and converted into PyTorch Tensor format.

[0044] Finally, the dataset was split chronologically to avoid leakage: training set (70%, fitting parameters), validation set (10%, tuning hyperparameters), and test set (20%, evaluating performance).

[0045] Furthermore, this model uses 16-dimensional pulse wave temporal features as input and achieves end-to-end prediction of blood pressure values ​​through an encoder-decoder architecture. The overall architecture, core module design, and training strategy of the model are described in detail below. The core advantage of this example is that it takes into account both the local subtle features of the pulse wave (such as peak shape and dicrotic wave) and the global temporal trend, while enhancing the robustness to noise and individual differences.

[0046] (I) Overall Model Architecture The model employs a classic encoder-decoder architecture, with the overall process following a data flow of "input preprocessing → global feature extraction → feature embedding → encoder-decoder feature transformation → output prediction." The input is a temporal feature matrix processed by a sliding window, with each sample containing 16-dimensional pulse wave features across several consecutive time steps. These features originate from 16 spatial markers extracted by a binocular vision system. The output is the predicted arterial blood pressure value for the corresponding time step, achieving a mapping from multi-dimensional pulse signals to single-dimensional blood pressure values. Specifically, this includes... Figure 4 The steps shown.

[0047] 1. Mamba2 Feature Preprocessing: Multi-scale extraction and adaptive filtering are performed on the original 16-dimensional pulse wave features. This module extracts high-frequency details and low-frequency trends from the signal in parallel using convolutional kernels of different scales, and uses a gating mechanism to dynamically enhance key time segments related to blood pressure changes, suppressing noise interference, thereby improving the representation quality of the input features.

[0048] 2. Input Embedding Layer: This layer maps the preprocessed features to a high-dimensional space through a linear transformation, providing a richer feature base for the subsequent attention mechanism. The embedding dimension is kept consistent with the input feature dimension to achieve a balance between computational efficiency and feature expressive power. 3. Transformer encoder: Consists of multiple stacked coding blocks, each with a hybrid attention mechanism at its core. This mechanism integrates local self-attention with global state-space modeling capabilities, enabling it to simultaneously capture the instantaneous morphological changes and long-term physiological dependencies of pulse wave signals.

[0049] 4. Transformer Decoder: Also composed of 3 stacked decoding blocks, each containing a self-attention, cross-attention, and feedforward network. The decoder focuses on the internal correlations of historical features through self-attention and fuses the global information output by the encoder through cross-attention, gradually establishing a mapping relationship from historical pulse features to the current blood pressure value.

[0050] 5. Output layer: The high-dimensional features output by the decoder are mapped to the final blood pressure prediction value through linear transformation.

[0051] The key parameters of the proposed model have been systematically designed to achieve a balance between computational efficiency, model expressive power, and generalization performance. The embedding dimension is set to match the input pulse wave feature dimension, ensuring sufficient preservation of feature information while controlling computational complexity. The attention mechanism employs a multi-head design, using four parallel attention heads to enable the model to flexibly capture temporal dependencies in different subspaces. The encoder and decoder each consist of three stacked modules, achieving progressively deeper feature extraction through a multi-layered structure, avoiding overfitting risks caused by excessive network depth. The hidden layer dimension of the feedforward network is consistent with the embedding dimension, providing appropriate network capacity for feature transformation. A low dropout ratio is introduced during training to suppress overfitting, while sufficient training epochs are set to ensure full model convergence, and an early stopping strategy further guarantees generalization performance. The overall parameter configuration aims to support the model in efficiently and stably learning blood pressure mapping patterns from pulse wave temporal signals.

[0052] (II) Module Design 1. Mamba2 Feature Preprocessing Module To enhance the model's robustness to multi-scale noise in pulse wave signals (such as high-frequency noise from plateau jitter and low-frequency interference from baseline drift) and individual variability, a Mamba2 feature preprocessing module is introduced before input embedding, such as... Figure 5 As shown, it includes two innovative sub-modules: multi-scale feature extraction and adaptive computation path.

[0053] Specifically, this includes: Multi-scale feature extraction submodule: Addressing the physiological significance of different frequency components in pulse wave signals (e.g., high-frequency components reflect vascular elasticity, low-frequency components reflect hemodynamic trends), it employs parallel feature extraction using grouped convolutions with three different kernels (kernel_size=3, 5, 7). (1) The input features are first adapted to the convolution input format by dimensionality transformation ([batch_size, 10, 16]→[batch_size, 16, 10]); (2) Three sets of convolution kernels perform convolution operations on the features respectively (nn.Conv1d(16, 16, kernel_size=k,padding=k / / 2, groups=16)), where group convolution ensures that each feature dimension is extracted independently and avoids cross-dimensional interference; (3) The convolution output is converted back to the temporal format ([batch_size, 10, 16]) and then concatenated into [batch_size, 10, 48] (3 groups × 16 dimensions). After layer normalization (nn.LayerNorm(48)) and linear transformation (nn.Linear(48,16)), it is fused into 16-dimensional features.

[0054] This submodule effectively separates useful features from noise in pulse waves by covering signal components of different frequencies through multi-scale convolution.

[0055] Adaptive computation path submodule: To enable the model to dynamically focus on key time segments related to blood pressure (such as the main pulse wave and dicrotic wave), a feature selection path based on a gating mechanism is designed. (1) The multi-scale fused features are fed into the gate control network (nn.Linear(16, 8)→ReLU→nn.Linear(8, 1)→Sigmoid) to generate a gate control vector with the same shape as the input (gate_value∈[0,1]). (2) The input features are multiplied element-wise with the gate vector (x * gate_value) to enhance important features and suppress minor features (such as noise). For example, at the moment of sudden change in blood pressure, the gate vector will automatically increase the weight of the corresponding time step, making the model pay more attention to the feature changes of that segment.

[0056] The final output of the preprocessing module is [batch_size, 10, 16], which is fed into the encoder after passing through the embedding layer (nn.Linear(16,embed_dim)) to lay a high-quality foundation for subsequent feature extraction.

[0057] 2. Transformer Encoder and Decoder The encoder and decoder are the core framework for the model to realize temporal feature transformation. Its design draws on the classic structure of Transformer, but by replacing the self-attention mechanism module with a hybrid attention mechanism, the adaptability to pulse wave signals is enhanced.

[0058] The encoder consists of three identical stacked coding blocks, each with a core of a "hybrid attention mechanism layer + feedforward network" (e.g., ...). Figure 4 (As shown). The specific process is as follows: The input features are first processed by a hybrid attention mechanism to generate enhanced features that integrate local details and global trends; the enhanced features and the original input are fused through residual connections and then stably distributed by layer normalization (nn.LayerNorm(embed_dim)); then fed into a feedforward network (two-layer linear transformation + ReLU activation) to achieve nonlinear feature transformation, and the output is optimized again by residual connections and normalization. The three coding blocks deepen the capture of pulse wave temporal correlation layer by layer, and the final output dimension is maintained at [batch_size, 10, 16].

[0059] The decoder consists of three stacked decoding blocks, each containing a self-attention layer + a cross-attention layer + a feedforward network (e.g., ...). Figure 4(As shown). The self-attention layer focuses on the internal temporal correlation of the decoder input, capturing the local dependence between the prediction target and historical features; the cross-attention layer associates the self-attention output with the encoder output, enabling the decoder to utilize both local features and global trends; the feedforward network structure is consistent with the encoder, further optimizing feature representation. The decoder gradually focuses on the mapping pattern of "historical features → blood pressure value" through multi-layer stacking, and the final output shape is [batch_size, 10, 16].

[0060] The synergistic effect of the encoder and decoder enables end-to-end conversion from raw pulse wave features to blood pressure predictions, preserving the dynamic characteristics of time-series data while enhancing the capture of key physiological correlations.

[0061] 3. Hybrid Attention Mechanism Module The hybrid attention mechanism is the core innovation of the encoder, aiming to address the shortcomings of traditional self-attention mechanisms in capturing subtle local features and the weakness of single temporal models in modeling global correlations, thereby achieving collaborative learning of "local details and global trends." Its structure includes two parallel branches and a feature fusion stage: The local self-attention mechanism branch employs a multi-head self-attention mechanism to capture local dynamics (such as the peak-trough interval and amplitude change rate of adjacent pulse waves) within 10 time steps by calculating the association weights between each time step and other time steps in the input features. Specifically, this branch maps the input features into query (Q), key (K), and value (V) matrices, respectively, using attention weight formulas (2) and (3): ; Calculate the weighted features at each time step. Here, Q is the query matrix obtained from the input feature mapping, used as the "query" part to initiate attention calculation; K is the key matrix obtained from the input feature mapping, used to match and calculate the correlation with the query matrix; and V is the value matrix obtained from the input feature mapping, which is the final feature part used for weighted aggregation. It is the query vector at the i-th time step in the query matrix; It is the key vector at the j-th time step of the key matrix. It is its transpose; It is the dimension of the key matrix K, used to scale the dot product result to avoid the value being too large; It is the intermediate value of the similarity calculation between the query vector at time step i and the key vector at time step j; Yes The attention weights obtained after Softmax normalization are used to measure the importance of the features at time step j to time step i. It is the value vector of the j-th time step of the value matrix; Attention(Q,K,V) is the attention output feature obtained by weighting and aggregating the value vectors through attention weights. The output is normalized by dropout(nn.Dropout(dropout_rate)) and layers to preserve the fine features of local temporal sequence.

[0062] The global feature extraction branch incorporates the Mamba2 model (configured as Mamba2Config(d_model=16,n_layer=6,d_state=64,d_conv=4,chunk_size=10)). Its core advantage lies in employing a Structured State-Space Duality (SSD) framework, a key upgrade to the traditional State-Space Model (SSM), particularly well-suited for highly dynamic and noisy physiological signals such as pulse waves. Mamba2 selectively updates the state vector, allocating greater attention to important temporal segments (such as moments of rapid blood pressure rise) while suppressing noise interference. Its output has the same dimension as the local branches, focusing on long-term temporal correlations.

[0063] The local self-attention mechanism branch and the global feature extraction branch are fused through residual connections, and the numerical distribution is optimized by layer normalization, ultimately generating enhanced features that combine local details and global trends. This design enables the model to identify both instantaneous morphological changes in the pulse wave and correlate them with long-term blood pressure fluctuation patterns, significantly improving the completeness of feature representation.

[0064] (III) Model Training Strategy Model training is based on preprocessed time-series window data. The core configuration and process are as follows: (1) Data loading: DataLoader is used to load the training set, validation set and test set, batch_size=16; shuffle=True is enabled for the training set to enhance generalization, and shuffle=False is kept for the validation set and test set to preserve the time sequence.

[0065] (2) Equipment and optimization configuration: The model automatically adapts to the hardware environment (GPU is preferred, and CPU is used if there is no GPU); the loss function is the mean squared error (MSE), and the optimizer is Adam (learning rate lr=0.001).

[0066] (3) Training process: A total of 50 epochs were trained. In each iteration: during the training phase, the model was enabled in train() mode and the parameters were updated through backpropagation; during the validation phase, the model was switched to eval() mode and gradient calculation was disabled to evaluate the generalization performance. The model with the lowest validation loss was saved as the final model to ensure optimal performance.

[0067] The model was trained using the early stopping method with a batch size of 128. The hardware for network training consisted of an i7-14650HX processor, a 1000GB hard drive, and 16GB of RAM. The software used Anaconda with PyCharm, employing a deep learning framework with a PyTorch backend.

[0068] Experimental Results and Comparative Analysis To verify the effectiveness of the proposed hybrid attention mechanism combined with the Transformer+Mamba2 model in non-invasive arterial blood pressure measurement, the experiment selected six pre-processed simulated pulse wave data as inputs—clinically representative pulse signals. Among them, rapid and fast pulses reflect cardiovascular rhythm function, slippery and choppy pulses are related to vascular elasticity and blood flow status, and even and soft pulses reflect the difference between myocardial contractility and peripheral resistance, covering the core physiological information dimensions required for blood pressure prediction. Simulated arterial blood pressure values ​​acquired and pre-processed by a pressure transmitter were used as the ground truth. The experiment used time-series data from 20 consecutive pulse cycles for model training and validation. Figure 6 The comparison between the predicted results of the proposed model and the actual values ​​of some blood pressure waveforms is shown. Among them, (a) is blood pressure waveform 1, (b) is blood pressure waveform 2, (c) is blood pressure waveform 3, (d) is blood pressure waveform 4, (e) is blood pressure waveform 5, and (f) is blood pressure waveform 6.

[0069] As can be seen from the waveform comparison in Figure 6, for the six different types of pulse signals, the arterial blood pressure prediction curves output by the proposed model maintain a high degree of consistency with the actual blood pressure curves in terms of morphological characteristics (such as peak systolic blood pressure, trough diastolic blood pressure, and position of the dicrotic notch) and temporal trends. Even for abnormal pulse patterns with more complex waveform fluctuations (such as the irregular amplitude of a choppy pulse and the accelerated rhythm of a rapid pulse), the prediction curves can still accurately reproduce the dynamic changes of actual blood pressure without significant phase delay or amplitude deviation, which intuitively verifies the model's adaptability to pulse signals with different physiological characteristics.

[0070] The coefficient of determination (R²) for linear regression is a core statistical indicator for evaluating the goodness of fit of a regression model, and its calculation formula is shown in equation (4). This indicator objectively reflects the degree of linear correlation between predicted and true values ​​by quantifying the proportion of variation of the dependent variable (arterial blood pressure value in this invention) explained by the model, and has wide application value in time-series physiological signal regression tasks. ; in, This represents the actual blood pressure value at the i-th time step. For the corresponding predicted value, The mean of the true blood pressure values ​​is given, and n represents the total number of time steps. The closer the value is to 1, the better the model fit and the stronger its explanatory power for blood pressure variability.

[0071] Table 1 shows the R² results of the predicted and actual blood pressure values ​​under six pulse types: all R² values ​​are greater than 0.95 and close to 1. Among them, the fitting effect is the best for the regular pulse and the characteristic pulse (R²>0.97). Even for the irregular pulse, the R² remains above 0.95.

[0072] The quantitative results further corroborate the qualitative observation conclusions in Figure 6, indicating that the proposed model can not only reproduce blood pressure changes from the intuitive waveform, but also achieve high-precision fitting from the statistical level, fully verifying its reliability and accuracy in the non-invasive prediction of arterial blood pressure, especially its significant advantage in adapting to common abnormal pulse patterns in clinical practice.

[0073] (V) Analysis of Key Error Indicators In order to evaluate the accuracy and stability of the model in arterial blood pressure measurement tasks more comprehensively and from multiple perspectives, this invention introduces three key indicators in addition to R²: mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) to form a multi-level error evaluation system. The calculation formulas are shown in equations (5), (6), and (7).

[0074] ; MSE measures the average of the squared differences between predicted and true values, and is a commonly used loss function in model optimization. A smaller MSE value indicates a lower overall prediction error. RMSE, as the square root of MSE, maintains the same dimensions as the original data and is more sensitive to larger deviations in predictions, effectively revealing the model's performance under extreme conditions. MAE, as the average of the absolute differences between predicted and true values, more intuitively and robustly reflects the average deviation of prediction results, and is less affected by individual outliers. Combining MSE, RMSE, and MAE allows for a comprehensive assessment of the model's overall accuracy and error distribution robustness from three dimensions: loss optimization, sensitivity to large errors, and average deviation. This provides a more comprehensive and multi-dimensional evaluation of the model's predictive performance. Table 1 lists the blood pressure prediction performance indicators of the proposed model under different pulse conditions, including detailed numerical results for R², MSE, RMSE, and MAE for blood pressure prediction under six different pulse conditions.

[0075] Table 1. Blood pressure prediction performance indicators of the proposed model under different pulse conditions.

[0076] Furthermore, Figure 7 Performance evaluation results under different pulse patterns, from Figure 7Observations show that the model's R² for all six pulse types is greater than 0.95 and close to 1, while RMSE and MAE remain at extremely low levels. The rapid pulse exhibits the best predictive accuracy (RMSE=0.0117, MAE=0.0090), benefiting from its "fast and regular rhythm" signal characteristics—high pulse wave temporal stability, and the hybrid attention mechanism accurately capturing the correlation between peak-trough intervals and blood pressure. The hesitant pulse, due to its irregular fluctuations of "thin and slow, difficult to move," results in a more complex mapping relationship between local features and blood pressure, leading to the highest relative error (RMSE=0.0357, MAE=0.0462), but still within a clinically acceptable range. These quantitative results not only corroborate the qualitative conclusions of the waveform comparison but also statistically confirm the model's robustness to different pulse types, laying the foundation for further in-depth performance analysis.

[0077] (vi) Comparison To quantify the necessity of the three core components—hybrid attention mechanism, Mamba2 module, and multi-scale feature preprocessing—four ablation experiments were designed to compare the overall and pulse-specific performance differences between the original model and each ablation variant (removing a single core component). The results are shown in Table 2, comparing the performance of the ablation experiments. The experiments focused on the "universality of performance changes after component removal" and the "specificity of impact on different pulse types" to reveal the core mechanisms of action of the components.

[0078] Table 2 Comparison of Ablation Test Performance

[0079] Furthermore, Figure 8The graph shows a comparison of R² performance in ablation experiments based on thermograms. It can be seen that removing mixed attention results in the most significant performance degradation for abnormal pulse patterns with complex waveforms (hesitant pulse, soft pulse): for hesitant pulse, R² decreases by 0.0693 and RMSE increases by 61.85% (0.0594 vs 0.0367); for soft pulse, R² decreases by 0.0612 and RMSE increases by 48.10% (0.0508 vs 0.0343), confirming that its "local-global collaborative modeling" can compensate for the shortcomings of a single Transformer in modeling irregular pulse patterns. Removing Mamba2 has a significant impact on pulse patterns with strong rhythm (rapid pulse, rapid pulse): for rapid pulse, R² decreases by 0.0123 and RMSE increases by 60.68% (0.0188 vs 0.0117); for rapid pulse, R² decreases by 0.0201 and RMSE increases by 44.94% (0.0416). The structured state-space duality (SSD) framework of the component (vs. 0.0287) can avoid the decay of long-term temporal information. After removing multi-scale preprocessing, the accuracy of pulses (flat pulses and slippery pulses) that are susceptible to noise interference decreases significantly: the RMSE of flat pulses increases by 33.96% (0.0422 vs. 0.0315), and the RMSE of slippery pulses increases by 15.05% (0.0344 vs. 0.0299). This component is crucial for separating noise from physiological features through multi-level convolution kernels.

[0080] In summary, the three core components have a “specific adaptation” feature in supporting the performance of six pulse types: hybrid attention adapts to complex waveforms, Mamba2 adapts to rhythmic timing, and multi-scale preprocessing adapts to noisy interference scenarios. The synergistic effect of the three provides core technical support for high-precision blood pressure prediction of different pulse types.

[0081] Furthermore, to objectively verify the advantages of the proposed model in pulse signal blood pressure prediction, a two-layer comparison framework of "traditional baseline + advanced baseline" was constructed. All models were trained and tested based on data of six pulse signals (rapid pulse, rapid pulse, slippery pulse, normal pulse, soft pulse, and hesitant pulse) collected on the research group's self-built experimental platform. The input feature dimensions and training hyperparameters were standardized to ensure fairness in the comparison. The performance comparison results are shown in Table 3, which compares the performance of the proposed model with the baseline model. Figure 9 The evaluation indexes of different algorithm models on pulse promotion are shown in the figure.

[0082] Table 3 shows the performance comparison between the proposed model and the baseline model.

[0083]

[0084] As shown in Table 3, the model proposed in this invention achieves optimal performance across all six pulse types, and its advantages exhibit a gradient effect. First, compared to the traditional baseline model: for the most error-prone pulse, the proposed model reduces RMSE by 49.36% (0.0705→0.0357) and improves R² by 24.44% (0.7689→0.9568); even for the least difficult pulse to predict, the RMSE is still lower than that of SVM. The decrease of 47.68% (0.0224→0.0117) confirms that traditional models are unable to capture the nonlinear temporal correlation of pulse signals; Secondly, compared to advanced baseline models: Compared to the current mainstream Mamba-SSM, the proposed model reduces the RMSE of noise-sensitive flat pulses by 34.24% (0.0479→0.0315) and increases the R² of irregularly shaped slippery pulses by 22.93% (0.7845→0.9644); the RMSE of BiLSTM-Attention on slippery pulses (0.0388) is 22.94% higher than that of the proposed model (0.0299), highlighting the advantages of the proposed model's "hybrid attention + multi-scale preprocessing" in adapting to complex pulse patterns.

[0085] Furthermore, Figure 10 This is a structural block diagram of the arterial blood pressure wave prediction device according to an embodiment of the application, such as... Figure 10 As shown, the device includes: The information acquisition module is used to acquire dynamic pulse images of the target area and the corresponding arterial blood pressure waveform labels; The dataset construction module is used to preprocess dynamic pulse images and extract feature points to construct a multi-dimensional pulse wave time-series feature dataset. The arterial blood pressure waveform prediction module is used to input multi-dimensional pulse wave temporal features into the constructed and trained blood pressure prediction model to output predicted continuous arterial blood pressure waveforms. The blood pressure prediction model adopts an encoder-decoder architecture. Its encoder core contains a hybrid attention mechanism module. This hybrid attention mechanism module captures the local detail features and long-range dependencies of the pulse wave signal through a parallel local self-attention branch and a global feature extraction branch based on the Mamba2 model, and then fuses the outputs of the two branches.

[0086] The application of the relevant modules of the device in this example can be referred to the relevant introduction of the method principle above, and will not be repeated here.

[0087] Figure 10 The arterial blood pressure wave prediction device in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The electronic device in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0088] Figure 11This is a schematic diagram of the structure of an electronic device 1100 provided in an embodiment of the present invention. The electronic device 1100 can vary significantly due to different configurations or performance characteristics. It may include one or more processors (CPUs) 1110 (e.g., one or more processors) and a memory 1120, and one or more storage media 1130 (e.g., one or more mass storage devices) for storing application programs 1133 or data 1132. The memory 1120 and storage media 1130 may be temporary or persistent storage. The program stored in the storage media 1130 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the electronic device 1100. Furthermore, the processor 1110 may be configured to communicate with the storage media 1130 and execute the series of instruction operations in the storage media 1130 on the electronic device 1100.

[0089] Electronic device 1100 may also include one or more power supplies 1140, one or more wired or wireless network interfaces 1150, one or more input / output interfaces 1160, and / or one or more operating systems 1131, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 11 The illustrated electronic device structure does not constitute a limitation on electronic devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0090] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of any of the above-described arterial blood pressure wave prediction methods.

[0091] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0092] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting arterial blood pressure waves, characterized in that, The method includes: Acquire dynamic pulse images of the target region and corresponding arterial blood pressure waveform labels; The dynamic pulse images are preprocessed and feature points are extracted to construct a multi-dimensional pulse wave time-series feature dataset; A blood pressure prediction model is established, and the multi-dimensional pulse wave temporal features are input into the trained blood pressure prediction model so that the blood pressure prediction model outputs continuous arterial blood pressure waveforms. The blood pressure prediction model adopts an encoder-decoder architecture, and its encoder core includes a hybrid attention mechanism module. This hybrid attention mechanism module captures the local detail features and long-range dependencies of the pulse wave signal through a parallel local self-attention branch and a global feature extraction branch based on the Mamba2 model, and fuses the outputs of the two branches.

2. The arterial blood pressure wave prediction method according to claim 1, characterized in that, The acquisition of dynamic pulse images and corresponding arterial blood pressure waveform labels for the target region includes: A binocular stereo vision acquisition system and a high-precision pressure transmitter are used to acquire data from a target area, which is a simulated radial artery on a simulated radial artery platform. A thin-film probe is placed on the target area to sense and visualize the pulsation of the radial artery in the target area. After the binocular stereo vision acquisition system and the high-precision pressure transmitter are triggered synchronously, the binocular stereo vision acquisition system records a sequence of dynamic pulsation images of the target area, and the pressure transmitter synchronously records the radial artery blood pressure waveform in the target area to establish a one-to-one correspondence between each frame of the image and the corresponding blood pressure value at that moment.

3. The arterial blood pressure wave prediction method according to claim 1, characterized in that, The preprocessing and feature point extraction of the dynamic pulse image to construct a multi-dimensional pulse wave time-series feature dataset includes: The dynamic pulse image is subjected to median filtering to suppress illumination differences and texture noise; The preprocessed image is sequentially divided into blocks for binarization and color inversion, and abnormal regions are removed based on the area threshold of connected regions. Standardized connectivity analysis is performed on the filtered regions to extract the centroids of the target dimension as feature markers. By combining the calibrated camera parameters, the feature markers in the left and right camera images are matched, and their three-dimensional spatial coordinates are solved by a three-dimensional reconstruction algorithm. Thus, the coordinate changes of each feature marker in time series are constructed as a one-dimensional feature, generating a multi-dimensional pulse wave time series feature dataset of the target dimension.

4. The arterial blood pressure wave prediction method according to claim 1, characterized in that, Before inputting the multi-dimensional pulse wave time-series features into the constructed and trained blood pressure prediction model, the method further includes: The pulse wave time series features were processed using a local weighted regression scatter smoothing method, and the synchronously acquired arterial blood pressure waveform labels were subjected to low-pass filtering for noise reduction. At the same time, Butterworth low-pass filters were used to remove baseline drift from both to obtain the initial time series. A normalization method based on the fitting parameters of the training set is adopted to normalize the pulse wave time sequence features and blood pressure wave label amplitudes after the signal purification to the [0,1] interval; and the pulse waves and blood pressure waves of the same type are processed into equal-length sequences by interpolation to establish a one-to-one mapping relationship in order to achieve the standardization of the initial time series. The standardized data of the sequence is organized into feature-label pairs, with the first 16 columns of pulse wave features serving as the model input features and the last column of blood pressure values ​​serving as the target labels. A sliding window method is used to generate continuous input-output sample pairs, and the total dataset is divided into training set, validation set and test set in chronological order.

5. The arterial blood pressure wave prediction method according to claim 1, characterized in that, The multi-dimensional pulse wave temporal features are input into the constructed and trained blood pressure prediction model so that the blood pressure prediction model outputs continuous arterial blood pressure waveforms, including: The multi-dimensional pulse wave temporal features are input into the constructed and trained blood pressure prediction model, so that the blood pressure prediction model sequentially receives the temporal feature matrix processed by the sliding window, performs multi-scale extraction and adaptive filtering on the temporal feature matrix, maps the preprocessed features to a high-dimensional space through linear transformation, extracts and fuses the deep features of the pulse wave signal, generates a prediction sequence based on the encoder output and its own contextual information, and maps the high-dimensional features output by the decoder to the corresponding time step of the arterial blood pressure prediction value output.

6. The arterial blood pressure wave prediction method according to claim 1, characterized in that, The local self-attention branch is configured to employ a multi-head self-attention mechanism, with its attention weights calculated over a consecutive 10 time steps. This is used to capture the local dynamic features of the pulse wave signal. The local self-attention branch maps the input features to a query matrix Q, a key matrix K, and a value matrix V, respectively, and calculates the output according to the attention weight formula. ； ； Where Q is the query matrix obtained from the input feature mapping, K is the key matrix obtained from the input feature mapping, and V is the value matrix obtained from the input feature mapping. It is the query vector at the i-th time step in the query matrix. It is the key vector at the j-th time step of the key matrix. It is its transpose; It is the dimension of the key matrix K. It is the intermediate value of the similarity calculation between the query vector at time step i and the key vector at time step j; Yes The attention weights obtained after Softmax normalization are used to measure the importance of the features at time step j to time step i. It is the value vector of the value matrix at the j-th time step; It is the attention output feature obtained by weighting and aggregating the value vector by attention weights.

7. The arterial blood pressure wave prediction method according to claim 1, characterized in that, The method also includes training a blood pressure prediction model, including: The preprocessed pulse wave time series feature dataset is divided into training set, validation set and test set in chronological order, and continuous input-output sample pairs are generated using the sliding window method. A batch data loader is used to load the training set, validation set and test set in batches of 16. The training set data is randomly shuffled at the beginning of each training cycle, while the validation set and test set data maintain their original temporal order. Set the loss function to mean squared error, the optimizer to Adam optimizer, and the learning rate to 0.001; During the training phase, the model calculates the predicted values ​​through forward propagation, calculates the error through the loss function, and updates the model parameters through backpropagation. During the validation phase, the model is switched to evaluation mode, gradient calculation is stopped, and its generalization performance is evaluated. Throughout the training process, the model parameters with the minimum loss on the validation set are saved as the parameters of the blood pressure prediction model.

8. An arterial blood pressure wave prediction device, characterized in that, The device includes: The information acquisition module is used to acquire dynamic pulse images of the target area and the corresponding arterial blood pressure waveform labels; The dataset construction module is used to preprocess and extract feature points from the dynamic pulse images to construct a multi-dimensional pulse wave time-series feature dataset. The waveform prediction module is used to input the multi-dimensional pulse wave temporal features into the constructed and trained blood pressure prediction model so that the blood pressure prediction model outputs a continuous arterial blood pressure waveform. The blood pressure prediction model adopts an encoder-decoder architecture, and its encoder core includes a hybrid attention mechanism module. This hybrid attention mechanism module captures the local detail features and long-range dependencies of the pulse wave signal through a parallel local self-attention branch and a global feature extraction branch based on the Mamba2 model, and fuses the outputs of the two branches.

9. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the steps of the arterial blood pressure wave prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the arterial blood pressure wave prediction method as described in any one of claims 1-7.

Images