Central arterial pressure reconstruction method and device based on deep learning model

Abstract

The embodiment of the invention discloses a central artery pressure reconstruction method and device based on a deep learning model. The method comprises the following steps: acquiring a current radial artery pressure waveform; the method comprises the following steps: performing data preprocessing on a current radial artery pressure waveform through a preprocessing layer of a model to obtain first time series data, and performing the following processing on the first time series data in sequence through a plurality of residual modules stacked in the model: performing local feature extraction on the input time series data by each residual module through a causal expansion convolution layer of the residual module, intermediate time sequence data is obtained, global feature extraction is conducted on the intermediate time sequence data through an attention layer of the intermediate time sequence data, output time sequence data of the corresponding residual modules are obtained, and the output time sequence data of the last residual module serves as second time sequence data; mapping the second time sequence data into a normalized central arterial pressure waveform through a full connection layer of the model; and performing reverse normalization processing on the normalized central arterial pressure waveform through reverse normalization operation of the model to obtain a central arterial pressure reconstruction result.

Description

Technical Field

[0001] This application relates to the field of biomedical engineering technology, and relates to, but is not limited to, a method and apparatus for central arterial pressure reconstruction based on a deep learning model. Background Technology

[0002] Data shows that cardiovascular diseases account for a significant proportion of deaths worldwide. Cardiovascular diseases are diverse, including coronary heart disease, hypertensive heart disease, and arrhythmias. These diseases not only seriously threaten life and health but also impose a heavy economic and psychological burden on patients and their families, while also placing enormous pressure on social medical resources. Among the pathogenic factors of cardiovascular diseases, abnormal blood pressure is extremely critical. Blood pressure, as an important physiological indicator of cardiovascular health, reflects the heart's pumping function and the pressure exerted on the blood vessel walls. Maintaining normal blood pressure is essential for the normal blood perfusion and function of various organs. When blood pressure remains abnormally high for a long period, whether high or low, it significantly increases the risk of cardiovascular disease. For example, high blood pressure increases the workload on the heart, leading to myocardial hypertrophy, while also damaging the vascular endothelium, accelerating the process of atherosclerosis, and ultimately causing serious complications such as coronary heart disease and stroke.

[0003] Meanwhile, numerous studies have demonstrated that central aortic pressure (CAP) is superior to peripheral arterial pressure in assessing the effectiveness of antihypertensive treatment, predicting cardiovascular events, and guiding clinical decision-making. CAP is the lateral pressure exerted on the proximal aortic vascular structures and is influenced by factors such as left ventricular ejection, arterial resistance, and peripheral arterial load. Compared to peripheral arterial blood pressure, CAP is more closely related to cardiovascular disease and can more directly and accurately reflect the status of the coronary arteries and aorta, making it a primary tool for assessing cardiovascular risk. Furthermore, CAP can also serve as a reference indicator for evaluating the efficacy of antihypertensive drugs. Therefore, accurate CAP measurement has significant clinical value in the prevention, diagnosis, and treatment of cardiovascular disease.

[0004] Invasive central aortic blood pressure measurement is considered the "gold standard" for assessing left ventricular and coronary artery load, but its invasiveness and high cost of consumables prevent its widespread adoption in primary healthcare institutions and among large populations. Traditional non-invasive methods also have limitations such as low measurement accuracy. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and apparatus for reconstructing central aortic pressure based on a deep learning model, which at least solves the problem of low accuracy in central aortic pressure measurement.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a central arterial pressure reconstruction method based on a deep learning model, comprising: The current radial artery pressure waveform is acquired; the current radial artery pressure waveform is preprocessed using the preprocessing layer of the trained TCN-Attention model to obtain first time-series data, the data preprocessing including normalization; the first time-series data is processed sequentially using multiple residual modules stacked in the trained TCN-Attention model as follows: each residual module extracts local features from the input time-series data through its causal dilated convolutional layer to obtain intermediate time-series data, and extracts global features from the intermediate time-series data through its attention layer to obtain the corresponding residual module. The output time-series data of the block is used as the output time-series data of the last residual module as the second time-series data; wherein, the input time-series data of each residual module is the output time-series data of the previous residual module, and the input time-series data of the first residual module is the first time-series data; the second time-series data is mapped to a normalized central arterial pressure waveform through the fully connected layer of the trained TCN-Attention model; the normalized central arterial pressure waveform is denormalized through the denormalization operation of the trained TCN-Attention model to obtain the central arterial pressure reconstruction result.

[0007] Secondly, embodiments of this application provide a central arterial pressure reconstruction device based on a deep learning model, comprising: A first acquisition module is used to acquire the current radial artery pressure waveform; a preprocessing module is used to preprocess the current radial artery pressure waveform through the preprocessing layer of a trained TCN-Attention model to obtain first time-series data, wherein the data preprocessing includes normalization processing; an extraction module is used to sequentially process the first time-series data through multiple residual modules stacked in the trained TCN-Attention model as follows: each residual module performs local feature extraction on the input time-series data through its causal dilated convolutional layer to obtain intermediate time-series data, and performs global feature extraction on the intermediate time-series data through its attention layer to obtain the first time-series data. The output time-series data of the residual module is used as the second time-series data; wherein, the input time-series data of each residual module is the output time-series data of the previous residual module, and the input time-series data of the first residual module is the first time-series data; the mapping module is used to map the second time-series data into a normalized central arterial pressure waveform through the fully connected layer of the trained TCN-Attention model; the denormalization module is used to perform denormalization processing on the normalized central arterial pressure waveform through the denormalization operation of the trained TCN-Attention model to obtain the central arterial pressure reconstruction result.

[0008] The beneficial effects of the technical solutions provided in this application include at least the following: The TCN-Attention model effectively extracts local features from the radial artery pressure waveform using causal dilated convolutional layers, while combining this with an attention mechanism to focus on global features. This allows the model to focus on the long-range dependencies between the radial artery pressure waveform and the central artery pressure waveform, accurately capturing their physiological relationship and achieving efficient and accurate reconstruction of central artery pressure. Residual module stacking and modular design optimize deep feature learning, reducing training difficulty and ensuring model stability and efficiency. This non-invasive reconstruction method is safer and more convenient than traditional invasive measurements, providing a reliable reference for cardiovascular disease diagnosis and combining technological advancement with clinical practicality. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A schematic flowchart of a central arterial pressure reconstruction method based on a deep learning model provided in this application embodiment; Figure 2This is a schematic diagram of the structure of a TCN-Attention model provided in an embodiment of this application; Figure 3 This application provides a schematic diagram comparing the accuracy of prediction results of different models on different performance indicators. Figure 4 A schematic diagram of a central aortic pressure waveform estimated based on different models is provided for embodiments of this application; Figure 5 A schematic diagram of blood pressure values ​​based on a measurement dataset is provided as an embodiment of this application; Figure 6 A histogram based on the blood pressure distribution of a measurement dataset is provided in an embodiment of this application; Figure 7 This is a schematic diagram of the composition of a central arterial pressure reconstruction device based on a deep learning model, provided in an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0012] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0013] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0014] This application provides a method for central arterial pressure reconstruction based on a deep learning model, applied to electronic devices. These electronic devices include, but are not limited to, mobile phones, laptops, tablets, handheld internet devices, multimedia devices, streaming media devices, mobile internet devices, wearable devices, or other types of electronic devices. The functions implemented by this method can be achieved by a processor in the electronic device calling program code. The program code can be stored in a computer storage medium; therefore, the electronic device includes at least a processor and a storage medium. The processor can be used to process the central arterial pressure reconstruction process based on the deep learning model, and the memory can be used to store the data required and generated during the central arterial pressure reconstruction process based on the deep learning model.

[0015] Figure 1 A flowchart illustrating a central arterial pressure reconstruction method based on a deep learning model, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes at least the following steps: Step S110: Obtain the current radial artery pressure waveform; Peripheral arteries refer to the arteries distributed in the trunk, limbs, head and neck, excluding the central arteries (such as the aorta and pulmonary artery) that are directly connected to the heart. They are branches and extensions of the central arteries and are responsible for transporting blood to all tissues and organs of the body. The central arteries and peripheral arteries together constitute the human arterial system. Peripheral arteries include peripheral arteries of the head and neck, peripheral arteries of the upper limbs, peripheral arteries of the trunk, and peripheral arteries of the lower limbs. Among them, peripheral arteries of the upper limbs include the subclavian artery, axillary artery, and brachial artery. The radial artery is an important branch of the brachial artery.

[0016] The radial artery pressure waveform refers to the waveform formed by the regular changes in arterial pressure at the radial artery as the heart contracts and relaxes periodically. This waveform can reflect physiological information such as the heart's pumping function, peripheral vascular resistance, and elasticity. Figure 2 As shown in part (A), the acquired current radial arterial pressure waveform (RAP) can be input into the trained TCN-Attention model.

[0017] The TCN-Attention model is proposed by combining the Temporal Convolutional Network (TCN) model with an attention mechanism. This model further optimizes the TCN model, enhances the model's attention to channel information, and effectively integrates the advantages of TCN and the attention mechanism.

[0018] Step S120: Perform data preprocessing on the current radial artery pressure waveform through the preprocessing layer of the trained TCN-Attention model to obtain first time-series data. The data preprocessing includes normalization processing. like Figure 2 As shown in part (A), the preprocessing layer includes "Normalization" and "TimeAlignment". Normalization is a preprocessing operation for the range of data values. Its purpose is to map data of different magnitudes and units to a unified data range (usually [0,1] or [1,1]) to eliminate interference caused by differences in magnitude. Time alignment is a preprocessing operation for time-series data. Its purpose is to eliminate the misalignment or asynchrony of different data in the time dimension and keep the data synchronized or corresponding on the time axis.

[0019] Step S130: The first time-series data is processed sequentially through multiple residual modules stacked in the trained TCN-Attention model as follows: Each residual module extracts local features from the input time-series data through its causal dilated convolutional layer to obtain intermediate time-series data, and extracts global features from the intermediate time-series data through its attention layer to obtain the output time-series data of the corresponding residual module. The output time-series data of the last residual module is used as the second time-series data. Wherein, the input timing data of each residual module is the output timing data of the previous residual module, and the input timing data of the first residual module is the first timing data; Here, the TCN-Attention model can include multiple stacked residual modules (Residual Module 1 to Residual Module L). The residual modules include causal dilated convolutional layers and attention layers. The residual modules are the core "hierarchical feature extraction" part of the TCN-Attention model. Since there is an inherent physiological transmission relationship between the radial artery pressure waveform and the central artery pressure waveform (the former waveform is transformed from the latter through the arterial system), the bottom layer extracts the local temporal pattern of the radial artery pressure waveform (such as waveform details such as peaks, troughs, and ascending branch slope within a single cycle) through stacked convolutional layers (i.e., causal dilated convolutional layers). The higher-level attention layer further aggregates global contextual information based on the extracted local features, such as the dynamic change pattern between multiple cycles, and finally realizes the global reconstruction of the central artery pressure waveform. The dilation mechanism of the causal dilated convolutional layer makes the receptive field grow exponentially with the number of convolutional layers, enabling the model to extract fine local features and implicitly model long-distance dependencies.

[0020] Local features focus on the details of the radial artery pressure waveform on a short time scale or within a single cycle (such as the systolic peak, adjacent points of the diastolic trough, and the steep rise of the blood pressure waveform), which can be used to distinguish different physiological states (such as changes in heart rate and differences in vascular elasticity). Global features focus on the overall patterns of the radial artery pressure waveform on a long time scale or within multiple cycles (such as the correlation of peaks in multiple cycles, the position of reflected waves, the periodic trend of the blood pressure waveform, and the influence of early systolic waves on the superposition position of later reflected waves), which can be used to reflect the overall dynamics and regulatory function of the cardiovascular system (the overall trend of peak height changes in multiple cardiac cycles can indicate short-term fluctuations in blood pressure (such as changes in blood pressure caused by changes in body position or emotional fluctuations)).

[0021] like Figure 2 As shown in part (A), the input of Residual Module 1 is the preprocessed first timing data, the input of Residual Module 2 is the output of Residual Module 1, the input of Residual Module 3 is the output of Residual Module 2, the input of Residual Module L is the output of Residual Module L-1, and the output of Residual Module L is the second timing data.

[0022] Step S140: Map the second time series data into a normalized central artery pressure waveform through the fully connected layer of the trained TCN-Attention model; The fully connected layer is a key component that connects feature extraction and the final task output. Its core function is to integrate, abstract, and map the features extracted by the preceding network layers (such as causal dilated convolutional layers and attention layers) and finally output a result that meets the task objective.

[0023] Step S150: The normalized central arterial pressure waveform is denormalized by the denormalization operation of the trained TCN-Attention model to obtain the central arterial pressure reconstruction result.

[0024] The parameters for denormalization are the same as those for normalization during data preprocessing; the central arterial pressure reconstruction result is the central arterial pressure waveform predicted by the trained TCN-Attention model, such as... Figure 2 As shown in part (A), the trained TCN-Attention model outputs the central aortic pressure reconstruction result (CAP).

[0025] In some embodiments, the input sequence (i.e., the current radial artery pressure waveform) can be represented as: The output sequence to be predicted (i.e., the central arterial pressure reconstruction result) can be represented as: Thus, the sequence model between the input and output sequences (i.e., the trained TCN-Attention model) can be represented by the following formula (1): Formula (1); In formula (1), G represents the sequence model obtained through the training dataset. This represents the value of the predicted output sequence at time n.

[0026] The stacked convolutional layers of the TCN-Attention model can progressively expand the receptive field, effectively extracting local features from the radial artery pressure waveform, such as the peak value of systolic blood pressure and the trough value of diastolic blood pressure at key time points. The attention mechanism focuses on extracting global patterns. By calculating the correlation between all time steps in the time series data, a global weight distribution is generated, enabling the model to pay attention to the long-range dependencies related to the reconstruction of the radial artery pressure waveform and the central artery pressure waveform (such as the influence of the early systolic waveform on the superposition position of the later reflected wave). The model can simultaneously pay attention to global patterns of different feature dimensions (such as waveform amplitude, slope, curvature, etc.), making up for the limitation of TCN's ability to only perceive locally through the convolutional kernel, thereby achieving efficient reconstruction of CAP. In this embodiment, by introducing an attention mechanism inside the residual module, the model can not only utilize stacked convolutional layers and increase the receptive field of the convolutional kernel to effectively capture local dependencies in the sequence data, but also allow the neural network to assign different weights to the sequence data input at different times, thereby paying more attention to historical moments related to the prediction results, thus achieving a balance between the extraction of local and global features.

[0027] In the above embodiments, the causal dilated convolutional layer of the TCN-Attention model effectively extracts local features from the radial artery pressure waveform, while the attention mechanism focuses on global features. This allows the model to focus on the long-range dependencies between the radial artery pressure waveform and the central artery pressure waveform, accurately capturing their physiological relationship and thus achieving efficient and accurate reconstruction of central artery pressure. The stacking of residual modules and modular design optimize deep feature learning, reducing training difficulty and ensuring model stability and efficiency. This non-invasive reconstruction method is safer and more convenient than traditional invasive measurements, providing a reliable reference for cardiovascular disease diagnosis and combining technological advancement with clinical practicality.

[0028] In some embodiments, each residual module includes a main branch and a skip connection branch, wherein the main branch includes two causal dilated convolutional layers, two weight normalization layers, three activation function layers, two Dropout layers and one attention layer; the skip connection branch includes a 1×1 convolutional layer; The residual module, also known as the residual network, has two branches. The main branch consists of sequentially stacked module components. Two dilated causal convolutional layers ensure no future information leakage during temporal modeling, with the dilation coefficient gradually increasing to expand the receptive field for efficient extraction of local and global temporal features. Two weight normalization layers alleviate gradient explosion and accelerate convergence. Three ReLU activation function layers introduce nonlinear transformations to increase the model's expressive power, and two dropout layers (random deactivation) prevent overfitting. Simultaneously, the main branch uses an attention mechanism to enhance the contribution of key feature points in the blood pressure waveform. Skip connections directly add to the output of the main branch. For the TCN-Attention model, the input and output of the residual module may have different dimensions (channel mismatch). Therefore, a 1×1 convolutional layer is used to ensure dimensionality increase of the input, allowing input and output sequences with the same dimension to be added together.

[0029] like Figure 2 As shown in section (B), the main branch of the residual module is set up from bottom to top as follows: causal dilation convolutional layer, weight normalization layer (WeightNorm), activation function layer (ReLU), dropout layer, causal dilation convolutional layer, weight normalization layer, activation function layer, dropout layer, attention layer, activation function layer.

[0030] The main branch of each residual module is used to perform feature transformation processing on the input timing data of the corresponding residual module to obtain output data; the jump connection branch of each residual module is used to pass the input timing data of the corresponding residual module to the output end of the corresponding residual module, and add it with the output data of the main branch of the corresponding residual module to obtain the output timing data of the corresponding residual module.

[0031] Among them, such as Figure 2 As shown in section (B), the input timing data of residual module i can be represented as follows: The output timing data of residual module i can be represented as , This represents the value at time 0 in the input time series data of residual module i. This represents the value at time step 1 in the input time series data of residual module i. This represents the value at time n-1 in the input time series data of residual module i. This represents the value at time n in the input time series data of residual module i; similarly, This represents the value at time 0 in the output timing data of residual module i. This represents the value at time 1 in the output timing data of residual module i. This represents the value at time n-1 in the output time series data of residual module i. This represents the value at time n in the output timing data of residual module i.

[0032] In order to solve the problems of "gradient vanishing" and "feature degradation" in deep neural network training, the input time series data can be added to the output data to obtain the final output time series data of the residual module, which can enhance the preservation and transmission of the original time series features.

[0033] In the above embodiments, the main branch of the residual module deepens feature extraction through multi-layer causal dilated convolutions and an attention mechanism. The 1×1 convolutional layer with skip connections ensures input-output dimension matching and avoids feature degradation. The dual-branch design enhances the model's ability to learn complex waveform features, improves gradient propagation efficiency, and ensures stable training of deep networks.

[0034] In some embodiments, step S130 above, "each residual module extracts local features from the input time series data through its causal dilated convolutional layer to obtain intermediate time series data," includes the following steps: Step S302, for each moment of the output intermediate time series data, perform the following operations: Step S304: Select historical data from the input time series data of the causal dilated convolutional layer at intervals of k, according to the dilation factor; The input time series data can be represented as follows: The output value of causal dilation convolution at time n It can be described as shown in the following formula (2): Formula (2); in, It is the expansion factor (l = 0, 1, ..., L-1); It is the number of convolutional layers in the network; The filter (f(i), i.e., the convolution kernel, each filter is a vector of length k) is represented. ,in It is the kernel size; () indicates the direction of past time.

[0035] Here, the input temporal data of the causal dilated convolutional layer can be determined according to the dilation factor d. Select k historical data points: x(n), x(nd), x(n-2d), ..., x(n-(k-1)d).

[0036] Step S306: Determine the position of the convolution kernel corresponding to each historical data point; Here, the position of the convolution kernel corresponding to each historical data can be determined. For example, x(n) corresponds to i=0, x(nd) corresponds to i=1, x(n-2d) corresponds to i=2, and x(n-(k-1)d) corresponds to i=k-1.

[0037] Step S308: Multiply each piece of historical data by the weight at the corresponding position of the convolution kernel to obtain the product corresponding to the past time. Step S310: Summate the products of k past time points to obtain the value of the causal dilated convolutional layer at the corresponding time point; Here, as shown in formula (2) above, each historical data point can be multiplied by the weight at the corresponding position of the convolution kernel, and the sum can be obtained to get the value of the intermediate time series data at time n. .

[0038] Step S312: Based on the value of the causal dilated convolutional layer at each time step, obtain the intermediate time series data output by the causal dilated convolutional layer.

[0039] Here, the intermediate time series data can be represented as [F(1),F(2),…,F(n)].

[0040] In the above embodiments, the local feature extraction logic of causal dilated convolution is clearly defined. By selecting historical data at intervals and performing weighted summation, temporal dependencies are accurately captured. This process preserves local waveform details (such as peaks / troughs) while adhering to the causality of physiological signals, laying a reliable foundation for subsequent global feature aggregation.

[0041] In some embodiments, the method includes the following steps: Step S102: Based on the kernel size, dilation factor, and length of the input time series data of the causal dilated convolutional layer, determine the lower limit of the number of convolutional layers of the causal dilated convolutional layer using the receptive field formula to ensure that the receptive field of the trained TCN-Attention model can cover the length of the input time series data.

[0042] To effectively utilize the information from long input sequence signals, the receptive field of the TCN model can be increased by adding a larger filter size and dilation factor. The effective historical receptive field of each layer is... The sum of the receptive fields R of all convolutional layers can be expressed by the following formula (3): Formula (3); like Figure 2 The example shown in section (C) is a causal dilated convolution with 4 hidden layers L, a kernel size k of 2 (as indicated by the arrow), and a dilation factor d = 2. l =1,2,4,8.

[0043] To cover the length of the input sequence signal As shown in the following formula (4), feel the wild and the It must be greater than or equal to N: Formula (4); As can be derived from formula (4), the number of convolutional layers Must be greater than or equal to To ensure that the receptive field covers N, to ensure It can obtain information about the entire input sequence, which can be obtained by solving formula (4).

[0044] In the above embodiments, a formula is used to determine the lower limit of the number of convolutional layers to ensure that the model's receptive field covers the complete input time sequence and avoids the omission of long-distance features due to insufficient receptive field. This design enables the model to fully capture multi-cycle waveform dynamics (such as heart rate variability) and improves the global consistency of central arterial pressure reconstruction.

[0045] In some embodiments, the method further includes the following steps: Step S3011: Based on the kernel size and dilation factor of the causal dilated convolutional layer, determine the padding length as follows; ; Where p is the length, k is the kernel size, and d is the dilation factor; Step S3012: Zero padding of the length is performed on the left side of the input time series data of the causal dilated convolutional layer so that the length of the input time series data of the causal dilated convolutional layer is the same as that of the intermediate time series data.

[0046] The TCN-Attention model requires the output to have the same length as the input. To ensure the output tensor has the same length as the input tensor, zero-padding is applied to the left of the input tensor, while also guaranteeing causal convolution. In convolutional neural networks, p represents padding, k' represents the adjusted kernel size, and s represents the stride; in this embodiment, s = 1. The size of the sequence after causal dilation convolution is... As shown in the following formula (5): Formula (5); After the convolution operation, the TCN-Attention model has a "chomp1d" function to cut off the excess padding. The output size after "chomp1d" processing is... for: Formula (6); Therefore, in order to ensure that the input and output lengths are the same, it is required that ,Right now .

[0047] In the above embodiments, left-side zero padding maintains the consistency of input and output timing lengths, avoiding waveform truncation caused by convolution operations and ensuring timing integrity. The padding strategy neither introduces future information to disrupt causality nor hinders feature alignment across layers, thus improving the model's stability in processing continuous waveforms.

[0048] In some embodiments, the method further includes the following steps: Step S160: Obtain the central arterial pressure reconstruction result and the measured central arterial pressure waveform; Step S161: Based on the central arterial pressure reconstruction result and the measured central arterial pressure waveform, determine the mean square error between the two. The central arterial pressure reconstruction result can be expressed as follows: The measured central arterial pressure waveform can be expressed as: The model's parameters can be minimized by the loss function. Therefore, in order to train the model to the optimal level, the loss function defined in this application embodiment can be the mean squared error (MSE), as shown in formula (7): Formula (7); in, For the first The true value of each sample For the first The predicted value for each sample.

[0049] Step S162: Based on the mean squared error, adjust the parameters of the trained TCN-Attention model through backpropagation to minimize the mean squared error until the mean squared error meets a preset condition.

[0050] here, and These can represent the measured and estimated (i.e., predicted) central arterial pressure waveforms, respectively, with a preset condition that the mean squared error is less than a preset threshold. To evaluate the accuracy of the TCN-Attention model reconstructing the central arterial pressure waveform proposed in this application, the root mean squared error (RMSE) is used as the evaluation metric. A paired t-test (IBM SPSS Statistics, Version-23) is used to analyze the significance between the estimated and actual measured central arterial pressure waveforms. Spearman's correlation coefficient is used to analyze the correlation. The P-value represents the probability of observing the current sample result (or a more extreme result) under the premise that the null hypothesis (H0) is true. P < 0.05 is considered statistically significant. P < 0.05 indicates that the difference between the model estimate and the actual value is unlikely to be caused by random fluctuations, but rather that the model can indeed stably capture the characteristics of the actual waveform (or there is a systematic bias, but this bias is real and not accidental).

[0051] In the above embodiments, the mean squared error is used as the optimization target, and the model parameters are dynamically adjusted through backpropagation to make the reconstructed result approximate the true central arterial pressure waveform. This mechanism achieves end-to-end optimization of the model, reduces systematic errors, improves prediction accuracy, and provides reliable data support for clinical applications.

[0052] In some embodiments, the method further includes the following steps: Step S163: Based on the central arterial pressure reconstruction results and the measured central arterial pressure waveform, determine the mean absolute error, root mean square error, maximum error, and Spearman correlation coefficient between the two. Among them, Mean Absolute Error (MAE) is a commonly used evaluation index. It calculates the average of the absolute values ​​of the prediction errors for each sample, reflecting the magnitude of the average error. The advantage of MAE is that it can intuitively show the difference between the model's predicted value and the actual value. It does not consider the positive or negative of the predicted value, and it pays more attention to the magnitude of the absolute error, which makes it better than Mean Squared Error in reflecting the degree of deviation of the predicted value. The range of MAE is [0, +∞). It is equal to 0 when the predicted value and the actual value are completely consistent, that is, a perfect model; the larger the error, the larger the value. As shown in the following formula (8), N is the number of samples. For the first The true value of each sample For the first The predicted value for each sample.

[0053] Root Mean Square Error (RMSE) is a commonly used metric for evaluating the accuracy of a prediction model. It measures the degree of deviation between the predicted and actual values, and is calculated as the square root of the sum of the squares of the differences between the predicted and actual values. The range of RMSE is [0, +∞). The smaller the value, the smaller the prediction error of the model and the stronger the prediction ability of the model. The specific calculation formula is shown in formula (9) below: Maximum error (MAX Error) is one of the metrics for evaluating the predictive performance of a model. It represents the maximum absolute error between the predicted and actual values ​​across all samples. Its core focus is on the model's performance in the worst-case scenario, rather than the average or overall error level. The specific formula is shown in formula (10) below: in, For the first The true value of each sample For the first The predicted value for each sample.

[0054] The Spearman's Correlation Coefficient (SCC) is a nonparametric statistic used to measure the monotonic relationship between two variables. It is based on the rank of the data rather than the original values, therefore it does not require assumptions that the data follows a normal distribution or a linear relationship. This makes it different from the Pearson correlation coefficient; the Spearman's correlation coefficient is less sensitive to outliers. At the same time, the Spearman's correlation coefficient effectively captures monotonic relationships between variables, whether linear or nonlinear. Where N is the sample size, ... For the first of two variables The rank difference of each sample, that is, the difference between the rank of the true value and the rank of the predicted value, is specifically calculated using the following formula (11): "True value rank" refers to arranging observed real data (such as actual measured central arterial pressure waveform data, actual blood pressure values, etc.) in ascending (or descending) order, and the position number of each data point in this ordered sequence is its rank. "Predicted value rank" is obtained by sorting the model-predicted data (such as central arterial pressure waveform data predicted by the TCN-Attention model, model-predicted blood pressure values, etc.) according to the same rule (ascending or descending), and the position number of each predicted data point in the sorted sequence is its predicted value rank. After calculating the true value rank and the predicted value rank, the rank difference for each sample can be obtained. The Spearman correlation coefficient uses these rank differences to measure the monotonic relationship between two variables (true value and predicted value).

[0055] Step S164: Based on the mean absolute error, the root mean square error, the maximum error, and the Spearman correlation coefficient, evaluate the trained TCN-Attention model to obtain the evaluation result; like Figure 3 As shown, the orange bars represent the performance parameters of the TCN-Attention model, and the blue bars represent the performance parameters of the TCN model. MAE, RMSE, and MAX represent the mean absolute error, root mean square error, and maximum error, respectively. Figure 3 Part (A) represents the performance metrics of the entire waveform (TW). Figure 3 Part (B) represents the performance indicators of systolic blood pressure (SBP). Figure 3 Part (C) represents the performance indicators of diastolic blood pressure (DBP). Figure 3 The (D) part represents the performance index of mean pressure (MAP).

[0056] The TCN-Attention model outperforms the TCN model in reconstructing central arterial pressure. The TCN-Attention model shows significantly lower prediction errors (P<0.05) for mean absolute error, root mean square error, maximum error assessment index, overall waveform, systolic blood pressure, and mean blood pressure, validating the effectiveness of the attention mechanism in extracting temporal features. For diastolic blood pressure prediction, the two models perform similarly, with the TCN-Attention model showing a slight advantage in the results metrics.

[0057] Table 1 below shows that the TCN-Attention model significantly outperformed the TCN-based model in reconstructing the entire central arterial pressure waveform (MAE: 0.57±0.46 mmHg) (MAE: 0.71±0.71 mmHg); the TCN-Attention model (RMSE: 0.69±0.57 mmHg) significantly outperformed the TCN model (RMSE: 0.89±0.86 mmHg); and the TCN-Attention model (MAX: 1.94±1.73 mmHg) significantly outperformed the TCN model (MAX: 2.91±2.09 mmHg). The TCN-Attention model outperformed the TCN model in overall waveform prediction, especially in maximum error control, reducing the error by 33%.

[0058] The TCN-Attention model significantly outperformed the TCN-based model (MAE and RMSE: 1.01±1.73 mmHg) in estimating systolic blood pressure (SBP) (MAE and RMSE: 0.27±0.25 mmHg); the TCN-Attention model (MAX: 0.86±0.24 mmHg) also significantly outperformed the TCN model (MAX: 5.74±3.51 mmHg). The TCN-Attention model performed exceptionally well in SBP prediction, reducing the error by over 70%, with a maximum error reduction of 85%.

[0059] The TCN-Attention model slightly outperformed the TCN-based model (MAE and RMSE: 0.42±0.39 mmHg) in estimating diastolic blood pressure (DBP) (MAE and RMSE: 0.39±0.28 mmHg), but the difference was not significant. The TCN-Attention model (MAX: 1.75±0.51 mmHg) slightly outperformed the TCN model (MAX: 1.16±0.43 mmHg), but the difference was also not significant. The TCN-Attention model performed exceptionally well in DBP prediction, reducing the error by over 70%, with a maximum error reduction of 85%. The TCN model slightly outperformed the TCN-Attention model in DBP prediction, but this did not reach statistical significance, possibly due to the smaller physiological fluctuations in DBP.

[0060] The TCN-Attention model significantly outperformed the TCN-based model in estimating mean pressure (MAE and RMSE: 0.31±0.28 mmHg) (MAE and RMSE: 0.48±0.58 mmHg); the TCN-Attention model also significantly outperformed the TCN model (MAX: 2.02±1.02 mmHg) in estimating mean pressure (MAE and RMSE: 0.31±0.28 mmHg). In MAP prediction, the TCN-Attention model significantly outperformed the TCN model, reducing the error by more than 35% and the maximum bias by 51%, demonstrating a significant improvement in stability.

[0061] The SCC of the entire waveform for both models is as high as 0.9985 and 0.9970, indicating that they can capture the trend of blood pressure changes very well.

[0062] Table 1. Comparison of the accuracy of prediction results of different models on different indicators (Mean ± Standard Deviation, Unit: mmHg) Note: * indicates P < 0.05; ** indicates P < 0.01 like Figure 4 As shown, the black curve represents the actual measured central aortic pressure waveform (CAP), the blue curve represents the central aortic pressure waveform (CAP) estimated by the TCN model in related technologies, and the red curve represents the central aortic pressure waveform (CAP) estimated by the TCN-Attention model in this embodiment. The central aortic pressure waveform estimated by the two different methods shows that the blood pressure estimation based on the TCN-Attention model is closer to the actual measured central aortic pressure waveform than the TCN model across the entire waveform and key feature points.

[0063] Step S165: Update the trained TCN-Attention model based on the evaluation results.

[0064] In the above embodiments, the TCN-Attention model outperforms the traditional TCN model on most metrics, particularly in systolic blood pressure estimation, demonstrating that the attention mechanism effectively captures key features in blood pressure signals. Due to smaller physiological fluctuations in DBP, the model has limited room for improvement. TCN-Attention is a deep learning architecture combining temporal convolutional networks and an attention mechanism, specifically designed for time series prediction tasks. This model utilizes temporal convolutional layers to capture local and long-term dependencies in time series data. The attention mechanism dynamically allocates weights at different time steps, highlighting important features. The standard deviations of TCN-Attention's metrics are generally smaller than those of TCN, indicating more stable and reliable prediction results. Both models have high SCC (Spearman correlation coefficient) values ​​of 0.9985 and 0.9970, respectively, indicating that they can effectively capture blood pressure trends.

[0065] The entire waveform typically encompasses the full-cycle characteristics of the blood pressure signal, such as systole, diastole, and pulse wave morphology. Its numerical variation range is large and highly dynamic. Model prediction requires capturing multiple features simultaneously, leading to different sensitivities of MAE and RMSE to different types of errors (such as systematic bias and random fluctuations). The error distribution of the entire waveform index may exhibit a long-tailed distribution, with a few samples showing extremely large errors. This makes RMSE sensitive to outliers, while MAE is more robust.

[0066] Systolic blood pressure, diastolic blood pressure, and mean blood pressure are all estimated using single keypoints. If the model's estimation bias for a single value mainly stems from systematic biases, such as overall overestimation or underestimation, the error distribution of a single keypoint may approximate a uniform or fixed-bias distribution, causing the calculated MAE and RMSE results to converge. Considering individual differences and the predictive needs of special populations (such as hypertensive patients), multimodal data fusion should be explored, such as combining PPG signals. Future research could explore incorporating more physiological features, such as heart rate variability, to further improve performance.

[0067] In the above embodiments, multi-dimensional evaluation indicators (MAE, RMSE, etc.) comprehensively reflect model performance, quantifying both the magnitude of error and waveform correlation. Updating the model based on the evaluation results allows for targeted optimization of weaknesses (such as extreme value prediction bias), enhancing the model's generalization ability and clinical applicability.

[0068] In some embodiments, the method further includes the following steps: Step S104: Obtain the average value of multiple systolic / diastolic blood pressure readings collected by the mercury sphygmomanometer, and the raw radial artery pulse wave collected by the pressure sensor; Among them, the original radial artery pulse wave is the original radial artery pressure waveform.

[0069] The experiment included 1032 participants with an average age of 28.4 ± 16.4 cm, height of 171.0 ± 33.4 cm, weight of 69.6 ± 21.1 kg, and other parameters, as shown in Table 2. The data acquisition process consisted of two main steps: First, the subjects maintained a stable sitting posture, and their systolic and diastolic blood pressure were collected three times using a mercury sphygmomanometer, with the average value taken. Second, the radial artery pulse wave was collected at least twice. After each collection, if the instrument's built-in signal quality assessment parameter, Operator Index, was less than 80 (e.g., due to improper operation leading to waveform distortion or excessive noise), the data was re-collected to ensure that two sets of data had an Operator Index greater than or equal to 80 (indicating proper operation, minimal interference, and complete waveform characteristics). A force sensor was used to collect the radial artery pulse wave at a sampling frequency of 128 Hz. This device can record pulse wave waveforms in superficial arteries, most commonly the radial artery, followed by the carotid, brachial, or femoral arteries (all peripheral arteries), using a high-fidelity pressure sensor.

[0070] Table 2 Clinical characteristics of the subjects like Figure 5 The figure shows the distribution of systolic blood pressure (SBP) and diastolic blood pressure (DBP) values ​​in the measurement dataset (NEU-PWDB). Figure 5 Part (A) represents the systolic blood pressure of the radial artery (rSBP) and the systolic blood pressure of the central aorta (aSBP). Figure 5 Part (B) represents the diastolic blood pressure (DBP of the radial artery, rDBP) and the diastolic blood pressure (DBP of the central aorta, aDBP). Experimental results show that the systolic blood pressure (rSBP) of the radial artery is higher than that of the central aorta, ranging from 0 mmHg to 43 mmHg (Mean ± SD, 14.59 ± 5.95 mmHg). The diastolic blood pressure (rDBP) of the radial artery is basically the same as that of the central aorta, and in some individuals, the aDBP is even slightly higher than the rDBP, ranging from 0.20 mmHg to 14 mmHg (Mean ± SD, 1.40 ± 1.02 mmHg). The experimental statistical results also conform to the phenomenon of "arterial pressure amplification," with the average difference between the systolic blood pressure of the central and peripheral arteries being approximately 15 mmHg.

[0071] Step S106: Based on the average value of the multiple systolic / diastolic blood pressures, the pressure amplitude of the original radial artery pulse wave collected by the pressure sensor at the same time is calibrated to obtain the calibrated radial artery pressure waveform. Step S108: Based on the calibrated radial artery pressure waveform and the estimated central aortic waveform, the initial TCN-Attention model is trained to obtain the trained TCN-Attention model.

[0072] The process includes calibrating the pressure amplitude of the original radial artery pulse wave to obtain the calibrated radial artery pressure waveform, as well as preprocessing the radial artery pressure waveform and the estimated central aortic waveform.

[0073] In studies using pulse wave waveforms to predict blood pressure (the waveform characteristics of the pulse wave directly correspond to the values ​​of systolic and diastolic blood pressure, which is the core basis for blood pressure prediction; heart rate directly determines the period of the pulse wave), a significant challenge lies in the vast differences in blood pressure values ​​among individuals. For example, the systolic blood pressure of healthy adults typically fluctuates between 90 and 140 mmHg, while the values ​​of patients with hypertension or hypotension may be far above or below this range. This broad numerical distribution threatens the training stability of deep learning models, potentially leading to gradient explosion or vanishing gradients, thus hindering effective model convergence and prediction accuracy. Therefore, standardizing the raw pulse wave data before model training is a crucial step. One of the core tasks of this process is to normalize the numerical range representing blood pressure amplitude in the pulse wave signal. This is significant not only because it makes the data format conform to the model input requirements, but also because it significantly improves training efficiency. By compressing the numerical span, the optimization algorithm can find stable and effective solutions more quickly, accelerating the model to reach a performance plateau. Meanwhile, standardization is also a key strategy for improving data quality. It helps suppress various noise interferences mixed in the signal (including physiological fluctuations, equipment measurement errors, and individual baseline differences), thereby more clearly extracting effective waveform features reflecting blood pressure status. To ensure that the training data format meets the model's requirements, this embodiment standardizes the acquired data sequence before model training, normalizing the range of blood pressure amplitude in the pulse wave waveform sequence. This accelerates the model's training speed and allows it to quickly reach an optimized stable state. The aim is to improve data quality, reduce noise interference, and extract effective features.

[0074] Common time series data preprocessing methods include min-max standardization, mean-variance standardization, mean standardization, and unitization.

[0075] In the following four different data preprocessing methods, x represents the original data. This represents the data obtained after transformation. max indicates the maximum value operation; min indicates the minimum value operation; mean indicates the mean operation; std indicates the variance operation; norm indicates the modulus operation.

[0076] Minimax standardization, also known as deviation standardization, performs a linear transformation on the original data, mapping the data features to the range [0, 1]. The transformation function is shown in the following formula (12): Formula (12); Mean-variance standardization, also known as Z-score standardization, standardizes data by taking the mean and standard deviation of the original data. The preprocessed data follows a standard normal distribution, i.e., the mean is 0 and the standard deviation is 1. The transformation function is shown in the following formula (13): Formula (13); Mean standardization maps data features to the range [-0.5, 0.5], and the transformation function is shown in the following formula (14): Formula (14); Normalization is performed by setting the modulus of the data features to 1, and the transformation function is shown in the following formula (15): Formula (15); Studies have shown that after preprocessing pulse wave sequence data using mean-variance standardization, mean standardization, and unit standardization, the pulse wave waveform predicted by the model exhibits a peak-clipping phenomenon, meaning that peaks and troughs cannot be estimated, resulting in a straight line. This phenomenon is because the mean-variance standardization, mean standardization, and unit standardization preprocessing methods are easily affected by the minimum and maximum values ​​of the pulse wave waveform. If these values ​​exceed the extreme values ​​of the training set, they cannot be effectively processed. In this application embodiment, minimum-maximum standardization is selected as the best preprocessing method. However, the minimum-maximum standardization method also has drawbacks. When new data is added, it may cause changes in the calculated max and min values, requiring redefinition. In this application embodiment, minimum-maximum standardization is specifically used as shown in formulas (16) and (17), where the minimum and maximum values ​​are taken from all sample inputs and outputs.

[0077] Formula (16); Formula (17); here, and These represent the input radial artery pressure waveform signal and the actual central artery pressure waveform signal, respectively. and These represent the normalized radial artery pressure waveform signal and the actual central artery pressure waveform signal, respectively, where 'a' represents... and The minimum value in, b represents and The difference between the maximum and minimum values ​​in the waveform should be noted. It should be pointed out that invasive measurement of central aortic blood pressure cannot be widely used in primary healthcare institutions and among large populations due to cost and safety issues. The actual central arterial pressure waveform signal here is the estimated central aortic waveform.

[0078] like Figure 6 As shown, Figure 6 Part (A) is the systolic pressure of the radial artery. Figure 6 Part (B) represents the diastolic pressure of the radial artery. Figure 6 Part (C) represents the systolic blood pressure of the central artery. Figure 6 Part (D) represents the diastolic pressure of the central artery. For the Pressure Wave Database (PWDB) measurement dataset, the average diastolic pressure at the radial artery is 49 mmHg, and the average systolic pressure is 178 mmHg. Considering the dataset, in formulas (16) and (17), a is 45 mmHg and b is 135 mmHg.

[0079] In the above embodiments, by combining the calibration value of the mercury sphygmomanometer with the original waveform of the sensor, accurate calibration of the radial artery pressure waveform is achieved, eliminating equipment errors. The model is trained based on the calibrated data, improving the reliability of the input features and providing a high-quality data source for subsequent central arterial pressure reconstruction.

[0080] In some embodiments, the method further includes the following steps: Step S1071: Based on the original radial artery pulse wave, generate the estimated central aortic waveform using a general transformation function; or, Step S1072: Based on the calibrated radial artery pressure waveform, the estimated central aortic waveform is generated using a patient-specific calibration model.

[0081] Here, a general conversion function or a patient-specific calibrated mathematical model can be applied to convert this peripheral artery waveform into an estimated central aortic waveform, thereby obtaining parameters such as central systolic blood pressure, diastolic blood pressure, and pulse pressure.

[0082] In the above embodiments, two methods for estimating the central aortic waveform are provided: a general transformation function adapts to the commonalities of the population, while a patient-specific model takes into account individual differences. Flexible waveform generation strategies enhance the diversity of training data and improve the model's adaptability to different populations (such as healthy / disease-prone individuals).

[0083] Based on the foregoing embodiments, this application further provides a central arterial pressure reconstruction device based on a deep learning model. The device includes various modules and units included in each module, which can be implemented by a processor in an electronic device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0084] Figure 7 A schematic diagram of the composition structure of a central arterial pressure reconstruction device based on a deep learning model provided in this application embodiment is shown below. Figure 7 As shown, the device 700 includes a first acquisition module 710, a preprocessing module 720, an extraction module 730, a mapping module 740, and an inverse normalization module 750, wherein: The first acquisition module 710 is used to acquire the current radial artery pressure waveform; The preprocessing module 720 is used to preprocess the current radial artery pressure waveform through the preprocessing layer of the trained TCN-Attention model to obtain first time-series data, wherein the data preprocessing includes normalization processing. The extraction module 730 is used to process the first time-series data sequentially through multiple residual modules stacked in the trained TCN-Attention model as follows: each residual module extracts local features from the input time-series data through its causal dilated convolutional layer to obtain intermediate time-series data, extracts global features from the intermediate time-series data through its attention layer to obtain the output time-series data of the corresponding residual module, and uses the output time-series data of the last residual module as the second time-series data; Wherein, the input timing data of each residual module is the output timing data of the previous residual module, and the input timing data of the first residual module is the first timing data; The mapping module 740 is used to map the second time-series data into a normalized central artery pressure waveform through the fully connected layer of the trained TCN-Attention model. The inverse normalization module 750 is used to perform inverse normalization processing on the normalized central arterial pressure waveform through the inverse normalization operation of the trained TCN-Attention model to obtain the central arterial pressure reconstruction result.

[0085] In some possible embodiments, each residual module includes a main branch and a skip connection branch, wherein the main branch includes two causal dilated convolutional layers, two weight normalization layers, three activation function layers, two dropout layers, and one attention layer; the skip connection branch includes a 1×1 convolutional layer; the main branch of each residual module is used to perform feature transformation processing on the input temporal data of the corresponding residual module to obtain output data; the skip connection branch of each residual module is used to pass the input temporal data of the corresponding residual module to the output end of the corresponding residual module, and add it with the output data of the main branch of the corresponding residual module to obtain the output temporal data of the corresponding residual module.

[0086] In some possible embodiments, the extraction module includes: a selection submodule, configured to perform the following operations for each moment of the output intermediate time series data: selecting historical data from the input time series data of the causal dilated convolutional layer at intervals of k past moments according to the dilation factor; a first determination submodule, configured to determine the position of the convolutional kernel corresponding to each historical data; a first calculation submodule, configured to multiply each historical data by the weight corresponding to the position of the convolutional kernel to obtain the product of the corresponding past moment; a second calculation submodule, configured to sum the products of the k past moments to obtain the value of the causal dilated convolutional layer at the corresponding moment; and a second determination submodule, configured to obtain the intermediate time series data output by the causal dilated convolutional layer based on the value of the causal dilated convolutional layer at each moment.

[0087] In some possible embodiments, the apparatus 700 further includes: a first determining module, configured to determine a lower limit of the number of convolutional layers of the causal dilated convolutional layer based on the kernel size, dilation factor, and length of the input time-series data, using a receptive field formula, to ensure that the receptive field of the trained TCN-Attention model can cover the length of the input time-series data.

[0088] In some possible embodiments, the apparatus 700 further includes: a second determining module, configured to determine the length of the padding based on the kernel size and dilation factor of the causal dilated convolutional layer in the following manner; Where p is the length, k is the kernel size, and d is the dilation factor; the padding module is used to zero-padding the left side of the input time series data of the causal dilation convolutional layer to make the length of the input time series data of the causal dilation convolutional layer the same as that of the intermediate time series data.

[0089] In some possible embodiments, the device 700 further includes: a second acquisition module for acquiring the central arterial pressure reconstruction result and the measured central arterial pressure waveform; a third determination module for determining the mean square error between the central arterial pressure reconstruction result and the measured central arterial pressure waveform; and an adjustment module for adjusting the parameters of the trained TCN-Attention model based on the mean square error through backpropagation to minimize the mean square error until the mean square error meets a preset condition.

[0090] In some possible embodiments, the device 700 further includes: a fourth determining module, configured to determine the mean absolute error, root mean square error, maximum error, and Spearman correlation coefficient between the central arterial pressure reconstruction result and the measured central arterial pressure waveform; an evaluation module, configured to evaluate the trained TCN-Attention model based on the mean absolute error, the root mean square error, the maximum error, and the Spearman correlation coefficient, and obtain an evaluation result; and an updating module, configured to update the trained TCN-Attention model based on the evaluation result.

[0091] In some possible embodiments, the device 700 further includes: a third acquisition module, configured to acquire the average value of multiple systolic / diastolic blood pressure measurements collected by a mercury sphygmomanometer and the raw radial artery pulse wave collected by a pressure sensor; a calibration module, configured to perform pressure amplitude calibration on the raw radial artery pulse wave collected by the pressure sensor simultaneously based on the average value of the multiple systolic / diastolic blood pressure measurements, to obtain a calibrated radial artery pressure waveform; and a training module, configured to train an initial TCN-Attention model based on the calibrated radial artery pressure waveform and an estimated central aortic waveform, to obtain the trained TCN-Attention model.

[0092] In some possible embodiments, the device 700 further includes: a first generation module for generating the estimated central aortic waveform based on the original radial artery pulse wave using a general transformation function; or a second generation module for generating the estimated central aortic waveform based on the calibrated radial artery pressure waveform using a patient-specific calibration model.

[0093] It should be noted that the descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the system embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0094] It should be noted that, in the embodiments of this application, if the above-mentioned pipeline corrosion state prediction method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, 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 an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0095] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps in any of the deep learning model-based central arterial pressure reconstruction methods described in the above embodiments. Correspondingly, embodiments of this application also provide a computer program product, which, when executed by a processor of an electronic device, is used to implement the steps in any of the deep learning model-based central arterial pressure reconstruction methods described in the above embodiments.

[0096] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0097] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0098] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0099] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. In addition, each functional unit in the embodiments of this application may be fully integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in the form of hardware plus software functional units.

[0100] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, 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 the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0101] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.

[0102] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for reconstructing central arterial pressure based on a deep learning model, characterized in that, include: Obtain the current radial artery pressure waveform; The current radial artery pressure waveform is preprocessed using the preprocessing layer of the trained TCN-Attention model to obtain first time-series data, the data preprocessing including normalization. The first time-series data is processed sequentially by multiple residual modules stacked in the trained TCN-Attention model as follows: each residual module extracts local features from the input time-series data through its causal dilated convolutional layer to obtain intermediate time-series data, and extracts global features from the intermediate time-series data through its attention layer to obtain the output time-series data of the corresponding residual module. The output time-series data of the last residual module is used as the second time-series data. Wherein, the input timing data of each residual module is the output timing data of the previous residual module, and the input timing data of the first residual module is the first timing data; The second time-series data is mapped to a normalized central arterial pressure waveform through the fully connected layer of the trained TCN-Attention model; By performing an inverse normalization operation on the trained TCN-Attention model, the normalized central arterial pressure waveform is inversely normalized to obtain the central arterial pressure reconstruction result.

2. The method according to claim 1, characterized in that, Each residual module includes a main branch and a skip connection branch, wherein the main branch includes two causal dilated convolutional layers, two weight normalization layers, three activation function layers, two Dropout layers and one attention layer; the skip connection branch includes a 1×1 convolutional layer. The main branch of each residual module is used to perform feature transformation processing on the input time series data of the corresponding residual module to obtain output data; The jump connection branch of each residual module is used to pass the input timing data of the corresponding residual module to the output terminal of the corresponding residual module, and add it with the output data of the main branch of the corresponding residual module to obtain the output timing data of the corresponding residual module.

3. The method according to claim 1, characterized in that, Each of the residual modules extracts local features from the input time-series data through its causal dilated convolutional layer to obtain intermediate time-series data, including: For each moment of the output intermediate time series data, perform the following operations: Based on the inflation factor, select historical data from the input time series data of the causal dilated convolutional layer at intervals of k past moments; Determine the position of the convolution kernel corresponding to each historical data point; Multiply each piece of historical data by the weight at the corresponding position of the convolution kernel to obtain the product at the corresponding past time. The summation of the products of k past time points yields the value of the causal dilated convolutional layer at the corresponding time point; Based on the value of the causal dilated convolutional layer at each time step, the intermediate time series data output by the causal dilated convolutional layer is obtained.

4. The method according to claim 3, characterized in that, The method further includes: Based on the kernel size, dilation factor, and length of the input time series data of the causal dilated convolutional layer, the lower limit of the number of convolutional layers of the causal dilated convolutional layer is determined by the receptive field formula to ensure that the receptive field of the trained TCN-Attention model can cover the length of the input time series data.

5. The method according to claim 1, characterized in that, The method further includes: Based on the kernel size and dilation factor of the causal dilated convolutional layer, the padding length is determined as follows: ； Where p is the length, k is the kernel size, and d is the dilation factor; Zero padding of the length is applied to the left side of the input time series data of the causal dilated convolutional layer so that the length of the input time series data of the causal dilated convolutional layer is the same as that of the intermediate time series data.

6. The method according to claim 1, characterized in that, The method further includes: Obtain the central arterial pressure reconstruction results and the measured central arterial pressure waveform; Based on the central arterial pressure reconstruction results and the measured central arterial pressure waveform, the mean square error between the two is determined; Based on the mean squared error, the parameters of the trained TCN-Attention model are adjusted through backpropagation to minimize the mean squared error until the mean squared error meets a preset condition.

7. The method according to claim 6, characterized in that, The method further includes: Based on the central arterial pressure reconstruction results and the measured central arterial pressure waveform, the mean absolute error, root mean square error, maximum error, and Spearman correlation coefficient between the two are determined. The trained TCN-Attention model is evaluated based on the mean absolute error, the root mean square error, the maximum error, and the Spearman correlation coefficient to obtain the evaluation results. Based on the evaluation results, the trained TCN-Attention model is updated.

8. The method according to claim 1, characterized in that, The method further includes: Obtain the average value of multiple systolic / diastolic blood pressure readings from a mercury sphygmomanometer, as well as the raw radial artery pulse wave from a pressure sensor; Based on the average value of the multiple systolic / diastolic blood pressures, the pressure amplitude of the original radial artery pulse wave collected simultaneously by the pressure sensor is calibrated to obtain the calibrated radial artery pressure waveform. Based on the calibrated radial artery pressure waveform and the estimated central aortic waveform, the initial TCN-Attention model is trained to obtain the trained TCN-Attention model.

9. The method according to claim 8, characterized in that, The method further includes: Based on the original radial artery pulse wave, the estimated central aortic waveform is generated using a general transformation function; or, Based on the calibrated radial artery pressure waveform, the estimated central aortic waveform is generated using a patient-specific calibration model.

10. A central arterial pressure reconstruction device based on a deep learning model, characterized in that, include: The first acquisition module is used to acquire the current radial artery pressure waveform; The preprocessing module is used to preprocess the current radial artery pressure waveform through the preprocessing layer of the trained TCN-Attention model to obtain first time-series data, wherein the data preprocessing includes normalization processing. The extraction module is used to process the first time-series data sequentially through multiple residual modules stacked in the trained TCN-Attention model as follows: each residual module extracts local features from the input time-series data through its causal dilated convolutional layer to obtain intermediate time-series data, extracts global features from the intermediate time-series data through its attention layer to obtain the output time-series data of the corresponding residual module, and uses the output time-series data of the last residual module as the second time-series data; Wherein, the input timing data of each residual module is the output timing data of the previous residual module, and the input timing data of the first residual module is the first timing data; The mapping module is used to map the second time-series data into a normalized central artery pressure waveform through the fully connected layer of the trained TCN-Attention model; The inverse normalization module is used to inverse normalize the normalized central arterial pressure waveform through the inverse normalization operation of the trained TCN-Attention model to obtain the central arterial pressure reconstruction result.

Images