CBL-iTransform model-based central arterial pressure waveform reconstruction method

Abstract

The invention discloses a central arterial pressure waveform reconstruction method based on a CBL-iTransform model, and relates to the technical field of central arterial pressure evaluation. According to the method, a waveform reconstruction model (CML-iTransform) is obtained through the design of the method, the waveform reconstruction model is based on an improved iTransform architecture, and a one-dimensional CNN and a BiLSTM network are fused, so that the data feature extraction capability in the waveform reconstruction process is remarkably improved, and reconstruction from an RAP waveform to a CAP waveform is realized; wherein the RAP waveform is firstly input into the feature extraction module, then the reconstructed CAP waveform is output through the waveform reconstruction module, the data sample size is expanded through data enhancement, and the model performance is improved; compared with a traditional method and several deep learning network models, the model capable of capturing long and short-term memory dependence is superior to other methods in the aspect of reconstructing the central arterial pressure from the radial arterial pressure, and it is proved that the model is good in reconstruction performance.

Description

Technical Field

[0001] The present invention relates to the technical field of central arterial pressure waveform reconstruction, and in particular to a central arterial pressure waveform reconstruction method based on a CBL-iTransformer model. Background Art

[0002] Cardiovascular disease (CVD) is a vascular disorder affecting the heart and other tissues caused by factors such as hyperlipidemia, increased blood viscosity, and atherosclerosis. Despite significant advances in preventive strategies over the past few decades, CVD remains a leading cause of death and a major global threat. Blood pressure, as a parameter used to measure the status of the cardiovascular system, plays a crucial role in patient health assessment, disease treatment, and other areas. CAP, or aortic pressure, refers to the lateral pressure exerted on the blood vessels at the root of the ascending aorta. In clinical practice, the CAP waveform contains rich physiological and pathological information. Extracting CAP waveform parameters can accurately reflect important parameters such as central aortic systolic pressure, central aortic diastolic pressure, and augmentation index. These parameters are more strongly correlated with target organ damage and the risk of cardiovascular events. Therefore, accurate and rapid CAP waveform reconstruction methods are crucial.

[0003] CAP measurement methods are categorized as invasive and non-invasive. Traditional CAP measurement typically relies on invasive cardiac catheterization. A catheter is inserted into a large artery (such as the aorta or carotid artery) to directly measure blood pressure. However, this method places high demands on both the equipment and the person performing the measurement, making it difficult to popularize clinically. Currently, non-invasive measurement methods have become a common approach for measuring CAP, including direct substitution, transfer function, blind recognition, and deep learning model-based measurement methods.

[0004] To avoid invasive surgery, early researchers developed several alternative methods to estimate CAP. Kelly et al. found that the carotid pressure waveform and the CAP waveform were highly similar, suggesting the possibility of using carotid pressure to estimate CAP. Studies have confirmed that carotid pressure can be used to estimate CAP. Similar to the direct substitution method for the carotid pressure waveform, brachial artery pressure can also be used as a surrogate for CAP. However, CAP levels do not always correspond to peripheral blood pressure levels, and the direct substitution of peripheral blood pressure for CAP has many drawbacks. Therefore, researchers have proposed establishing a functional relationship between peripheral blood pressure and CAP to reconstruct the CAP waveform.

[0005] The generalized transfer function (GTF) method is a widely used approach that can be generated from both the time-domain and frequency-domain characteristics of a system. Karamanoglu et al. proposed a GTF in the frequency domain and demonstrated that it can be used to establish a relationship between central and peripheral arterial pressures at varying levels of accuracy. Chen et al. established an autoregressive model (ARX) and verified that the GTF can accurately estimate CAP using a radial artery sphygmomanometer. The reconstructed waveform can provide an estimate of central artery pressure, but an accurate CAP waveform requires a more realistic reproduction of the waveform profile, and therefore, the estimated value may be lower than the true value. Although the transfer function method is simple and can achieve good measurement accuracy, each transfer function is specific to a specific population. Changes in the population require the construction of a new transfer function, which is often difficult for clinicians to accomplish independently. Consequently, blind system identification techniques have been introduced.

[0006] Blind system identification (BSI), as a blind signal processing method, can use the output signal of an unknown system model to identify the input signal, thereby obtaining the required unknown information, such as the system model parameters. Ahmed et al. proposed a blind system identification method based on Malliavin calculus, effectively demonstrating its superiority over traditional methods. Swamy et al. estimated CAP from less invasive peripheral blood pressure using multichannel blind system identification. Liu et al. proposed a multichannel Newton blind system identification algorithm to reconstruct aortic pressure waveforms. The root mean square error of the CAP waveform measured and reconstructed using this method is smaller than that of canonical correlation analysis and is less sensitive to noise. BSI, when used to measure central aortic pressure, does not require a predefined model and is applicable to a wide range of populations. However, this method relies on a large amount of peripheral arterial data, and some of the underlying theories of BSI are still imperfect, so this method is still under exploration.

[0007] In recent years, due to the high specificity and complexity of blood pressure signals, traditional blood pressure feature extraction methods have become inadequate to meet the needs of modern medicine. Machine learning, the study of patterns from data using computational models and algorithms, has numerous applications in various fields requiring the discovery of patterns from complex data and has become one of the core technologies in the broader field of artificial intelligence. Deep learning, a branch of machine learning, uses multi-layered nonlinear transformations to extract complex patterns and features from data, thereby addressing the challenge of information processing in large datasets. Deep learning has been widely applied in fields such as speech recognition, machine vision, image processing, signal processing, and natural language processing. Convolutional neural networks (CNNs) are a very popular and widely used deep learning network. They excel at extracting spatial features from data that effectively represent the class or quantity of objects. The key lies in performing multi-layered, continuous transformations on the input data (such as pooling operations) and convolutional processing on the data at different spatial scales. Long short-term memory networks (LSTMs) possess excellent predictive performance due to their ability to retain information over long distances, making them widely used in time series forecasting. BiLSTM is a variant of LSTM that can fully capture and utilize the forward and backward correlation characteristics of time series.

[0008] With the advancement of machine learning, more and more researchers are applying deep learning models to CAP waveform reconstruction. Xiao et al. proposed using a CNN-BiLSTM network to reconstruct the central aortic pressure waveform from the radial pressure waveform, validating that this network is a feasible and accurate method for reconstructing the supine CAP waveform from the RAP waveform. Fen et al. proposed continuous blood pressure measurement using a single-channel electrocardiogram (ECG) signal and verified the reliability of data from ICU patients and patients with arrhythmias. Furthermore, Xiao et al. also used an artificial neural network to estimate CAP from radial artery systolic and diastolic pressures alone. Experiments demonstrated that CAP reconstructed from RAP is highly consistent with invasive CAP measurements. Pan et al. proposed a Bi-Unet deep learning model to reconstruct arterial blood pressure signals from photoplethysmography signals, while also improving its predictive performance. While these deep learning models improve CAP reconstruction performance, they still face computational efficiency challenges.

[0009] In recent years, the Transformer has made significant progress in time series prediction through its outstanding parallel computing capabilities. It processes sequence data in parallel, improving processing efficiency and employing a self-attention mechanism to allow for interaction between elements in the input sequence, effectively capturing contextual information. However, the Transformer's positional encoding mechanism lacks flexibility when processing long sequences of data. To address this issue, Liu et al. summarized the Transformer, arguing that its components have not been fully applied to time series. They proposed the iTransformer model, which inverts the global representation of the embedded sequence to be more variable-centric, enhancing the temporal representation of the input information.

[0010] In summary, the accuracy of waveform reconstruction by traditional methods is low, and some deep learning models have difficulty in extracting features, which hinders their clinical application and promotion. Therefore, a new solution to the above problems is needed. Summary of the Invention

[0011] The purpose of the present invention is to provide a central arterial pressure waveform reconstruction method based on the CBL-iTransformer model, aiming to improve the model feature extraction capability and reconstruction accuracy, so as to solve the technical problems raised in the background technology.

[0012] To achieve the above objectives, the present invention provides the following technical solution: a central arterial pressure waveform reconstruction method based on the CBL-iTransformer model, comprising at least the following steps:

[0013] S1: Use a pre-validated dataset that includes CAP (central aortic pressure) and RAP (radial aortic pressure) samples to support comparison and analysis of CAP waveform estimation methods;

[0014] S2: Data preprocessing: The total samples are divided into training and test sets in a ratio of 9:1. To improve the generalization ability and prediction accuracy of the model and the reliability of model evaluation, the sliding window technique is used to perform data augmentation on the training and test sets: starting from the first data point of each sample, segments of a fixed length of 500 data points are cyclically intercepted, and the generated subsequences are added to the training and test sets respectively to form new training and test sets;

[0015] S3: Build a CBL-iTransformer model. The CBL-iTransformer model is based on the improved iTransformer network. By analyzing the characteristics of the CAP waveform, a corresponding feature extraction module is designed to capture key time-frequency information from the RAP waveform, and a waveform reconstruction module is used to establish a nonlinear mapping relationship between RAP and CAP, ultimately realizing the reconstruction of the CAP waveform from the RAP waveform.

[0016] Preferably, the CBL-iTransformer model includes a feature extraction module and a waveform reconstruction module;

[0017] The feature extraction module consists of two one-dimensional convolutions and BiLSTM;

[0018] The waveform reconstruction module is composed of an improved iTransformer module;

[0019] The RAP waveform is the input of the CBL-iTransformer model, and the CAP waveform is the label corresponding to the training CBL-iTransformer model.

[0020] Preferably, the application of the CBL-iTransformer model comprises at least the following steps:

[0021] First, the RAP data after S2 preprocessing is input into the feature extraction module, and a one-dimensional convolution unit is used to perform convolution operation on the long-term RAP time series data using a sliding window to extract local waveform features.

[0022] Then, the feature sequence output by the convolution is input into the BiLSTM unit, and the long-term and short-term context dependencies are captured through bidirectional temporal modeling, thereby enhancing the representation ability of the features in the time dimension.

[0023] Finally, the temporal features output by BiLSTM are input into the improved iTransformer network, and global dependencies are established through the improved iTransformer module to achieve end-to-end waveform reconstruction from PAP to CAP.

[0024] Preferably, the feature extraction module includes a convolutional layer and a bidirectional long short-term memory network, and the convolutional layer includes two one-dimensional convolutions, namely Conv1 and Conv2;

[0025] The application of the feature extraction module comprises at least the following steps:

[0026] The RAP waveform contains a lot of useful information. The convolution layer uses different convolution kernels to extract multi-scale local features of the input RAP waveform through a sliding window.

[0027] Perform batch normalization on the convolution output to accelerate training convergence and prevent overfitting;

[0028] Enhance the nonlinear expression ability of features through activation functions;

[0029] Finally, the multi-scale local features are spliced ​​in the channel dimension to obtain a comprehensive feature representation containing multi-scale information;

[0030] The fused features are input into BiLSTM, and long-term and short-term dependencies are established through bidirectional time series analysis.

[0031] Preferably, the feature extraction module uses BiLSTM to extract time series features. For the sequence X(x1, x2, ... x n ,n=1,2,…), the feature extraction process is as follows:

[0032] h1=Conv1(X)h2=Conv2(X)(1)

[0033]

[0034]

[0035] f=Concat(a1,a2)(4)

[0036] y t =BiLSTM(f t ),t=1,2,…,T (5)

[0037] Among them, h 1, h2 is the local feature obtained by different convolution kernels; is the data after batch normalization; a 1, a2 is the feature after activation function; f is the multi-scale feature after fusion; y t It is the temporal feature representation output by BiLSTM at each time step t.

[0038] Preferably, the waveform reconstruction module is based on the iTransformer architecture, which is a variant of the Transformer. By inverting the global representation of the embedding sequence and introducing a multi-head attention mechanism, it can better process the extracted features and use these features to reconstruct the central arterial pressure waveform. An LSTM layer is added after the embedding layer of the iTransformer architecture to learn and save important time information in the input sequence through a gating mechanism, further enhancing data processing.

[0039] Preferably, the application of the waveform reconstruction module includes at least the following steps:

[0040] The high-dimensional features output by the feature extraction module are converted into fixed-dimensional vector representations through the Embedding layer to establish the semantic encoding of the input features;

[0041] Use LSTM to process the embedding vector and learn the temporal dynamic characteristics of the feature sequence;

[0042] The data is passed to the multivariate attention layer, which calculates the attention weight of each time step and assigns different weights to different features, thereby capturing the most important part of the sequence for reconstructing the target and ignoring irrelevant parts;

[0043] LayerNorm is used to normalize the attention output, reduce the scale differences between different features, ensure a more stable distribution of data between different layers, improve training stability, and reduce the gradient vanishing or exploding problem of the model;

[0044] Finally, the processed features are mapped to the final output space to obtain the final reconstruction result, namely the CAP waveform.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention obtains a waveform reconstruction model (CBL-iTransformer) through method design. The waveform reconstruction model is based on the improved iTransformer architecture and integrates a one-dimensional CNN and BiLSTM network. It significantly improves the data feature extraction capability during the waveform reconstruction process and realizes the reconstruction from RAP waveform to CAP waveform. First, the RAP waveform is input into the feature extraction module, and then the reconstructed CAP is output through the waveform reconstruction module.

[0047] Since the cost of invasive radial artery pressure data and central arterial pressure data measurement is high, resulting in a small training dataset, data augmentation is used to expand the data sample size and improve model performance;

[0048] By comparing with traditional methods and several deep learning network models, the CBl-iTransformer model used in this invention, which captures long-term and short-term memory dependencies, is superior to other methods in reconstructing central arterial pressure from radial artery pressure, which proves the good reconstruction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0050] Figure 1 This is the CBL-iTransformer model diagram of the present invention;

[0051] Figure 2 This is a structural diagram of the feature extraction module of the present invention;

[0052] Figure 3 This is a comparison chart between LSTM and BiLSTM of the present invention;

[0053] Figure 4 This is a waveform reconstruction module diagram of the present invention;

[0054] Figure 5 This is a schematic diagram of central arterial pressure systolic and diastolic pressure in the present invention;

[0055] Figure 6 Schematic diagram of the reconstruction effect of the model of the present invention on different samples;

[0056] Figure 7 This is a Bland-Altman comparison diagram of the CASP predicted values ​​and true values ​​reconstructed by different models under the same conditions of the present invention. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0058] See also Figure 1 The central arterial pressure waveform reconstruction method based on the CBL-iTransformer model at least includes the following steps:

[0059] S1: A pre-validated dataset containing CAP (central aortic pressure) and RAP (radial artery pressure) samples was used to support the comparison and analysis of CAP waveform estimation methods.

[0060] The data set used in this embodiment is derived from a previous study by Pauca et al., which recorded invasive radial artery and central arterial blood pressure data of 62 patients (mean age 61±11 years, 45 males, 17 females, 58 hypertensive patients) before and after medication. The study was approved by the institutional ethics committee, and informed consent was obtained from all participants before the examination. Each subject underwent two CAP and RAP measurements before and during vascular intervention, forming 124 sets of data. The pressure gauge system used has a frequency response of more than 20 Hz, ensuring high-fidelity recording of the aorta and radial artery. Each set of samples was resampled to 128 Hz by linear interpolation from the data originally collected at 200 Hz, and 1024 sampling points were extracted for CAP waveform reconstruction training and testing. Finally, 124 sets of CAP and RAP samples were obtained to support the comparison and analysis of CAP waveform estimation methods.

[0061] S2: Data preprocessing: The total samples are divided into training and test sets in a ratio of 9:1. To improve the generalization ability and prediction accuracy of the model and the reliability of model evaluation, the sliding window technique is used to perform data augmentation on the training and test sets: starting from the first data point of each sample, segments of a fixed length of 500 data points are cyclically intercepted, and the generated subsequences are added to the training and test sets respectively to form new training and test sets;

[0062] S3: Build a CBL-iTransformer model. The CBL-iTransformer model is based on the improved iTransformer network. By analyzing the characteristics of the CAP waveform, a corresponding feature extraction module is designed to capture key time-frequency information from the RAP waveform, and a waveform reconstruction module is used to establish a nonlinear mapping relationship between RAP and CAP, ultimately realizing the reconstruction of the CAP waveform from the RAP waveform.

[0063] The CBL-iTransformer model includes a feature extraction module and a waveform reconstruction module;

[0064] The feature extraction module consists of two one-dimensional convolutions and BiLSTM;

[0065] The waveform reconstruction module is composed of the improved iTransformer module;

[0066] The RAP waveform is the input to the CBL-iTransformer model, and the CAP waveform is the corresponding label used to train the CBL-iTransformer model. The dataset contains both RAP and CAP waveforms. The RAP waveform is processed by the CBL-iTransformer network model (feature extraction module and waveform reconstruction module) to obtain the predicted CAP. The predicted CAP is then compared with the CAP labels in the dataset to obtain the error.

[0067] The application of the CBL-iTransformer model includes at least the following steps:

[0068] First, the RAP data after S2 preprocessing is input into the feature extraction module, and a one-dimensional convolution unit is used to perform convolution operation on the long-term RAP time series data using a sliding window to extract local waveform features.

[0069] The radial artery pressure signal includes at least systolic and diastolic pressure features. The radial artery pressure signal varies in time and amplitude. The one-dimensional convolution unit can recognize these different patterns by learning the weights of the convolution kernel and encode them into feature vectors. In addition, multiple independent and parallel one-dimensional convolution units can be combined with different convolution kernels to learn more complex blood pressure signal features. Using convolution kernels of different sizes, local features of different scales can be captured simultaneously and fused in the channel to obtain a more comprehensive feature representation.

[0070] Then, the feature sequence output by the convolution is input into the BiLSTM unit, and the long-term and short-term context dependencies are captured through bidirectional temporal modeling, thereby enhancing the representation ability of the features in the time dimension.

[0071] Finally, the temporal features output by BiLSTM are input into the improved iTransformer network, and global dependencies are established through the improved iTransformer module to achieve end-to-end waveform reconstruction from PAP to CAP.

[0072] The feature extraction module includes convolutional layers and bidirectional long short-term memory networks, such as Figure 2 As shown, the convolution layer includes two one-dimensional convolutions, namely Conv1 and Conv2;

[0073] The application of the feature extraction module includes at least the following steps:

[0074] The RAP waveform contains a lot of useful information. The convolution layer uses different convolution kernels to extract multi-scale local features of the input RAP waveform through a sliding window.

[0075] Perform batch normalization on the convolution output to accelerate training convergence and prevent overfitting;

[0076] Enhance the nonlinear expression ability of features through activation functions;

[0077] Finally, the multi-scale local features are spliced ​​in the channel dimension to obtain a comprehensive feature representation containing multi-scale information;

[0078] The fused features are input into BiLSTM, and long-term and short-term dependencies are established through bidirectional time series analysis.

[0079] LSTM and BiLSTM are both variants of recurrent neural networks (RNNs) used to process time series data. They excel in capturing temporal dependencies and processing long sequences of data. LSTM controls the flow of information through input gates, forget gates, and output gates. BiLSTM is an extension of LSTM, implemented through two parallel LSTM networks: one from the past to the future (forward), and the other from the future to the past (backward), meaning that data is processed in both directions. Figure 3 shown.

[0080] The feature extraction module uses BiLSTM to extract time series features. For the sequence X(x1,x2,…x n ,n=1,2,…), the feature extraction process is as follows:

[0081] h1=Conv1(X)h2=Conv2(X)(1)

[0082]

[0083]

[0084] f=Concat(a1,a2)(4)

[0085] y t =BiLSTM(f t ),t=1,2,…,T (5)

[0086] Among them, h1 and h2 are local features obtained by different convolution kernels; is the data after batch normalization; a1, a2 are the features after activation function; f is the multi-scale feature after fusion; y t It is the temporal feature representation output by BiLSTM at each time step t.

[0087] The waveform reconstruction module is based on the iTransformer architecture, which is a variant of the Transformer. By inverting the global representation of the embedding sequence and introducing a multi-head attention mechanism, it can better process the extracted features and use these features to reconstruct the central arterial pressure waveform. The LSTM layer is added after the embedding layer of the iTransformer architecture to learn and save important time information in the input sequence through a gating mechanism, further strengthening data processing. Figure 4 shown.

[0088] The application of the waveform reconstruction module includes at least the following steps:

[0089] The high-dimensional features output by the feature extraction module are converted into fixed-dimensional vector representations through the Embedding layer to establish the semantic encoding of the input features;

[0090] Use LSTM to process the embedding vector and learn the temporal dynamic characteristics of the feature sequence;

[0091] The data is passed to the multivariate attention layer, which calculates the attention weight of each time step and assigns different weights to different features, thereby capturing the most important part of the sequence for reconstructing the target and ignoring irrelevant parts;

[0092] Layer normalization is used to standardize the attention output, reduce the scale differences between different features, ensure a more stable distribution of data between different layers, improve training stability, and reduce the gradient vanishing or exploding problem of the model;

[0093] Finally, the processed features are mapped to the final output space to obtain the final reconstruction result, namely the CAP waveform.

[0094] The characteristic sequence zt is input into the waveform reconstruction module and is expressed as follows:

[0095]

[0096]

[0097] z=TemBlock(h) (8)

[0098] a=Projection(z) (9)

[0099] Among them, zt = [zt1, zt2, …, ztn] is a feature sequence. After the sequence is processed through each layer, the final reconstruction result a is obtained, that is, the CAP waveform.

[0100] Performance indicators

[0101] The present invention uses the mean absolute error between the CAP waveform prediction value and the true value, the root mean square error, the root mean square error between the CASP and CADP waveform prediction values ​​and the true value (MAECAP, RMSECAP, RMSECASP, RMSECADP) as evaluation indicators of waveform reconstruction effect.

[0102] MAE: Mean Absolute Error is a common metric for evaluating the accuracy of a forecasting model. It measures the average difference between the predicted value and the true value. The smaller the value, the smaller the difference between the forecasting model and the true value, and the higher the forecast accuracy. The specific formula is as follows:

[0103]

[0104] RMSE: Root mean square error is the square root of the average of the squares of the differences between the model's predicted values ​​and the true values. The specific formula is as follows:

[0105]

[0106] Among them, h(x i ) represents the true value, y i represents the predicted value, and m represents the number of samples.

[0107] (3)RMSE CASP / CADP :CASP is the maximum pressure exerted by blood on the arterial wall during cardiac contraction, and CADP is the minimum pressure during cardiac relaxation. Figure 5 Comparing the RMSE of the predicted and true values ​​of systolic and diastolic blood pressure can further evaluate the performance of the model.

[0108] Based on the above embodiments, the following results and discussions are proposed:

[0109] Experimental results

[0110] The radial artery pressure waveform in the test set samples is input into the model and the prediction performance after data processing, feature extraction and reconstruction is as follows: Figure 6 As shown in the figure, different samples show good reconstruction performance on this model.

[0111] As can be seen in the figure, the CBL-iTransformer model's predictions (red curve) for reconstructing the central arterial pressure waveform closely match the actual invasively measured waveform (blue curve). Despite slight local deviations, the overall trend remains consistent, demonstrating good prediction accuracy. This consistency across multiple samples further demonstrates the model's reliability and generalization capabilities in reconstructing central arterial pressure waveforms.

[0112] Performance Comparison

[0113] To verify the effectiveness of the CBL-iTransformer model for CAP waveform reconstruction, we compared it with traditional methods and several deep learning models. The MAE and RMSE between the predicted and true CAP waveforms, and the RMSE between the predicted and true CASP and CADP values, were used as evaluation metrics to verify the model's performance in reconstructing central aortic pressure. The results are shown in Table 1.

[0114] Table 1 Performance comparison of CBL-iTransformer model and other models

[0115]

[0116] The results show that the CBL-iTransformer model has good reconstruction performance for central aortic pressure. The mean absolute error and root mean square error between the predicted value and the true value of the central aortic pressure waveform reconstructed by the model, and the root mean square error (MAE) between the predicted value and the true value of the central aortic systolic pressure and central aortic diastolic pressure waveforms are shown in Figure 2. CAP :0.93±0.90mmHg,RMSE CAP :1.30±0.90mmHg,RMSE CASP :1.44±0.84mmHg, RMSE CADP :1.29±0.78mmHg) are all improved compared with other models.

[0117] In order to further verify whether the CBL-iTransformer model can achieve the desired effect, we show the BlandAltman comparison of the CASP predicted values ​​and true values ​​of the model and three methods (ARX, CNN-BiLSTM, CBi-SAN), and perform consistency analysis, such as Figure 7 shown.

[0118] Figure 7 Bland-Altman comparison plots of the CASP predictions and true values ​​reconstructed by the ARX, CNN-BiLSTM, CBi-SAN, and CBL-iTransformer models under the same conditions are presented to assess the consistency between the CASP predictions and true values ​​of each model. The x-axis represents the average of the CASP predictions and true values ​​for each team reconstructed by different models under the same conditions. The y-axis represents the difference. The two red horizontal solid lines represent the upper and lower limits of the mean difference (±1.96 times the standard deviation), respectively. These limits show the range of consistency between the two measurements under ideal conditions. The horizontal dashed line represents the mean difference between the two methods, indicating systematic bias.

[0119] As can be seen, the errors of all four methods fall within the limits of consistency, indicating good consistency between the predictions of these models and the invasive measurements. The CBL-iTransformer model, in particular, shows a high degree of agreement between its predicted CASP and the actual measured waveform, demonstrating higher prediction accuracy and stability.

[0120] In summary:

[0121] The present invention proposes a central arterial pressure reconstruction method based on the CBl-iTransformer model, which reconstructs the CAP from the RAP waveform, further improving the CAP reconstruction performance. Comparative experiments were conducted on traditional methods and deep learning methods, and the advantages of the iTransformer model in reconstructing the central arterial pressure waveform were found. LSTM was introduced to enhance the dependencies between local data, and CNN and BiLSTM were introduced as feature extraction modules of the model. Data augmentation was used to increase the number of training samples, avoid overfitting, and improve the training ability of the model. Experimental results show that the CBl-iTransformer model has good CAP reconstruction performance and is expected to be effectively applied and promoted in future clinical practice.

[0122] The traditional central arterial pressure reconstruction method mainly establishes a functional relationship between peripheral blood pressure (carotid artery, brachial artery, etc.) and central arterial pressure. The ARX autoregressive model establishes a functional relationship between historical measurement values ​​and current data. NPMA performs square wave convolution on radial artery data to reconstruct central arterial pressure. Although these methods can reconstruct central arterial pressure, their accuracy is low. With the development of machine learning, researchers continue to improve deep learning models (LSTM-BiLSTM, CNN-BiLSTM, CBi-SAN, etc.) to continuously improve the accuracy of central arterial pressure reconstruction. Based on previous research, the present invention proposes a waveform reconstruction model CBl-iTransformer, which is based on the iTransformer architecture and integrates CNN and LSTM networks to improve the feature extraction ability of data during waveform reconstruction. The results show that the reconstruction effect of central arterial pressure of the model of the present invention is better than that of traditional methods and some deep learning network models on the same data set, and is expected to be effectively applied and promoted in future clinical practice.

[0123] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A central arterial pressure waveform reconstruction method based on the CBL-iTransformer model, characterized by: At least the following steps are included: S1: A pre-confirmed data set is used, which includes CAP samples and RAP samples to support the comparison and analysis of CAP waveform estimation methods. The CAP is the central aortic pressure and the RAP is the radial artery pressure. S2: Data preprocessing: The total samples are divided into training and test sets in a ratio of 9:

1. To improve the generalization ability and prediction accuracy of the model and the reliability of model evaluation, the sliding window technique is used to perform data augmentation on the training and test sets: starting from the first data point of each sample, segments of a fixed length of 500 data points are cyclically intercepted, and the generated subsequences are added to the training and test sets respectively to form new training and test sets; S3: Build a CBL-iTransformer model. The CBL-iTransformer model is based on the improved iTransformer network. By analyzing the characteristics of the CAP waveform, a corresponding feature extraction module is designed to capture key time-frequency information from the RAP waveform, and a waveform reconstruction module is used to establish a nonlinear mapping relationship between RAP and CAP, ultimately realizing the reconstruction of the CAP waveform from the RAP waveform.

2. The central arterial pressure waveform reconstruction method based on the CBL-iTransformer model according to claim 1, characterized in that: The CBL-iTransformer model includes a feature extraction module and a waveform reconstruction module; The feature extraction module consists of two one-dimensional convolutional networks and a bidirectional long short-term memory network; The waveform reconstruction module is composed of an improved iTransformer module; The RAP waveform is the input of the CBL-iTransformer model, and the CAP waveform is the label corresponding to the training CBL-iTransformer model.

3. The central arterial pressure waveform reconstruction method based on the CBL-iTransformer model according to claim 1, characterized in that: The application of the CBL-iTransformer model includes at least the following steps: First, the RAP data after S2 preprocessing is input into the feature extraction module, and a one-dimensional convolution unit is used to perform convolution operation on the long-term RAP time series data using a sliding window to extract local waveform features. Then, the feature sequence output by the convolution is input into the BiLSTM unit, and the long-term and short-term context dependencies are captured through bidirectional temporal modeling, thereby enhancing the representation ability of the features in the time dimension. Finally, the temporal features output by BiLSTM are input into the improved iTransformer network, and global dependencies are established through the improved iTransformer module to achieve end-to-end waveform reconstruction from PAP to CAP.

4. The central arterial pressure waveform reconstruction method based on the CBL-iTransformer model according to claim 2, characterized in that: The feature extraction module includes a convolutional layer and a bidirectional long short-term memory network, and the convolutional layer includes two one-dimensional convolutions, namely Conv1 and Conv2; The application of the feature extraction module comprises at least the following steps: The RAP waveform contains a lot of useful information. The convolution layer uses different convolution kernels to extract multi-scale local features of the input RAP waveform through a sliding window. Perform batch normalization on the convolution output to accelerate training convergence and prevent overfitting; Enhance the nonlinear expression ability of features through activation functions; Finally, the multi-scale local features are spliced ​​in the channel dimension to obtain a comprehensive feature representation containing multi-scale information; The fused features are input into BiLSTM, and long-term and short-term dependencies are established through bidirectional time series analysis.

5. The central arterial pressure waveform reconstruction method based on the CBL-iTransformer model according to claim 3, characterized in that: The feature extraction module uses BiLSTM to extract time series features. n ,n=1,2,…), the feature extraction process is as follows: h1=Conv1(X)h2=Conv2(X)(1) f=Concat(a1,a2)(4) y t =BiLSTM(f t ),t=1,2,…,T(5) Among them, h1 and h2 are local features obtained by different convolution kernels; is the data after batch normalization; a1, a2 are the features after activation function; f is the multi-scale feature after fusion; y t It is the temporal feature representation output by BiLSTM at each time step t.

6. The central arterial pressure waveform reconstruction method based on the CBL-iTransformer model according to claim 2, characterized in that: The waveform reconstruction module is based on the iTransformer architecture, which is a variant of the Transformer. By inverting the global representation of the embedding sequence and introducing a multi-head attention mechanism, it can better process the extracted features and use these features to reconstruct the central arterial pressure waveform. A long short-term memory network layer is added after the embedding layer of the iTransformer architecture to learn and save important time information in the input sequence through a gating mechanism, further enhancing data processing.

7. The central arterial pressure waveform reconstruction method based on the CBL-iTransformer model according to claim 6, characterized in that: The application of the waveform reconstruction module comprises at least the following steps: The high-dimensional features output by the feature extraction module are converted into fixed-dimensional vector representations through the Embedding layer to establish the semantic encoding of the input features; Use LSTM to process the embedding vector and learn the temporal dynamic characteristics of the feature sequence; The data is passed to the multivariate attention layer, which calculates the attention weight of each time step and assigns different weights to different features, thereby capturing the most important part of the sequence for reconstructing the target and ignoring irrelevant parts; Layer normalization is used to standardize the attention output, reduce the scale differences between different features, ensure a more stable distribution of data between different layers, improve training stability, and reduce the gradient vanishing or exploding problem of the model; Finally, the processed features are mapped to the final output space to obtain the final central arterial pressure reconstruction result, namely the CAP waveform.

Images