A method and system for invasive blood pressure waveform estimation based on generative adversarial networks

Abstract

This invention provides a method and system for invasive blood pressure waveform estimation based on generative adversarial networks (GANs), comprising: acquiring photoplethysmography (PPG) signals from a user; inputting the PPG signals into a trained blood pressure waveform estimation and monitoring model to obtain the user's invasive blood pressure waveform. The model construction method includes: acquiring PPG signals and continuous blood pressure waveform signals and performing data preprocessing to obtain preprocessed data; constructing a generator and a discriminator based on GAN principles, wherein the generator incorporates a spatial-temporal attention mechanism to improve the accuracy of blood pressure waveform generation, and the discriminator increases the blood pressure waveform perception field through fully convolutional layers to improve waveform discrimination capability; and training the network model using the preprocessed data to obtain a trained blood pressure waveform estimation and monitoring model. This invention achieves high-precision blood pressure waveform estimation using GAN principles, obtaining high-precision invasive blood pressure waveforms based on non-invasively acquired PPG signals.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for invasive blood pressure waveform estimation based on generative adversarial networks. Background Technology

[0002] The latest World Health Organization database shows that cardiovascular disease mortality in low- and middle-income countries has increased by nearly 50% in the past 20 years, with an estimated morbidity rate about 30% higher than in high-income countries. Developing low-cost continuous flow blood pressure (BP) monitoring technology can alleviate the burden of care, effectively improve prognosis, and thus reduce mortality. Changes in blood pressure waveforms provide a wealth of information about an individual's dynamic cardiovascular status. This information can help in the rapid diagnosis of cardiovascular dysfunction and monitor responses to irregular medications used to correct hypotension or hypertension, preventing their progression.

[0003] Invasive brachial artery cannulation is the gold standard for obtaining continuous blood pressure waveforms, but it carries risks of pain and infection. Therefore, non-invasive blood pressure monitoring is more suitable for health management outside of intensive care units (ICUs). However, non-invasive monitoring methods typically require cuffs, such as auscultation and oscillometric methods, which can cause the cuff to inflate and obstruct the blood vessel during testing, hindering long-term blood pressure monitoring. Novel cuffless blood pressure measurement methods acquire pulse waveforms using advanced sensing technologies such as photoplethysmography (PPG), radar, or ultrasound. Based on pulse wave velocity (PWV) theory, ballistocardiography (BCG) or electrocardiography (ECG) signals can also be incorporated. In particular, blood pressure prediction methods based on pulse transient time (PTT) have the potential to achieve high-accuracy BP estimation because PPG can utilize optical properties to reflect underlying hemodynamic information. However, acquiring ECG signals requires the placement of additional sensors or adhesive electrodes, and the potential discomfort and risk of infection limit the need for long-term monitoring. Furthermore, these methods fit blood pressure values ​​to abstract statistical features (e.g., obtaining discrete systolic and diastolic blood pressure values), making them unsuitable for estimating continuous blood pressure waveforms.

[0004] Recently, inspired by the successful applications of deep learning in various fields, some technical methods and research have attempted to establish methods for non-invasive blood pressure waveform estimation. Compared with traditional machine learning methods that require manual extraction of PPG morphological features, the advantage of using deep learning methods lies in their data-driven nature of automatically learning the optimal features mapped to the BP waveform from the PPG signal. Existing methods for estimating the BP waveform include autoregressive models, deep convolutional autoencoders (DCAEs), and UNet and V-Net models in medical image segmentation. Autoregressive models perform sequence prediction using sampling points before and after the PPG signal, considering waveform details in only one direction. Furthermore, the sequence-to-sequence (Seq2Seq) model proposed by Aguiree et al. has a standard deviation error (SDE) of 15.67 mmHg for systolic blood pressure (SBP), which does not meet the Association for the Advancement of Medical Instrumentation (AAMI) standard (SDE should not exceed 8 mmHg).

[0005] On the other hand, autoencoders offer better predictive performance. This technique uses an encoder to extract features from the PPG signal, increasing its depth by reducing the length of the input signal; that is, the encoder performs a latent space representation of the PPG signal. The decoder then decodes the latent vectors in the bottleneck into the corresponding BP waveform. However, function mappings are lost when information is passed from the encoder to the decoder. Therefore, some existing techniques or methods concatenate the decoder layer with the corresponding encoder layer to compensate for the features lost during BP waveform reconstruction. For example, the VNet model based on skip connections has been shown in papers to reduce the SDE of SBP to 6.53 mmHg, but the mean error (ME) increases to 4.30 mmHg. These techniques still suffer from large BP waveform prediction errors, and even on large sample datasets, they sometimes fail to meet the medical-grade standards specified by AAMI, which may increase the risk of clinical monitoring. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a method and system for invasive blood pressure waveform estimation based on generative adversarial networks.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] This invention provides a method for invasive blood pressure waveform estimation based on generative adversarial networks, comprising:

[0009] Collect the photoplethysmography (PPG) signal of the user under test;

[0010] The photoplethysmography (PPG) signal of the user under test is input into the trained estimation and monitoring model to obtain the blood pressure signal of the user under test.

[0011] The method for constructing the estimation monitoring model includes:

[0012] Collect photoplethysmography (PPG) pulse wave signals and continuous blood pressure waveform data from sample users;

[0013] The photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users were preprocessed to obtain preprocessed data.

[0014] An estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography pulse waves is constructed. The estimation and monitoring model includes a signal waveform estimation model, an objective optimization function, an attention mechanism module, a generator, and a discriminator.

[0015] The preprocessed data is used to train the estimation and monitoring model for invasive blood pressure waveform based on photoplethysmography, resulting in a well-trained estimation and monitoring model.

[0016] Preferably, the construction of the estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography includes:

[0017] A signal waveform estimation model based on conditional generative adversarial network is constructed to estimate the invasive blood pressure waveform based on photoplethysmography pulse wave signal, and the objective optimization function is designed.

[0018] An attention mechanism module is constructed to implement a method for estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network.

[0019] The architecture of the generator and the discriminator for a method of estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network is constructed.

[0020] Preferably, the data preprocessing of the photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users to obtain preprocessed data includes:

[0021] The photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users were down-frequencyd to obtain a down-frequency signal.

[0022] The down-frequency signal is filtered using a finite impulse response filter with a preset cutoff frequency to obtain a filtered signal;

[0023] The filtered signal is phase-aligned to obtain an aligned signal;

[0024] The alignment signal is subjected to maximum-minimum normalization to obtain a normalized signal;

[0025] The normalized signal is windowed to obtain the preprocessed data.

[0026] Preferably, the construction of the signal waveform estimation model based on the conditional generative adversarial network to estimate the invasive blood pressure waveform from the photoplethysmography (PPG) signal, and the design of the objective optimization function, includes:

[0027] Construct a GAN model; the GAN model consists of a parameterized generator and a parameterized discriminator; the generator and the discriminator perform adversarial learning based on the game theory principle of the min-max optimization task;

[0028] The adversarial loss is determined based on the generator and the discriminator;

[0029] The mean absolute error between the generated BP waveform and the actual BP waveform is calculated based on the adversarial loss.

[0030] Calculate the mean absolute error between the feature map of the reference BP waveform and the generated BP waveform;

[0031] The perceptual loss is determined based on the mean absolute error between the feature map of the reference BP waveform and the generated BP waveform.

[0032] Calculate the mean-square error (MSE) between the reference blood pressure waveform and the estimated waveforms of the systolic and diastolic points;

[0033] The signal waveform estimation model is constructed based on the GAN model, the adversarial loss, the mean absolute error between the generated BP waveform and the real BP waveform calculated by the adversarial loss, the mean absolute error between the feature map of the reference BP waveform and the generated BP waveform, the perceptual loss, and the MSE.

[0034] The objective optimization function is determined based on the signal waveform estimation model.

[0035] Preferably, the attention mechanism module for constructing a method for estimating invasive blood pressure waveforms based on photoplethysmography (PPG) signals using conditional generative adversarial networks includes:

[0036] Obtain the first feature map;

[0037] The first feature map input is transformed into a first query channel and a first key-value channel using convolutional layer kernels;

[0038] Perform matrix multiplication, convolution, and layer normalization on the first query channel and the first key-value channel to obtain the first calculation result;

[0039] The spatial attention module is obtained based on the first calculation result and the feature channel;

[0040] Obtain the second feature map;

[0041] The second feature map input is transformed into a second query channel and a second key value channel using convolutional layer kernels;

[0042] The second query channel is enhanced using Softmax, and the enhanced information is then multiplied by the second key-value channel to obtain the second calculation result.

[0043] The time attention module is obtained based on the second calculation result and the time axis;

[0044] The temporal attention module and the spatial attention module are merged in parallel to obtain the attention mechanism module; the attention mechanism module is set in the skip layer and the connection layer.

[0045] Preferably, the architecture of the generator and the discriminator for constructing a method for estimating invasive blood pressure waveforms based on photoplethysmography (PPG) signals using conditional generative adversarial networks includes:

[0046] A generator based on UNet3+ is constructed; the generator based on UNet3+ is used to combine high-level semantics from feature maps of different scales with low-level semantics through full-scale skip connections;

[0047] Construct a discriminator based on PatchGAN.

[0048] Preferably, the step of training the estimation and monitoring model for invasive blood pressure waveform based on photoplethysmography (PPG) using the preprocessed data to obtain the trained estimation and monitoring model includes:

[0049] The preprocessed data is divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5.

[0050] The estimated monitoring model is trained using the training set and the validation set, and the signal waveform estimation model is validated using the test set to obtain the trained estimated detection model.

[0051] Preferably, it further includes:

[0052] During network training, the Adam optimizer is used to update the parameters; the learning rate is 1e-3, the weight decay is 0.999, and the momentum is 0.8.

[0053] Another aspect of the present invention provides a system for invasive blood pressure waveform estimation based on generative adversarial networks, comprising:

[0054] The first acquisition unit acquires the photoplethysmography (PPG) signal of the user under test.

[0055] The estimation unit is used to input the photoplethysmography (PPG) signal of the user under test into the trained estimation monitoring model to obtain the blood pressure signal of the user under test.

[0056] The second acquisition unit is used to acquire photoplethysmography (PPG) signals and continuous blood pressure waveform data of the sample users.

[0057] The preprocessing unit is used to preprocess the photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users to obtain preprocessed data.

[0058] The model building unit is used to build an estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography pulse waves; the estimation and monitoring model includes a signal waveform estimation model, an objective optimization function, an attention mechanism module, a generator, and a discriminator;

[0059] The training unit is used to train the estimation and monitoring model of invasive blood pressure waveform based on photoplethysmography pulse wave according to the preprocessed data, so as to obtain the trained estimation and monitoring model.

[0060] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0061] This invention provides a method and system for invasive blood pressure waveform estimation based on generative adversarial networks. The method includes: acquiring photoplethysmography (PPG) signals of a user under test; inputting the PPG signals of the user under test into a trained estimation monitoring model to obtain the user's blood pressure signal; the method for constructing the estimation monitoring model includes: acquiring PPG signals and continuous blood pressure waveform data of sample users; performing data preprocessing on the PPG signals and continuous blood pressure waveform data of the sample users to obtain preprocessed data; constructing an estimation monitoring model for invasive blood pressure waveforms based on PPG signals; the estimation monitoring model includes a signal waveform estimation model, an objective optimization function, an attention mechanism module, a generator, and a discriminator; training the estimation monitoring model for invasive blood pressure waveforms based on PPG signals according to the preprocessed data to obtain a trained estimation monitoring model. The input of this invention consists only of photoplethysmography (PPG) waves. Therefore, the acquisition circuit only needs to acquire PPG signals. Compared with the traditional method based on pulse wave propagation speed, it eliminates the step of acquiring electrocardiogram (ECG) signals and does not require excessive derivative calculations. This makes it easy to integrate into devices such as wristbands, eliminating the need for blood pressure measurement equipment such as cuffs, freeing the device from the constraints of cuffs, and making it more portable. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0064] Figure 2 This is a flowchart illustrating the main steps of a method for estimating invasive blood pressure waveforms based on PPG using generative adversarial networks, as provided in an embodiment of the present invention.

[0065] Figure 3 A specific experimental design flowchart is provided for embodiments of the present invention;

[0066] Figure 4 This is a schematic diagram of the overall architecture of the PPG2BP-cGAN model provided in an embodiment of the present invention;

[0067] Figure 5 A schematic diagram of the structure of the plug-and-play modified polarization self-attention (MPSA) module in the embedded model provided in the embodiments of the present invention;

[0068] Figure 6A schematic diagram of the modified polarization self-attention (MPSA) mechanism provided in an embodiment of the present invention;

[0069] Figure 7 A schematic diagram of a generator structure based on UNet3+ and incorporating the MPSA attention mechanism provided in an embodiment of the present invention.

[0070] Figure 8 A schematic diagram illustrating the second decoding layer provided in an embodiment of the present invention;

[0071] Figure 9 A schematic diagram illustrating the visualization and qualitative comparison of prediction results from different PPG-based BP waveform estimation models provided in embodiments of the present invention; wherein, Figure 9 (a) is a schematic diagram of the input normalized PPG signal. Figure 9 (b) is a schematic diagram of PPG2BP-cGAN. Figure 9 (c) is a schematic diagram of cGAN based on UNet++. Figure 9 (d) is a schematic diagram of VNet. Figure 9 (e) is a classic CycleGAN diagram. Figure 9 (f) is a schematic diagram of the BP waveform estimated by NARX. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0074] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, including a series of steps, processes, methods, etc., is not limited to the steps listed, but may optionally include steps not listed, or may optionally include other steps inherent to these processes, methods, products, or devices.

[0075] The purpose of this invention is to provide a method and system for invasive blood pressure waveform estimation based on generative adversarial networks (GANs). Since the BP waveform estimation problem can be viewed as a distribution transformation problem from the PPG signal to the BP waveform, this invention proposes to utilize a conditional generative adversarial network (cGAN), named PPG2BP-cGAN, as the backbone network for supervising BP waveform estimation, which has a certain degree of interpretability. Another important contribution of this technology is the introduction of full-scale skip connections into the cGAN, which can improve the computational efficiency of full-scale feature maps by reducing network parameters. Furthermore, this invention introduces perceptual loss to fuse the jointly optimized objective function, capturing the similarity of low-frequency beats between the PPG and the output BP waveform, thereby avoiding distortion and loss of detail.

[0076] The attention mechanisms in the human visual system have inspired recent developments in attention modules within deep neural networks. Humans use a series of partial glances, selectively focusing on salient parts to better capture visual structures. Furthermore, as... Figure 8 As shown, a similar rising and falling rhythm was found between the matched PPG and BP waveforms. The peaks and troughs in the BP waveform directly correspond to the physiological values ​​of SBP and diastolic BP (DBP). To address this phenomenon, the present invention designs a simple attention module that improves the network's representation by focusing only on the time nodes with peaks and troughs and their corresponding values. Specifically, the present invention proposes a modified polarized self-attention (MPSA) mechanism, a spatial-temporal attention mechanism that has been proposed and introduced into cGAN models. By using the MPSA attention mechanism, the deep network proposed in this invention can focus on predicting the peaks and troughs of the BP waveform without extracting other prior features, achieving the AAMI clinical standard.

[0077] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a method for invasive blood pressure waveform estimation based on generative adversarial networks, comprising:

[0079] Step 100: Acquire the photoplethysmography (PPG) signal of the user under test;

[0080] Step 200: Input the photoplethysmography (PPG) signal of the user to be tested into the trained estimation monitoring model to obtain the blood pressure signal of the user to be tested;

[0081] The method for constructing the estimation monitoring model includes:

[0082] Step 201: Collect photoplethysmography (PPG) wave signals and continuous blood pressure waveform signals from the sample users;

[0083] Step 202: Perform data preprocessing on the photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users to obtain preprocessed data;

[0084] Step 203: Construct an estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography pulse waves; the estimation and monitoring model includes a signal waveform estimation model, an objective optimization function, an attention mechanism module, a generator, and a discriminator;

[0085] Step 204: Train the estimation monitoring model for invasive blood pressure waveform based on photoplethysmography pulse wave according to the preprocessed data to obtain the trained estimation monitoring model.

[0086] Preferably, step 202 includes:

[0087] The photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users were down-frequencyd to obtain a down-frequency signal.

[0088] The down-frequency signal is filtered using a finite impulse response filter with a preset cutoff frequency to obtain a filtered signal;

[0089] The filtered signal is phase-aligned to obtain an aligned signal;

[0090] The alignment signal is subjected to maximum-minimum normalization to obtain a normalized signal;

[0091] The normalized signal is windowed to obtain the preprocessed data.

[0092] Figure 2 The main flowchart of the method for estimating invasive blood pressure waveforms based on PPG using generative adversarial networks provided in this embodiment of the invention is as follows: Figure 2 As shown, the process 1 in this embodiment is as follows:

[0093] First, PPG signals from the human body are acquired, and blood pressure waveforms are simultaneously collected as reference signals for the model. The method of this invention utilizes the MIMIC III (Multi-parameter Intelligent Monitoring III) dataset, which was jointly released in 2016 by Beth Israel Deaconess Medical Center and Philips Healthcare. The MIMIC III waveform database contains 67,830 records from approximately 30,000 ICU patients, including ECG waveforms, BP waveforms, respiratory waveforms, and PPG waveforms.

[0094] Adversarial learning-based BP waveform estimation can generally be viewed as a sequence-to-sequence transformation. Therefore, this model needs to learn and obtain the mapping relationship between two sequences through a time-delay window of fixed length. Here, the original PPG signal space... and the corresponding target BP signal space Paired samples are given. The total number of sample points in segment M is s. x , dimension d x The total number of PPG signals and sample points is s y The BP signal has a one-to-one correspondence with the data. The PPG2BP-cGAN generator proposed in this invention learns the mapping from the PPG signal x and the random noise vector z to y:

[0095] G:{x,z}→y (1)

[0096] At the same time, a discriminator is trained synchronously to determine whether the BP waveform estimated by the generator is a false signal.

[0097] Preferably, step 203 includes:

[0098] A signal waveform estimation model based on conditional generative adversarial network is constructed to estimate the invasive blood pressure waveform based on photoplethysmography pulse wave signal, and the objective optimization function is designed.

[0099] An attention mechanism module is constructed to implement a method for estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network.

[0100] The architecture of the generator and the discriminator for a method of estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network is constructed.

[0101] Specifically, process 2 in this embodiment is as follows:

[0102] The PPG signal and reference blood pressure signal fragments obtained in process 1 are preprocessed and used as training data for a method and monitoring system based on generative adversarial network to estimate invasive blood pressure waveforms based on PPG. Figure 3The experimental design flowchart of the method of this invention is illustrated. The PPG and blood pressure signals collected in the database vary in length and are affected by outliers and baseline drift; therefore, signal preprocessing is required, such as... Figure 4 As shown in the boxes, after obtaining all possible PPG-BP paired data, the method of this invention performs signal resampling, signal denoising, phase alignment, signal normalization, and signal segmentation in the preprocessing pipeline to facilitate the training and inference of the deep model. The method of this invention down-frequencys all signals to 125 Hz to reduce computational burden while ensuring signal quality. Next, the method of this invention selects a 200th-order finite impulse response filter with a cutoff frequency of 0.5-8 Hz, which can effectively filter out power frequency noise, high-frequency noise generated by muscle movement, and low-frequency noise generated by baseline drift. The method of this invention performs phase alignment on the filtered signals to address the phase delay and distortion problems caused by the filter. Furthermore, considering that the amplitude of the PPG signal may be affected by ambient light intensity, blood oxygen concentration, or venous volume, the method of this invention performs max-min normalization on the PPG signal. Finally, the method of this invention sets a sliding window of length 512 and step size 384 to segment the PPG-BP paired signals, thereby ensuring the consistent length of the input model.

[0103] Preferably, the construction of the signal waveform estimation model based on the conditional generative adversarial network to estimate the invasive blood pressure waveform from the photoplethysmography (PPG) signal, and the design of the objective optimization function, includes:

[0104] Construct a GAN model; the GAN model consists of a parameterized generator and a parameterized discriminator; the generator and the discriminator perform adversarial learning based on the game theory principle of the min-max optimization task;

[0105] The adversarial loss is determined based on the generator and the discriminator;

[0106] The mean absolute error between the generated BP waveform and the actual BP waveform is calculated based on the adversarial loss.

[0107] Calculate the mean absolute error between the feature map of the reference BP waveform and the generated BP waveform;

[0108] The perceptual loss is determined based on the mean absolute error between the feature map of the reference BP waveform and the generated BP waveform.

[0109] Calculate the MSE between the reference blood pressure waveform and the estimated waveform at the systolic and diastolic points;

[0110] The signal waveform estimation model is constructed based on the GAN model, the adversarial loss, the mean absolute error between the generated BP waveform and the real BP waveform calculated by the adversarial loss, the mean absolute error between the feature map of the reference BP waveform and the generated BP waveform, the perceptual loss, and the MSE.

[0111] The objective optimization function is determined based on the signal waveform estimation model.

[0112] Furthermore, process 3 in this embodiment is specifically as follows:

[0113] A model framework based on PPG2BP-cGAN is constructed to estimate invasive blood pressure waveforms from PPG, and an objective optimization function is designed, such as... Figure 4 As shown. The model architecture includes a generator based on UNet3+ and a discriminator architecture based on PatchGAN. Generator G θ The task of (·) is to gradually improve the encoder-decoder module to transform the input PPG signal from the source domain. Transform to target domain (i.e., the estimated BP waveform). Adversarial discriminant D φ (·) is trained to distinguish real BP waveforms. and the generated BP waveform Provides a correction for cGAN loss and perceived loss Furthermore, the method of this invention defines the L2 loss between the reference blood pressure waveform and the estimated waveform at the systolic and diastolic points. and This will help the model accurately predict SBP and DBP values.

[0114] Specifically, a GAN model consists of a parameterized generator G. θ (·) and a parameterized discriminant D φ Composed of (·). Generator G θ (·) Receives a sample z from a prior noise distribution as input, and its task is to map it to the target space. To generate a model distribution In addition, discriminator D φ (·) is a binary classifier whose goal is to classify the original space x ~ p(χ) as the true value D when the original space is true. φ (x) = 1, when the generated sample When it is false, classify it as false. These two networks learn adversarially using game theory principles of min-max optimization tasks.

[0115]

[0116] Among the losses in combat The following formula is given:

[0117]

[0118] In the task of estimating BP waveforms based on PPG signals, the generator aims to map the source domain signal through a mapping function. Mapped to its corresponding real target signal In this case, the adversarial loss is given by the following formula:

[0119]

[0120] Traditional image-to-image conversion methods add pixel reconstruction losses, such as L1 loss, to adversarial loss, potentially resulting in generated images with a global structure similar to the real image. Inspired by these works, the method of this invention calculates the mean absolute error (MAE) between the generated and real BP waveforms, as follows:

[0121]

[0122] Pixel-based reconstruction loss can preserve a refined global structure, but it leads to signal distortion and loss of detail, often causing shifts in the generated BP values ​​and phase. Accurate prediction of blood pressure patterns is crucial in analysis because fluctuations and hysteresis in blood pressure values ​​significantly impact clinical diagnostic decisions. Therefore, perceptual loss is used to capture the differences between the estimated blood pressure signal and the high-frequency components of the reference blood pressure signal. Figure 9 As shown, the perceptual loss extracts intermediate feature representations through a discriminator network and a trainable feature extractor. The reference BP waveform y's i-th feature map and the generated BP waveform are shown. The MAE between them is calculated as follows:

[0123]

[0124] in, This refers to the discriminator D φ The feature map extracted from the i-th hidden layer of (·), L i and C i Let represent the length of the feature space and the number of channels, respectively. For a discriminator with N hidden layers, the perceptual loss can be calculated as:

[0125]

[0126] in This corresponds to the prior optimization parameters of the i-th layer.

[0127] Furthermore, the PPG2BP-cGAN model framework proposed in this invention also calculates the MSE between the reference blood pressure waveform and the estimated waveforms of the systolic and diastolic points to help the model accurately predict SBP and DBP values.

[0128]

[0129]

[0130] By putting them together, the final optimization problem for BP waveform estimation using the method of this invention can be given by the following formula:

[0131]

[0132] in

[0133]

[0134] Where λ, γ, δ, ξ ≥ 0 are the hyperparameters of these corresponding losses. The resulting network structure is as follows: Figure 4 As shown.

[0135] Preferably, the attention mechanism module for constructing a method for estimating invasive blood pressure waveforms based on photoplethysmography (PPG) signals using conditional generative adversarial networks includes:

[0136] Obtain the first feature map;

[0137] The first feature map input is transformed into a first query channel and a first key-value channel using convolutional layer kernels;

[0138] Perform matrix multiplication, convolution, and layer normalization on the first query channel and the first key value channel to obtain the first calculation result;

[0139] The spatial attention module is obtained based on the first calculation result and the feature channel;

[0140] Obtain the second feature map;

[0141] The second feature map input is transformed into a second query channel and a second key value channel using convolutional layer kernels;

[0142] The second query channel is enhanced using Softmax, and the enhanced information is then multiplied by the second key-value channel to obtain the second calculation result.

[0143] The time attention module is obtained based on the second calculation result and the time axis;

[0144] The temporal attention module and the spatial attention module are merged in parallel to obtain the attention mechanism module; the attention mechanism module is set in the skip layer and the connection layer.

[0145] Optionally, process 4 in this embodiment is specifically as follows:

[0146] This invention constructs an attention mechanism module based on PPG2BP-cGAN to implement a method for estimating invasive blood pressure waveforms from PPG. Fine-grained BP waveform estimation requires a model that can map long-range dependencies of high-resolution input / output features with low computational complexity, which can be used to estimate the semantic features of nonlinear signal values ​​and trends. Therefore, this invention proposes an MPSA mechanism, such as... Figure 5 and Figure 6 As shown, the MPSA module maintains high internal resolution along the spatial dimension (C / 2) and the temporal dimension (L), while folding the input feature map ([C×L]) along its corresponding dimensions. Furthermore, the method of this invention uses a Sigmoid construct operation to fit the output distribution of sample-wise regression; where Conv1d(1) represents a one-dimensional convolution operator with a kernel size of 1, GlobalPool represents a global pooling operator, and LayerNorm represents layer normalization. The MPSA mechanism uses polarization filtering to finely fuse important features in both spatial and temporal dimensions. Specifically, we use... Represents the feature map, the feature map of interest. Calculated using matrix multiplication:

[0147]

[0148] in This corresponds to the spatial attention diagram. This is a time-sensitive diagram.

[0149] First, spatial attention mechanisms focus on feature channels that contribute more to the representation of BP waveforms. For example... Figure 5 As shown, we use a convolutional layer W q and a convolution kernel W of size 1 k The input feature map H is transformed into query Q and key K, where the Q channel is fully compressed and the information is enhanced by high dynamic range based on softmax. Then, Q and K are matrix multiplied and followed by convolution W. z Layer normalization (LN) increases the dimension C / 2 of the channel to C:

[0150]

[0151] in ψ1 and ψ2 are two tensor reshaping operators, and κ1 is the prior parameter for these channel convolutions. SG and SM represent the Sigmoid and Softmax operators, respectively. Figure 6 As shown, we can obtain the output of spatial attention as follows: Among them ⊙ ch It is a multiplication operator for the characteristic channels.

[0152] Next, a temporal attention module is constructed. This module assigns weights to time segments of the feature map to focus on segments with higher information content in the PPG signal. Similar to the spatial attention module, the input features are first converted into Q and K terms using a single-kernel convolution, such as... Figure 6 As shown. In this case, the spatial dimension of the Q-values ​​is compressed using global pooling. In contrast, the spatial dimension of the K-values ​​remains at its original size. Next, the information of the Q-values ​​is enhanced using Softmax. Then, Q and K are multiplied by matrices. The following results can be obtained:

[0153] A tm (H)=SG[ψ3(SM(ψ1(GP(W q (H))))×ψ2(W k (H)))] (14

[0154] Where ψ3 is the tensor reshaping operator, and GP represents the global set operator. The output of time attention is... Among them ⊙ tm It is a multiplication operator on the time axis.

[0155] Finally, MPSA combines spatial attention and temporal attention. These two modules can be combined in parallel:

[0156]

[0157] The + operator is the element-wise addition operator. For example... Figure 7 and Figure 8 As shown, we add MPSA attention modules to the skip layer and the connection layer to focus on the local feature representations along the spatial and temporal axes in the PPG signal.

[0158] Preferably, the architecture of the generator and the discriminator for constructing a method for estimating invasive blood pressure waveforms based on photoplethysmography (PPG) signals using conditional generative adversarial networks includes:

[0159] Construct a generator based on UNet3+; the generator based on UNet3+ is used to combine high-level semantics from feature maps of different scales with low-level semantics through full-scale skip connections;

[0160] Construct a discriminator based on PatchGAN.

[0161] Specifically, process 5 in this embodiment is as follows:

[0162] This invention constructs a generator and discriminator architecture based on PPG2BP-cGAN to estimate invasive blood pressure waveforms from PPG. The method of this invention customizes a special generator and discriminator architecture according to the characteristics of PPG-translated BP signals. Specifically, a generator based on UNet3+ is constructed. Existing research has shown that skip connections in UNet can avoid the bottleneck layer limitations of low-level shared information in encoder-decoder networks. Therefore, as... Figure 7 As shown, the method of this invention selects UNet3+ as the backbone of the generator, which can further capture fine-grained details and coarse-grained semantics. Furthermore, this generator chooses dilated convolutions to reduce the overlap area between different convolutional kernels, thereby expanding the receptive field of the convolutional layers. UNet3+ directly combines high-level semantics from feature maps of different scales with low-level semantics through full-scale skip connections. Therefore, the feature map along the encoding direction of the i-th downsampling layer... It can be calculated as:

[0163]

[0164] in, This is the feature map of the encoder, where W(·) represents the feature aggregation mechanism using convolution-batch Norm-ReLU. and These represent upsampling and downsampling operations, respectively. Specifically, in... Figure 8 The second-layer decoder is given in the text. The detailed description explains that it is used to integrate aggregated feature maps across the entire scale. Each decoder consists of N scale connections, resulting in a total of 64 × N channels. Compared to UNet and UNet++, UNet3+ has the fewest parameters.

[0165]

[0166] Where K W d is the kernel size of the one-dimensional convolution, and d(·) represents the node depth. It is worth noting that this invention uses dilated convolution instead of a standard convolution kernel. Dilated convolution has been proven effective in electrocardiogram analysis, achieving a larger receptive field without increasing the number of covariates.

[0167] Specifically, a PatchGAN-based discriminator is constructed. Markov discriminators can maintain extremely high resolution and detail during style transfer. In BP waveform prediction, the patch-based receptive field is scaled to the patch length; that is, the discriminator convolves the entire image and uses the average response of 1×N samples to provide the final output. Details of this PatchGAN-based discriminator are shown in [the diagram]. Figure 4 The PPG2BP-cGAN model framework diagram.

[0168] Preferably, the step of training the estimation and monitoring model for invasive blood pressure waveform based on photoplethysmography (PPG) using the preprocessed data to obtain the trained estimation and monitoring model includes:

[0169] The preprocessed data is divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5.

[0170] The estimated monitoring model is trained using the training set and the validation set, and the signal waveform estimation model is validated using the test set to obtain the trained estimated detection model.

[0171] Preferably, it further includes:

[0172] During network training, the Adam optimizer is used to update the parameters; the learning rate is 1e-3, the weight decay is 0.999, and the momentum is 0.8.

[0173] The last two processes in this embodiment are as follows:

[0174] Step 6: The data preprocessed in Step 2 is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The training and validation sets are used to train the system models established in Steps 3, 4, and 5 and to select parameters. The test set is used to finally verify the generalization ability of the optimal system model saved in Step 3. Data splitting adopts an inter-individual splitting method to ensure that each data set does not contain data from the same patient, thereby avoiding information leakage. During network training, the Adam optimizer is used to update parameters with a learning rate of 1e-3, a weight decay of 0.999, and a momentum of 0.8.

[0175] Step 7: Acquire the user's PPG signal and send the signal waveform as input to the cuffless blood pressure waveform continuous estimation monitoring system. The system can automatically output the user's blood pressure signal, such as... Figure 9 As shown, the PPG signal data to be measured is input into the invasive blood pressure waveform estimation and monitoring system based on photoplethysmography (PPG) pulse wave established in processes 3, 4, and 5 for non-invasive estimation of the blood pressure waveform.

[0176] Corresponding to the above method, another aspect of the present invention provides a system for invasive blood pressure waveform estimation based on generative adversarial networks, comprising:

[0177] The first acquisition unit acquires the photoplethysmography (PPG) signal of the user under test.

[0178] The estimation unit is used to input the photoplethysmography (PPG) signal of the user under test into the trained estimation monitoring model to obtain the blood pressure signal of the user under test.

[0179] The second acquisition unit is used to acquire photoplethysmography (PPG) signals and continuous blood pressure waveform data of the sample users.

[0180] The preprocessing unit is used to preprocess the photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users to obtain preprocessed data.

[0181] The model building unit is used to build an estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography pulse waves; the estimation and monitoring model includes a signal waveform estimation model, an objective optimization function, an attention mechanism module, a generator, and a discriminator;

[0182] The training unit is used to train the estimation and monitoring model of invasive blood pressure waveform based on photoplethysmography pulse wave according to the preprocessed data, so as to obtain the trained estimation and monitoring model.

[0183] The beneficial effects of this invention are as follows:

[0184] 1. The input of this invention consists only of photoplethysmography (PPG) waves. Therefore, the acquisition circuit only needs to acquire PPG signals. Compared with the traditional method based on pulse wave propagation speed, it eliminates the step of acquiring electrocardiogram (ECG) signals and does not require excessive derivative calculations. This makes it easy to integrate into devices such as wristbands, eliminating the need for blood pressure measurement equipment such as cuffs, freeing the device from the constraints of cuffs, and making it more portable.

[0185] 2. The blood pressure estimation algorithm of the present invention can realize continuous blood pressure waveform estimation, long-term blood pressure monitoring, and can be used for blood pressure measurement in daily life without causing trauma or discomfort to the human body.

[0186] 3. The method of this invention uses PPG2BP-cGAN as the network backbone to extract a large amount of information from PPG signals. This model consists of a generator and a discriminator. The generator, based on UNet3+, contains a full-scale skip connection structure and integrates a novel modified polarization self-attention (MPSA) module based on a space-time attention mechanism. The discriminator is composed of PatchGAN, which increases the receptive field of the generated BP waveform through dilated convolutional layers, thereby improving the discrimination ability. This framework can promote the model to utilize more discriminative features during the learning process, and the input signal contains richer information, thus making the measured blood pressure results more stable and achieving higher prediction accuracy.

[0187] 4. The PPG2BP-cGAN model framework proposed in this invention demonstrates superior BP waveform prediction performance compared to state-of-the-art work on two independent datasets, exhibiting high robustness and stability. The proposed model is first pre-trained on a dataset containing 683 subjects, and then transferred to a public dataset for testing. Experimental tests on the MIMIC III (Multi-Parameter Intelligent Monitoring III) dataset show that the proposed method achieves estimation errors of 0.72±4.34 mmHg for systolic blood pressure (SBP) and 0.41±2.48 mmHg for diastolic blood pressure (DBP), respectively, which conforms to the internationally recognized American Association for Medical Instrumentation (AAMI) standards.

[0188] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0189] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for invasive blood pressure waveform estimation based on generative adversarial networks, characterized in that, include: Collect the photoplethysmography (PPG) signal of the user under test; The photoplethysmography (PPG) signal of the user under test is input into the trained estimation and monitoring model to obtain the blood pressure signal of the user under test. The method for constructing the estimation monitoring model includes: Collect photoplethysmography (PPG) pulse wave signals and continuous blood pressure waveform data from sample users; The photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users were preprocessed to obtain preprocessed data. An estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography pulse waves is constructed. The estimation and monitoring model includes a signal waveform estimation model, an objective optimization function, an attention mechanism module, a generator, and a discriminator. The preprocessed data is used to train the estimation and monitoring model of invasive blood pressure waveform based on photoplethysmography to obtain a trained estimation and monitoring model. The construction of the estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography includes: A signal waveform estimation model based on conditional generative adversarial network is constructed to estimate the invasive blood pressure waveform based on photoplethysmography pulse wave signal, and the objective optimization function is designed. An attention mechanism module is constructed to implement a method for estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network. The architecture of the generator and the discriminator for a method of estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network is constructed. The construction of the signal waveform estimation model based on the conditional generative adversarial network to estimate the invasive blood pressure waveform from the photoplethysmography (PPG) signal, and the design of the objective optimization function, include: Construct a GAN model; the GAN model consists of a parameterized generator and a parameterized discriminator; the generator and the discriminator perform adversarial learning based on the game theory principle of the min-max optimization task; The adversarial loss is determined based on the generator and the discriminator; The mean absolute error between the blood pressure waveform generated based on the aforementioned resistance loss and the actual blood pressure waveform; Calculate the mean absolute error between the feature map of the reference blood pressure waveform and the generated blood pressure waveform; The perceived loss is determined based on the mean absolute error between the feature map of the reference blood pressure waveform and the generated blood pressure waveform. Calculate the mean square error between the reference blood pressure waveform and the estimated waveforms of the systolic and diastolic points; The signal waveform estimation model is constructed based on the GAN model, the adversarial loss, the mean absolute error between the generated blood pressure waveform and the real blood pressure waveform calculated by the adversarial loss, the mean absolute error between the feature map of the reference blood pressure waveform and the generated blood pressure waveform, the perceptual loss, and the mean square error. The objective optimization function is determined based on the signal waveform estimation model.

2. The method for invasive blood pressure waveform estimation based on generative adversarial networks according to claim 1, characterized in that, The data preprocessing of the photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users yields preprocessed data, including: The photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users were down-frequencyd to obtain a down-frequency signal. The down-frequency signal is filtered using a finite impulse response filter with a preset cutoff frequency to obtain a filtered signal; The filtered signal is phase-aligned to obtain an aligned signal; The alignment signal is subjected to maximum-minimum normalization to obtain a normalized signal; The normalized signal is windowed to obtain the preprocessed data.

3. The method for invasive blood pressure waveform estimation based on generative adversarial networks according to claim 1, characterized in that, The attention mechanism module of the method for estimating invasive blood pressure waveforms based on photoplethysmography (PPG) signals using conditional generative adversarial networks includes: Obtain the first feature map; The first feature map input is transformed into a first query channel and a first key-value channel using convolutional layer kernels; Perform matrix multiplication, convolution, and layer normalization on the first query channel and the first key-value channel to obtain the first calculation result; The spatial attention module is obtained based on the first calculation result and the feature channel; Obtain the second feature map; The second feature map input is transformed into a second query channel and a second key value channel using convolutional layer kernels; The second query channel is enhanced using Softmax, and the enhanced information is then multiplied by the second key-value channel to obtain the second calculation result. The time attention module is obtained based on the second calculation result and the time axis; The temporal attention module and the spatial attention module are merged in parallel to obtain the attention mechanism module; the attention mechanism module is set in the jump layer and the connection layer.

4. The method for invasive blood pressure waveform estimation based on generative adversarial networks according to claim 1, characterized in that, The architecture of the generator and the discriminator for constructing a method based on conditional generative adversarial networks to estimate invasive blood pressure waveforms from photoplethysmography (PPG) signals includes: A generator based on UNet3+ is constructed; the generator based on UNet3+ is used to combine high-level semantics from feature maps of different scales with low-level semantics through full-scale skip connections; Construct a discriminator based on PatchGAN.

5. The method for invasive blood pressure waveform estimation based on generative adversarial networks according to claim 1, characterized in that, The step of training the estimation and monitoring model of invasive blood pressure waveform based on photoplethysmography (PPG) using the preprocessed data to obtain the trained estimation and monitoring model includes: The preprocessed data is divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.

5. The estimated monitoring model is trained using the training set and the validation set, and the signal waveform estimation model is validated using the test set to obtain a trained estimated detection model.

6. The method for invasive blood pressure waveform estimation based on generative adversarial networks according to claim 5, characterized in that, Also includes: During network training, the Adam optimizer is used to update the parameters; The initial learning rate is 1e-3, the weight decay is 0.999, and the momentum is 0.

8.

7. A system for invasive blood pressure waveform estimation based on generative adversarial networks, characterized in that, include: The first acquisition unit acquires the photoplethysmography (PPG) signal of the user under test. The estimation unit is used to input the photoplethysmography (PPG) signal of the user under test into the trained estimation monitoring model to obtain the blood pressure signal of the user under test. The second acquisition unit is used to acquire photoplethysmography (PPG) signals and continuous blood pressure waveform data of the sample users. The preprocessing unit is used to preprocess the photoplethysmography (PPG) wave signal and continuous blood pressure waveform signal data of the sample users to obtain preprocessed data. The model building unit is used to build an estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography pulse waves; the estimation and monitoring model includes a signal waveform estimation model, an objective optimization function, an attention mechanism module, a generator, and a discriminator; The training unit is used to train the estimation and monitoring model of invasive blood pressure waveform based on photoplethysmography pulse wave according to the preprocessed data, so as to obtain the trained estimation and monitoring model. The construction of the estimation and monitoring model for invasive blood pressure waveforms based on photoplethysmography includes: A signal waveform estimation model based on conditional generative adversarial network is constructed to estimate the invasive blood pressure waveform based on photoplethysmography pulse wave signal, and the objective optimization function is designed. An attention mechanism module is constructed to implement a method for estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network. The architecture of the generator and the discriminator for a method of estimating invasive blood pressure waveforms based on photoplethysmography pulse wave signals using a conditional generative adversarial network is constructed. The construction of the signal waveform estimation model based on the conditional generative adversarial network to estimate the invasive blood pressure waveform from the photoplethysmography (PPG) signal, and the design of the objective optimization function, include: Construct a GAN model; the GAN model consists of a parameterized generator and a parameterized discriminator; the generator and the discriminator perform adversarial learning based on the game theory principle of the min-max optimization task; The adversarial loss is determined based on the generator and the discriminator; The mean absolute error between the blood pressure waveform generated based on the aforementioned resistance loss and the actual blood pressure waveform; Calculate the mean absolute error between the feature map of the reference blood pressure waveform and the generated blood pressure waveform; The perceived loss is determined based on the mean absolute error between the feature map of the reference blood pressure waveform and the generated blood pressure waveform. Calculate the mean square error between the reference blood pressure waveform and the estimated waveforms of the systolic and diastolic points; The signal waveform estimation model is constructed based on the GAN model, the adversarial loss, the mean absolute error between the generated blood pressure waveform and the real blood pressure waveform calculated by the adversarial loss, the mean absolute error between the feature map of the reference blood pressure waveform and the generated blood pressure waveform, the perceptual loss, and the mean square error. The objective optimization function is determined based on the signal waveform estimation model.

Images