Feature decoupling model training method and device, and electronic equipment

Abstract

Embodiments of the present application provide a feature decoupling model training method and device and electronic equipment, relating to the technical field of wearable devices, the training method comprising: obtaining a plurality of multi-channel signal segments and corresponding clean PPG signal labels; constructing a feature decoupling model comprising a feature decoupling unit and a signal reconstruction unit; separating the features of the multi-channel signal segments into physiological features and motion features in a feature space through the feature decoupling unit, and minimizing the correlation between the two; reconstructing the physiological features through the signal reconstruction unit to obtain reconstructed PPG signals; calculating a reconstruction loss according to the reconstructed PPG signals and the clean PPG signal labels, and calculating a decoupling loss according to the correlation between the physiological features and the motion features; and updating the model parameters with the aim of minimizing the reconstruction loss and the decoupling loss. The present application achieves effective stripping of motion artifacts and high-fidelity PPG signal reconstruction, and improves the measurement accuracy of blood pressure parameters.

Description

Technical Field

[0001] This application belongs to the field of wearable device technology, specifically relating to a feature decoupling model training method, device, and electronic device. Background Technology

[0002] Cuffless continuous blood pressure measurement technology based on photoplethysmography (PPG) is severely contaminated by motion artifacts during user movement. These motion artifacts exhibit deep nonlinear coupling with physiological pulse wave components, with their frequency bands highly overlapping with the effective signal, causing blood pressure calculation models based on features such as pulse wave conduction time (PTT) to fail. Traditional filtering or fixed threshold methods struggle to effectively separate these nonlinear mixed artifacts.

[0003] In recent years, deep learning technology has been used for motion artifact removal and signal reconstruction, but purely data-driven methods have obvious drawbacks: the models lack physiological interpretability and have weak generalization ability across individuals and scenarios. In addition, existing methods often separate signal purification and blood pressure calculation into independent modules, failing to utilize the physical priors in blood pressure calculation to guide the purification process, which leads to easy distortion of key temporal features of the reconstructed signal.

[0004] How to achieve high-precision and high-reliability continuous blood pressure monitoring in dynamic motion scenarios, while taking into account the interpretability and generalization ability of the model, remains a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0005] In view of this, in order to overcome at least one aspect of the above problems, this application proposes a feature decoupling model training method, device and electronic device for achieving high-precision and high-reliability continuous blood pressure monitoring in dynamic motion scenarios, while taking into account the interpretability and generalization ability of the model.

[0006] The first aspect of this application proposes a method for training a feature decoupling model, the method comprising: Acquire multiple multi-channel signal segments and their corresponding clean PPG signal tags; A feature decoupling model is constructed, which includes a feature decoupling unit and a signal reconstruction unit; The multi-channel signal segment is input into the feature decoupling model. The feature decoupling unit separates the features of the multi-channel signal segment into physiological features and motion features in the feature space, and minimizes the correlation between the physiological features and the motion features. The physiological characteristics are reconstructed by the signal reconstruction unit to obtain the reconstructed PPG signal; The reconstruction loss is calculated based on the reconstructed PPG signal and the clean PPG signal label, and the decoupling loss is calculated based on the correlation between the physiological feature and the motion feature. The parameters of the feature decoupling model are updated with the goal of minimizing the reconstruction loss and the decoupling loss.

[0007] Furthermore, the feature decoupling model also includes: The encoding unit is used to perform feature encoding on the multi-channel signal segment, extract local spatiotemporal features of each channel and fuse them to obtain multimodal fusion features, and then input the multimodal fusion features into the feature decoupling unit; and / or, A temporal modeling unit is used to perform temporal dependency modeling on the multimodal fusion features to obtain temporal enhanced features, and then input the temporal enhanced features into the feature decoupling unit.

[0008] Furthermore, the feature decoupling unit decouples the physiological features from the motion features by segmenting the input features into physiological features and motion features in the feature dimension and minimizing the Frobenius norm between the physiological features and the motion features.

[0009] Furthermore, the signal reconstruction unit is a multi-layer transposed convolutional network, used to take the physiological features as input and map them back to the original PPG signal space through step-by-step upsampling to obtain the reconstructed PPG signal.

[0010] Furthermore, updating the parameters of the feature decoupling model includes a pre-training phase and an adversarial training phase; The pre-training phase uses a weighted sum of the reconstruction loss, frequency loss, and decoupling loss to update the parameters of the feature decoupling model, wherein the frequency loss is the difference in amplitude spectrum between the reconstructed PPG signal and the clean PPG signal label in the central frequency band of the short-time Fourier transform domain. The adversarial training phase introduces a discriminator, which updates the parameters of the feature decoupling model using a weighted sum of the reconstruction loss, adversarial loss, frequency loss, and decoupling loss. The adversarial loss is calculated based on the discriminator's discrimination result of the reconstructed PPG signal.

[0011] A second aspect of this application provides a feature decoupling model training apparatus, comprising: The data acquisition module is used to acquire multiple multi-channel signal segments and their corresponding clean PPG signal tags; The model building module is used to build a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit, wherein: The feature decoupling unit is used to separate the input features into physiological features and motion features in the feature space, and to minimize the correlation between the physiological features and the motion features. The signal reconstruction unit is used to reconstruct the signal of the physiological feature to obtain the reconstructed PPG signal; The loss calculation module is used to calculate the reconstruction loss based on the reconstructed PPG signal and the clean PPG signal label, and to calculate the decoupling loss based on the correlation between the physiological feature and the motion feature. The parameter update module is used to update the parameters of the feature decoupling model with the goal of minimizing the reconstruction loss and the decoupling loss.

[0012] A third aspect of this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the feature decoupling model training method described in the first aspect of this application.

[0013] A fourth aspect of this application provides a method for determining blood pressure parameters resistant to exercise interference, the method comprising: Acquire multi-channel signal segments; The multi-channel signal segment is input into a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit. The feature decoupling unit separates the features of the multi-channel signal segment into physiological features and motion features in the feature space, and minimizes the correlation between the physiological features and the motion features. The physiological characteristics are reconstructed by the signal reconstruction unit to obtain a pure PPG signal; Extract pulse wave conduction parameters from the pure PPG signal; The pulse wave conduction parameters are input into the blood pressure calculation model to determine the blood pressure parameters.

[0014] Furthermore, the feature decoupling model also includes: The encoding unit is used to perform feature encoding on the multi-channel signal segment, extract local spatiotemporal features of each channel and fuse them to obtain multimodal fusion features, and then input the multimodal fusion features into the feature decoupling unit; and / or, A temporal modeling unit is used to perform temporal dependency modeling on the multimodal fusion features to obtain temporal enhanced features, and then input the temporal enhanced features into the feature decoupling unit.

[0015] Furthermore, the feature decoupling unit decouples the physiological features from the motion features by segmenting the input features into physiological features and motion features in the feature dimension and minimizing the Frobenius norm between the physiological features and the motion features.

[0016] Furthermore, the signal reconstruction unit is a multi-layer transposed convolutional network, used to take the physiological features as input and map them back to the original PPG signal space through step-by-step upsampling to obtain the pure PPG signal.

[0017] Furthermore, the extraction of pulse wave conduction parameters includes: Detect the position of the R wave in the ECG signal; Detect the pin location of the pure PPG signal; The time difference between the R-wave position and the foot position in the same period is taken as the pulse wave conduction time.

[0018] Furthermore, after extracting the pulse wave conduction parameters from the pure PPG signal, the process further includes: Pulse wave morphological features are extracted from the pure PPG signal. These pulse wave morphological features include systolic area, diastolic area, area ratio, rise time, systolic time, diastolic time, pulse wave amplitude, dicrotic wave amplitude, reflected wave enhancement index, waveform width, waveform asymmetry, and one or more of SDNN and RMSSD. The pulse wave morphological features are input into the blood pressure calculation model.

[0019] Furthermore, the blood pressure calculation model undergoes parameter adaptation for a specific object; The object-specific parameter adaptation includes: Train a basic blood pressure calculation model on a general dataset; Collect calibration samples from specific objects; The prediction offset of the baseline blood pressure calculation model is calculated based on the calibration sample. The output of the basic blood pressure calculation model is corrected based on the offset.

[0020] Furthermore, the blood pressure calculation model establishes a mapping relationship between blood pressure and pulse wave conduction parameters based on the Moens-Korteweg equation.

[0021] The fifth aspect of this application provides a system for determining blood pressure parameters against motion interference, comprising: The data acquisition module is used to acquire multi-channel signal segments; The model building module is used to build a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit, wherein: The feature decoupling unit is used to separate the features of the multi-channel signal segment into physiological features and motion features in the feature space, and to minimize the correlation between the physiological features and the motion features. The signal reconstruction unit is used to reconstruct the signal of the physiological characteristics to obtain a pure PPG signal. The feature extraction module is used to extract pulse wave conduction parameters from the purified PPG signal; The blood pressure determination module is used to input the pulse wave conduction parameters into the blood pressure calculation model to determine the blood pressure parameters.

[0022] The sixth aspect of this application provides an electronic device including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the motion-interference-resistant blood pressure parameter determination method as described in the fourth aspect of this application.

[0023] The seventh aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the motion-interference-resistant blood pressure parameter determination method as described in the fourth aspect of this application.

[0024] This application provides a feature decoupling model training method, device, and electronic device. By using a feature decoupling mechanism, motion features and physiological features are separated in the latent space of a neural network, and a high-quality, pure PPG signal is reconstructed based solely on physiological features. This solves the technical problem that the nonlinear mixing of motion artifacts and physiological signals in dynamic motion scenarios makes it difficult to separate them, which leads to the failure of the blood pressure calculation model. It has the beneficial effect of achieving continuous blood pressure monitoring with clinical accuracy in daily activities such as walking and running, while also taking into account the interpretability and generalization ability of the model. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the feature decoupling model training method of this application; Figure 2 This is a flowchart illustrating the method for determining exercise-induced blood pressure parameters according to this application. Figure 3 This is a structural block diagram of the feature decoupling model training device of this application; Figure 4 This is a block diagram of the motion-resistant blood pressure parameter determination system of this application; Figure 5 This is a structural block diagram of the electronic device of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be further described in detail below with reference to specific examples and accompanying drawings.

[0028] It should be noted that all uses of "first" and "second" in the embodiments of this application are for the purpose of distinguishing two entities with the same name but different names or different parameters. It can be seen that "first" and "second" are only for the convenience of expression and should not be construed as limiting the embodiments of this application. Subsequent embodiments will not explain this in detail.

[0029] Firstly, such as Figure 1 As shown, this application provides a feature decoupling model training method, including the following steps: Step S101: Acquire multiple multi-channel signal segments and their corresponding clean PPG signal tags.

[0030] In this embodiment, the feature decoupling model training method relates to the field of wearable device technology and can be widely applied to wearable devices such as smartwatches and smart bracelets. This method can be executed by the feature decoupling model training device of this application embodiment, which can be configured in any electronic device to execute the feature decoupling model training method of this application embodiment. The electronic device can be a server or a terminal; no specific limitation is made here.

[0031] In wearable devices that utilize photoplethysmography (PPG) for cuffless continuous blood pressure monitoring, human movement introduces strong motion artifacts into the PPG signal. These motion artifacts exhibit deep nonlinear coupling with the physiological pulse wave component, with their frequency bands highly overlapping with the effective signal, making them difficult to separate effectively using traditional filtering or fixed threshold methods. To address these issues, this application employs a feature decoupling model to separate motion features from physiological features in the feature space, reconstructing a pure PPG signal based solely on the physiological features.

[0032] The multi-channel signal segments are obtained by preprocessing and windowing the synchronously acquired raw PPG signals and raw motion reference signals. Specifically, the acquisition system includes the following sensors, all of which are hardware-synchronized using the same master clock: the PPG sensor sampling frequency is... The triaxial accelerometer sampling frequency is The ECG sensor (optional) has a sampling frequency of [frequency value missing]. .

[0033] Reference blood pressure monitor: Upper arm oscillometric method, measure reference blood pressure value every 30-60 seconds (i.e., ... (Measure once).

[0034] Define a discrete time series and a time index. Corresponding physical time ,in For frequency; The contaminated PPG signal can be represented as: ;in, For an ideal clean PPG signal, For motion artifacts, Sensor noise for acquiring PPG signals; The accelerometer signal is: ; ECG signal: ;in, For an ideal clean ECG signal, Sensor noise for acquiring ECG signals; Reference blood pressure value: ; Measurement time .

[0035] The preprocessing process includes the following steps: First, due to the different sampling rates of each sensor, samples are uniformly resampled to... Using linear interpolation, the maximum time delay requirement between channels is... Secondly, a zero-phase Butterworth bandpass filter with a passband of [0.5, 8] Hz was used on the PPG signal to remove baseline drift (below 0.5 Hz) and high-frequency electromyographic noise (above 8 Hz).

[0036] The filter difference equation is: ; The accelerometer signal was low-pass filtered with a cutoff frequency of 20Hz to remove high-frequency vibration noise. The ECG signal was band-pass filtered with a passband of [0.5, 40]Hz.

[0037] Then, calculate the PPG signal quality index (SQI): ; Calculate the magnitude of acceleration: ; Accelerometer signal quality index: ; in This is the empirical variance threshold, only when... and During the specified time period, clean PPG signal tags were extracted.

[0038] Next, we will perform sliding window splitting, with the window parameters set as follows: window length sampling points, step size Number of sampling points, overlap rate The data matrix for the k-th window is: ; in Used as the starting index for the window.

[0039] Finally, channel-wise Z-score normalization is performed, and for each channel c, the global statistic for the entire training set is calculated: , ; After standardization: , ; Note that the training set is used during testing. To avoid data leaks.

[0040] For the k-th window, the clean PPG label is: The same process is performed. standardization.

[0041] Reference blood pressure values ​​are aligned using time interpolation. If the time is the center time of the window, then: ,in, .

[0042] Dataset divided by subjects: Training set 10 subjects, validation set 10 subjects, test set Each set of subjects was selected to ensure a similar proportion of different motion states (resting, walking, running, etc.). To improve the model's generalization ability, the training data was augmented, including temporal scaling. Additive noise Amplitude scaling .

[0043] To ensure time alignment of signals from all channels, the cross-correlation function is calculated to verify synchronization: ; Require the time delay corresponding to the maximum cross-correlation. One sampling point (20ms).

[0044] Calculate the segmented signal-to-noise ratio to evaluate signal quality: ; Only keep The window is used for training.

[0045] Step S102: Construct a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit.

[0046] In this embodiment, the feature decoupling model further includes an encoding unit and / or a temporal modeling unit. The input tensor is obtained from preprocessing. Where B is the batch size, L is the time step (e.g., 500), and C is the number of channels (5: PPG, ACC). x ACC y ACC z (ECG). Target tensor For a clean PPG signal. Learn parameterization functions. f θ : → .

[0047] Encoding unit: Extracts local spatiotemporal features for each input channel using an independent one-dimensional convolutional neural network. ; Where Conv1D has a kernel size of K=3, padding P=1, and step size S=1; BN is batch normalization; D c Let be the feature dimension of the c-th channel (set to 32). Multi-level feature extraction is performed using two convolutional layers.

[0048] The features of each channel are concatenated along the feature dimension as follows: Multi-level feature extraction (2 layers): ; Feature splicing: ; in, .

[0049] Timing modeling unit: First, add sinusoidal position encoding: ; ; Among them, , ; Post-encoded features .

[0050] Temporal dependency modeling is performed using a Transformer encoder, which includes a 4-layer multi-head self-attention mechanism. h =8, The calculation for each layer is as follows: ; ; MSA stands for Multi-Head Self-Attention, and FFN stands for Feedforward Network.

[0051] Step S103: Input the multi-channel signal segment into the feature decoupling model, and use the feature decoupling unit to separate the features of the multi-channel signal segment into physiological features and motion features in the feature space, and minimize the correlation between the physiological features and the motion features.

[0052] Specifically, the feature decoupling unit decouples the output of the last layer of the Transformer encoder. Segmentation along the feature dimension: ; in Physiological characteristics , For motion characteristics ( ).

[0053] The decoupling loss function is defined as the Frobenius norm between physiological features and motion features: ; in, This is the Frobenius norm, which is the square root of the sum of the squares of the absolute values ​​of all elements of the matrix. Minimizing this loss forces the model to encode physiological information into... H physio In the process, motion information is encoded into H motion In this way, the two are decoupled.

[0054] Step S104: The physiological feature is reconstructed by the signal reconstruction unit to obtain the reconstructed PPG signal.

[0055] Specifically, the signal reconstruction unit uses a one-dimensional transposed convolutional network to reconstruct physiological features. H physio Mapping back to PPG signal space: Level 1: 80 input channels, 64 output channels, core size 3, step size 1, padding 1; Level 2: Input channels 64, output channels 32, core size 3, step size 1, padding 1; Layer 3 (Output Layer): Output channel 1 uses linear activation.

[0056] Motion characteristics H motion It does not participate in signal reconstruction, therefore the reconstructed PPG signal does not contain motion components.

[0057] Step S105: Calculate the reconstruction loss based on the reconstructed PPG signal and the clean PPG signal label, and calculate the decoupling loss based on the correlation between the physiological feature and the motion feature.

[0058] Specifically, the reconstruction loss (mean squared error): ; Decoupling loss: ; Step S106: Update the parameters of the feature decoupling model with the goal of minimizing the reconstruction loss and the decoupling loss.

[0059] Specifically, the total loss is the weighted sum of all losses: ; in, λ recon =1.0, λ adv =0.1, λ freq =0.5, λ dis =0.01, λ gp =10.0 .

[0060] Adversarial Loss (WGAN-GP): The discriminator is a 5-layer one-dimensional CNN, and the final output is a scalar. Discriminator loss: ; Among them, the gradient penalty term is: ; Generator-versus-loss: ; Frequency loss: Short-time Fourier transform (STFT) is used, with a window length of [missing information]. N win =256 Step length R=128 ,Hanming Window w[n]: ; Heart rate band loss (0.5~4.0Hz): ; in, , .

[0061] The parameters of the feature decoupling model are updated in two phases: pre-training and adversarial training. The pre-training phase (10 epochs) uses only a weighted sum of the reconstruction loss, frequency loss, and decoupling loss, with a learning rate of 10. -³. In the adversarial training phase (90 rounds), a discriminator is introduced and trained using the full loss function, with a purifier learning rate of 10. -4 The discriminator learning rate is 10. -4 The Adam optimizer is used; purifier optimizer parameters... β 1 =0.9、β 2 =0.999 Discriminator optimizer parameters β 1 =0.5、β 2 =0.9 During training, the discriminator is trained 5 times in each iteration, and then the purifier is trained once.

[0062] Secondly, such as Figure 2 As shown, this application provides a method for determining blood pressure parameters against exercise interference, comprising the following steps: Step S201: Acquire multi-channel signal segments.

[0063] The acquisition method of multi-channel signal segments is the same as the preprocessing method in step S101 of the first embodiment, including synchronous sampling, filtering, window segmentation and normalization processing, which will not be described again here.

[0064] Step S202: Input the multi-channel signal segment into the feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit.

[0065] In this embodiment, the feature decoupling model is the feature decoupling model trained in the first embodiment. This model has been described in detail in the first embodiment and will not be repeated here. The feature decoupling model further includes an encoding unit and / or a temporal modeling unit. The encoding unit is used to perform feature encoding on the multi-channel signal segment, extract the local spatiotemporal features of each channel and fuse them to obtain multimodal fusion features, and then input the multimodal fusion features into the feature decoupling unit. The temporal modeling unit is used to perform temporal dependency modeling on the multimodal fusion features to obtain temporal enhancement features, and then input the temporal enhancement features into the feature decoupling unit.

[0066] Step S203: The feature decoupling unit separates the features of the multi-channel signal segment into physiological features and motion features in the feature space, and minimizes the correlation between the physiological features and the motion features.

[0067] Specifically, the feature decoupling unit decouples the physiological features from the motion features by dividing the input features into physiological features and motion features in the feature dimension and minimizing the Frobenius norm between the physiological features and the motion features.

[0068] Step S204: The physiological characteristics are reconstructed by the signal reconstruction unit to obtain a pure PPG signal.

[0069] Specifically, the signal reconstruction unit is a multi-layer transposed convolutional network, which takes the physiological features as input and maps them back to the original PPG signal space through step-by-step upsampling to obtain the pure PPG signal.

[0070] Step S205: Extract pulse wave conduction parameters from the purified PPG signal.

[0071] Specifically, extracting pulse wave conduction parameters includes the following sub-steps: First, the R-wave position of the ECG signal is detected. An improved Pan-Tompkins algorithm is employed, including bandpass filtering (5~15Hz), differentiation, squaring, moving window integration (window length approximately 150ms), and adaptive threshold detection. The differentiation formula is as follows: ; The adaptive threshold is set as: signal threshold noise threshold Where M is the most recent The maximum signal value at each point. Detection condition is... RR interval greater than 0.3f s And in The interval contains a local maximum value. The detected R-wave location is denoted as... The position of the R wave is refined on the original ECG signal.

[0072] Next, the pin locations of the pure PPG signal are detected. A combination of the first-order difference method and the minimum value search method is used. First-order difference... The foot point is located at the starting point of the rising segment of PPG, satisfying... in, and Simultaneously in each heartbeat cycle The minimum point is searched within the interior. Finally, the foot position is determined by combining the two methods. .

[0073] Next, detect the main peak of PPG: ; Detection of PPG diphtheria wave notch: Second-order difference method ; Cutting point satisfy , , ,and It is a local minimum.

[0074] Detection of PPG diphtheria peak: ; Finally, the time difference between the R-wave position and the foot position in the same period is taken as the pulse wave conduction time: (Second); Outliers are removed (using the 3σ criterion), then median filtering is performed, and finally a weighted average of the most recent K=10 heartbeats is taken: ,in, , .

[0075] After extracting the pulse wave conduction parameters, the process also includes extracting pulse wave morphological features from the purified PPG signal. For the first... j One heartbeat cycle Extract the following features: Area during contraction: ; Diastolic area: ; Area ratio: ; Ascent time: ; Contraction time: ; Diastolic time: ; Pulse amplitude: ; Diptacardiogram amplitude: ; Reflection enhancement index: ; Waveform width: ( and waveform asymmetry: ASY j .

[0076] Heart rate variability time-domain index: SDNN: ; RMSSD: ; The pulse wave morphological features are input into the blood pressure calculation model.

[0077] Step S206: Input the pulse wave conduction parameters into the blood pressure calculation model to determine the blood pressure parameters.

[0078] Specifically, the blood pressure calculation model establishes a mapping relationship between blood pressure and pulse wave transmission parameters based on the Moens-Korteweg equation. According to the Moens-Korteweg equation: ; Where PWV is pulse wave velocity, E is Young's modulus of the blood vessel wall, h is blood vessel wall thickness, d is blood vessel diameter, ρ is blood density, and L is pulse wave propagation distance.

[0079] Combining Hughes' relationship (P represents blood pressure), derive the blood pressure-PTT relationship: ; Practical linearization model: ; ; ; Blood pressure calculation can also be performed using a multiple linear regression model. (Constructing feature vectors) ; Model: , ; in, This is the weight vector.

[0080] Blood pressure calculation models can also use support vector regression models. ε- Insensitive loss function and radial basis kernel function: ; The blood pressure calculation model undergoes parameter adaptation tailored to specific user groups. A two-stage calibration strategy is employed: the first stage trains a general model f0(x) on a large dataset; the second stage collects calibration data from new users (collection...). Group calibration data ), calculate the linear offset: ; ; After calibration: ; ; As an alternative, Bayesian adaptive calibration can be used. This assumes a prior distribution of blood pressure. Measurement model Bayesian update: Kalman gain: ; Posterior mean: ; Posterior covariance: ; in K For Kalman gain, μ 1 and Σ 1 represents the posterior mean and posterior covariance, respectively.

[0081] The blood pressure estimation results are smoothed using a Kalman filter. State vector. The state transition matrix F and the observation matrix H are respectively: ; ; Kalman filter update equation: .

[0082] To verify the technical effectiveness of the feature decoupling model training method and the method for determining blood pressure parameters against motion interference provided in this application, this embodiment uses the following evaluation indicators and testing methods to verify the system performance.

[0083] The accuracy of blood pressure parameter determination was evaluated using the following indicators: Mean error (ME): ; Standard deviation (SD): , Root Mean Square Error (RMSE): ; Bland-Altman analysis: Consistency bound is The percentage error is .

[0084] According to ANSI / AAMI / ISO 81060-2:2018 standard, the absolute value of the mean error of blood pressure measuring equipment should be mmHg, standard deviation should mmHg, sample size should Each subject must have at least three measurements.

[0085] The cumulative error percentage for the British Hypertension Society (BHS) standard grading is: A: 60%<5 mmHg, 85%<10 mmHg, 95%<15 mmHg; B: 50%<5 mmHg, 75%<10 mmHg, 90%<15 mmHg; C: 40%<5 mmHg, 65%<10 mmHg, 85%<15 mmHg; To verify the stability of this application under different noise conditions, the accuracy of feature point detection was tested under different signal-to-noise ratios. Gaussian noise was added to the clean PPG signal: ; Test PTT detection error: ; The system integrates three steps: preprocessing, feature decoupling and model purification, and temporal feature extraction and blood pressure calculation, forming a complete end-to-end system. The input is a raw, noisy PPG signal, which is purified, then features are extracted and blood pressure is calculated. Finally, the result is compared with the gold standard. The experiment uses a dataset containing multiple subjects for validation, covering various exercise states including resting, walking (low speed 3km / h, medium speed 5km / h), and running (8km / h).

[0086] Experimental results show that under different motion states, the original PPG signal is severely contaminated by motion artifacts, resulting in significant waveform distortion. After processing by the feature decoupling model of this application, the reconstructed PPG signal waveform is highly consistent with the clean reference signal, and motion artifacts are effectively removed. The signal-to-noise ratio (SNR) improvement ΔSNR reaches over 15 dB, and the correlation coefficient between the reconstructed signal and the clean reference signal is significantly improved. ρ>0.95 .

[0087] The average error of systolic blood pressure estimation was assessed according to AAMI / ISO 81060-2:2018 standard |ME SBP |<5mmHg, standard deviation SD SBP <8 mmHg; Mean error of diastolic blood pressure estimation | ME DBP |<5mmHg, standard deviation SD DBP <8 mmHg. According to BHS standards, this method meets Class A standards, meaning 60% of the measurement error is less than 5 mmHg, 85% is less than 10 mmHg, and 95% is less than 15 mmHg. This indicates that this application can achieve clinically accurate continuous blood pressure monitoring during daily activities such as walking and running.

[0088] Thirdly, such as Figure 3 As shown, this application provides a feature decoupling model training device 300, comprising: Data acquisition module 301 is used to acquire multiple multi-channel signal segments and their corresponding clean PPG signal tags; Model building module 302 is used to build a feature decoupling model, which includes a feature decoupling unit 303 and a signal reconstruction unit 304, wherein: The feature decoupling unit 303 is used to separate the input features into physiological features and motion features in the feature space, and minimize the correlation between the physiological features and the motion features; The signal reconstruction unit 304 is used to reconstruct the signal of the physiological feature to obtain the reconstructed PPG signal; The loss calculation module 305 is used to calculate the reconstruction loss based on the reconstructed PPG signal and the clean PPG signal label, and to calculate the decoupling loss based on the correlation between the physiological feature and the motion feature. The parameter update module 306 is used to update the parameters of the feature decoupling model with the goal of minimizing the reconstruction loss and the decoupling loss.

[0089] The feature decoupling model training device 300 provided in this application can implement all the processes implemented by the method of the first embodiment described above, and can achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0090] Fourthly, such as Figure 4 As shown, this application provides a motion-interference-resistant blood pressure parameter determination system 400, comprising: Data acquisition module 401 is used to acquire multi-channel signal segments; Model building module 402 is used to build a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit, wherein: The feature decoupling unit 403 is used to separate the features of the multi-channel signal segment into physiological features and motion features in the feature space, and minimize the correlation between the physiological features and the motion features; The signal reconstruction unit 404 is used to reconstruct the signal of the physiological characteristics to obtain a pure PPG signal. Feature extraction module 405 is used to extract pulse wave conduction parameters from the pure PPG signal; The blood pressure determination module 406 is used to input the pulse wave conduction parameters into the blood pressure calculation model to determine the blood pressure parameters.

[0091] The motion-interference-resistant blood pressure parameter determination system 400 provided in this application can realize all the processes implemented by the method of the second embodiment described above, and can achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0092] Fifthly, according to embodiments of this application, this application also provides an electronic device, a readable storage medium, and a computer program product.

[0093] Figure 5A schematic block diagram of an example electronic device 500 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present application described and / or claimed herein.

[0094] like Figure 5 As shown, device 500 includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 502 or a computer program loaded into random access memory (RAM) 503 from storage unit 508. RAM 503 can also store various programs and data required for the operation of device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.

[0095] Multiple components in device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0096] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as the feature decoupling model training method or the motion-disruption-resistant blood pressure parameter determination method. For example, in some embodiments, the feature decoupling model training method or the motion-disruption-resistant blood pressure parameter determination method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the feature decoupling model training method or the motion-disruption-resistant blood pressure parameter determination method described above can be performed. Alternatively, in other embodiments, the computing unit 501 may be configured by any other suitable method (e.g., by means of firmware) to perform a feature decoupling model training method or a motion-disruption-resistant blood pressure parameter determination method.

[0097] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0098] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0099] Within the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0100] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0101] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), the Internet, and blockchain networks.

[0102] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service ecosystem that addresses the management difficulties and weak business scalability inherent in traditional physical hosts and VPS services. Servers can also be servers for distributed systems or servers integrated with blockchain technology.

[0103] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0104] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

[0105] This application embodiment also provides a readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described functions. Figure 1 Any step in the corresponding method embodiment can achieve the same technical effect, and will not be repeated here to avoid repetition.

[0106] The computer-readable storage medium of this application embodiment can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate or transmit a program for use by or in connection with an instruction execution system, apparatus or device.

[0107] The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0108] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or terminal. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0109] The above description represents the preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for training a feature decoupling model, characterized in that, The method includes: Acquire multiple multi-channel signal segments and their corresponding clean PPG signal tags; A feature decoupling model is constructed, which includes a feature decoupling unit and a signal reconstruction unit; The multi-channel signal segment is input into the feature decoupling model. The feature decoupling unit separates the features of the multi-channel signal segment into physiological features and motion features in the feature space, and minimizes the correlation between the physiological features and the motion features. The physiological characteristics are reconstructed by the signal reconstruction unit to obtain the reconstructed PPG signal; The reconstruction loss is calculated based on the reconstructed PPG signal and the clean PPG signal label, and the decoupling loss is calculated based on the correlation between the physiological feature and the motion feature. The parameters of the feature decoupling model are updated with the goal of minimizing the reconstruction loss and the decoupling loss. The feature decoupling model further includes: The encoding unit is used to perform feature encoding on the multi-channel signal segment, extract the local spatiotemporal features of each channel and fuse them to obtain multimodal fusion features, and then input the multimodal fusion features into the feature decoupling unit; A temporal modeling unit is used to perform temporal dependency modeling on the multimodal fusion features to obtain temporal enhanced features, and then input the temporal enhanced features into the feature decoupling unit; The updating of the parameters of the feature decoupling model includes a pre-training phase and an adversarial training phase; The pre-training phase uses a weighted sum of the reconstruction loss, frequency loss, and decoupling loss to update the parameters of the feature decoupling model, wherein the frequency loss is the difference in amplitude spectrum between the reconstructed PPG signal and the clean PPG signal label in the central frequency band of the short-time Fourier transform domain. The adversarial training phase introduces a discriminator, which updates the parameters of the feature decoupling model using a weighted sum of the reconstruction loss, adversarial loss, frequency loss, and decoupling loss. The adversarial loss is calculated based on the discriminator's discrimination result of the reconstructed PPG signal.

2. The feature decoupling model training method according to claim 1, characterized in that, The feature decoupling unit decouples the physiological features from the motion features by dividing the input features into physiological features and motion features in the feature dimension and minimizing the Frobenius norm between the physiological features and the motion features.

3. The feature decoupling model training method according to claim 1, characterized in that, The signal reconstruction unit is a multi-layer transposed convolutional network used to take the physiological features as input and map them back to the original PPG signal space through step-by-step upsampling to obtain the reconstructed PPG signal.

4. A feature decoupling model training device, characterized in that, include: The data acquisition module is used to acquire multiple multi-channel signal segments and their corresponding clean PPG signal tags; The model building module is used to build a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit, wherein: The feature decoupling unit is used to separate the input features into physiological features and motion features in the feature space, and to minimize the correlation between the physiological features and the motion features. The signal reconstruction unit is used to reconstruct the signal of the physiological feature to obtain the reconstructed PPG signal; The loss calculation module is used to calculate the reconstruction loss based on the reconstructed PPG signal and the clean PPG signal label, and to calculate the decoupling loss based on the correlation between the physiological feature and the motion feature. The parameter update module is used to update the parameters of the feature decoupling model with the goal of minimizing the reconstruction loss and the decoupling loss. The feature decoupling model further includes: The encoding unit is used to perform feature encoding on the multi-channel signal segment, extract the local spatiotemporal features of each channel and fuse them to obtain multimodal fusion features, and then input the multimodal fusion features into the feature decoupling unit; A temporal modeling unit is used to perform temporal dependency modeling on the multimodal fusion features to obtain temporal enhanced features, and then input the temporal enhanced features into the feature decoupling unit; The updating of the parameters of the feature decoupling model includes a pre-training phase and an adversarial training phase; The pre-training phase uses a weighted sum of the reconstruction loss, frequency loss, and decoupling loss to update the parameters of the feature decoupling model, wherein the frequency loss is the difference in amplitude spectrum between the reconstructed PPG signal and the clean PPG signal label in the central frequency band of the short-time Fourier transform domain. The adversarial training phase introduces a discriminator, which updates the parameters of the feature decoupling model using a weighted sum of the reconstruction loss, adversarial loss, frequency loss, and decoupling loss. The adversarial loss is calculated based on the discriminator's discrimination result of the reconstructed PPG signal.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the feature decoupling model training method as described in any one of claims 1 to 3.

6. A method for determining blood pressure parameters resistant to motion interference, characterized in that, The method includes: Acquire multi-channel signal segments; The multi-channel signal segment is input into a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit. The feature decoupling unit separates the features of the multi-channel signal segment into physiological features and motion features in the feature space, and minimizes the correlation between the physiological features and the motion features. The physiological characteristics are reconstructed by the signal reconstruction unit to obtain a pure PPG signal; Extract pulse wave conduction parameters from the pure PPG signal; The pulse wave conduction parameters are input into the blood pressure calculation model to determine the blood pressure parameters; The feature decoupling model further includes: The encoding unit is used to perform feature encoding on the multi-channel signal segment, extract the local spatiotemporal features of each channel and fuse them to obtain multimodal fusion features, and then input the multimodal fusion features into the feature decoupling unit; A temporal modeling unit is used to perform temporal dependency modeling on the multimodal fusion features to obtain temporal enhanced features, and then input the temporal enhanced features into the feature decoupling unit; The feature decoupling model is obtained through a pre-training phase and an adversarial training phase. The pre-training phase uses a weighted sum of reconstruction loss, frequency loss, and decoupling loss to update the parameters of the feature decoupling model, wherein the frequency loss is the difference in amplitude spectrum between the reconstructed PPG signal and the clean PPG signal label in the central frequency band of the short-time Fourier transform domain. The adversarial training phase introduces a discriminator, which updates the parameters of the feature decoupling model using a weighted sum of the reconstruction loss, adversarial loss, frequency loss, and decoupling loss. The adversarial loss is calculated based on the discriminator's discrimination result of the reconstructed PPG signal.

7. The method for determining blood pressure parameters against exercise interference according to claim 6, characterized in that, The feature decoupling unit decouples the physiological features from the motion features by dividing the input features into physiological features and motion features in the feature dimension and minimizing the Frobenius norm between the physiological features and the motion features.

8. The method for determining motion-interference-resistant blood pressure parameters according to claim 6, characterized in that, The signal reconstruction unit is a multi-layer transposed convolutional network, which takes the physiological features as input and maps them back to the original PPG signal space through step-by-step upsampling to obtain the pure PPG signal.

9. The method for determining blood pressure parameters against motion interference according to claim 6, characterized in that, The extraction of pulse wave conduction parameters includes: Detect the R-wave position of the ECG signal; Detect the pin location of the pure PPG signal; The time difference between the R-wave position and the foot position in the same period is taken as the pulse wave conduction time.

10. The method for determining motion-interference-resistant blood pressure parameters according to claim 6, characterized in that, After extracting the pulse wave conduction parameters from the pure PPG signal, the process further includes: Pulse wave morphological features are extracted from the pure PPG signal. These pulse wave morphological features include systolic area, diastolic area, area ratio, rise time, systolic time, diastolic time, pulse wave amplitude, dicrotic wave amplitude, reflected wave enhancement index, waveform width, waveform asymmetry, and one or more of SDNN and RMSSD. The pulse wave morphological features are input into the blood pressure calculation model.

11. The method for determining motion-interference-resistant blood pressure parameters according to claim 6, characterized in that, The blood pressure calculation model is subject to object-specific parameter adaptation. The object-specific parameter adaptation includes: Train a basic blood pressure calculation model on a general dataset; Collect calibration samples from specific objects; The prediction offset of the baseline blood pressure calculation model is calculated based on the calibration sample. The output of the basic blood pressure calculation model is corrected based on the offset.

12. The method for determining motion-interference-resistant blood pressure parameters according to claim 6, characterized in that, The blood pressure calculation model establishes a mapping relationship between blood pressure and pulse wave conduction parameters based on the Moens-Korteweg equation.

13. A system for determining blood pressure parameters resistant to motion interference, characterized in that, include: The data acquisition module is used to acquire multi-channel signal segments; The model building module is used to build a feature decoupling model, which includes a feature decoupling unit and a signal reconstruction unit, wherein: The feature decoupling unit is used to separate the features of the multi-channel signal segment into physiological features and motion features in the feature space, and to minimize the correlation between the physiological features and the motion features. The signal reconstruction unit is used to reconstruct the signal of the physiological characteristics to obtain a pure PPG signal. The feature extraction module is used to extract pulse wave conduction parameters from the purified PPG signal; The blood pressure determination module is used to input the pulse wave conduction parameters into the blood pressure calculation model to determine the blood pressure parameters; The feature decoupling model further includes: The encoding unit is used to perform feature encoding on the multi-channel signal segment, extract the local spatiotemporal features of each channel and fuse them to obtain multimodal fusion features, and then input the multimodal fusion features into the feature decoupling unit; A temporal modeling unit is used to perform temporal dependency modeling on the multimodal fusion features to obtain temporal enhanced features, and then input the temporal enhanced features into the feature decoupling unit; The feature decoupling model is obtained through a pre-training phase and an adversarial training phase. The pre-training phase uses a weighted sum of reconstruction loss, frequency loss, and decoupling loss to update the parameters of the feature decoupling model, wherein the frequency loss is the difference in amplitude spectrum between the reconstructed PPG signal and the clean PPG signal label in the central frequency band of the short-time Fourier transform domain. The adversarial training phase introduces a discriminator, which updates the parameters of the feature decoupling model using a weighted sum of the reconstruction loss, adversarial loss, frequency loss, and decoupling loss. The adversarial loss is calculated based on the discriminator's discrimination result of the reconstructed PPG signal.

14. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the motion-interference-resistant blood pressure parameter determination method as described in any one of claims 6 to 12.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the motion-interference-resistant blood pressure parameter determination method as described in any one of claims 6 to 12.

Images