A cross-subject blood pressure estimation method based on physiological constraints and test source adaptation
Patent Information
- Application Number
- CN202611015926.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-09-29
AI Technical Summary
如果仅从纯数据驱动的角度进行迁移学习，而缺乏生理先验的引导，模型在测试阶段的在线自适应过程中极易因噪声干扰而陷入局部最优，导致预测结果偏离合理的生理范围
[0027]本发明从解决毫米波雷达血压估计中跨模态泛化难与受试者个体生理差异大的技术痛点出发，首先通过预训练机制内化了光电容积脉搏波信号的大规模血流动力学先验知识，使得模型在未接触目标雷达数据前即具备了捕捉通用心血管特征的基础能力，随后在推理阶段巧妙利用测试源自适应架构，仅依据当前受试者的无标签原始信号即可实现轻量化适配器参数的实时动态校准，极大地降低了系统在实际落地部署时的校准成本与数据标注依赖，同时引入多维度的生理规律约束与动态统计区间感知机制，确保了模型在对抗性扰动迭代过程中始终遵循人体循环系统的物理逻辑，有效抑制了非接触式测量环境下的随机噪声干扰，最终实现了在完全跨受试者场景下对收缩压与舒张压的高精度、高鲁棒性连续监测。
Smart Images
Figure CN122827641A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of physiological parameter monitoring, and more specifically to a cross-subject blood pressure estimation method based on physiological constraints and adaptive test sources. Background Technology
[0002] Cardiovascular disease has become one of the leading chronic diseases threatening human health worldwide. As a core physiological parameter reflecting cardiovascular function, blood pressure is of irreplaceable importance for accurate and continuous monitoring in clinical scenarios such as hypertension early warning and heart failure management, as well as in daily health monitoring.
[0003] Current mainstream blood pressure monitoring technologies can be divided into two categories: one is cuff-based intermittent measurement technology, which, while convenient to operate, cannot achieve continuous dynamic monitoring, and frequent cuff compression can easily cause discomfort in the user's arm, making it difficult to meet long-term monitoring needs; the other is continuous monitoring technology based on physiological signals. This technology achieves dynamic estimation of blood pressure parameters by analyzing the intrinsic correlation between physiological signals such as electrocardiogram (ECG), photoplethysmography (PPG), and arterial pulse waves and blood pressure. Among these, PPG signals have become the mainstream choice for consumer-grade health monitoring devices due to their convenient acquisition process and non-invasiveness. Existing technologies can already achieve preliminary blood pressure estimation by extracting the time-domain (e.g., pulse wave transit time) and frequency-domain characteristics of PPG signals.
[0004] However, traditional feature engineering methods heavily rely on the accuracy of prior knowledge, making it difficult to fully capture the complex nonlinear dynamic information in physiological signals. With the rapid development of deep learning technology, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and various fusion models are widely used in blood pressure prediction modeling due to their superior feature extraction capabilities. Deep learning models can mine deeper semantic representations than traditional handcrafted features through end-to-end learning on large-scale physiological signal datasets, significantly improving prediction accuracy within specific subject groups and making cuffless, real-time continuous blood pressure estimation possible at the algorithmic level.
[0005] Despite significant progress in deep learning-based blood pressure estimation, it still faces a serious challenge in practical applications due to its "cross-subject" generalization ability. The human physiological system is highly heterogeneous; different subjects exhibit varying vascular elasticity, peripheral resistance, blood viscosity, and cardiovascular physiological states. This leads to significant differences in pulse wave morphology even at the same blood pressure level. Conventional deep learning models, after being trained on source domain data, often exhibit "overfitting to known individuals"—a phenomenon where predictive performance declines drastically when faced with entirely new, unseen target subjects (the target domain) due to the significant feature distribution shift between the source and target domains.
[0006] To address this challenge, researchers have begun to explore techniques such as domain adaptation to bridge the distribution gap among subjects. However, most existing adaptive methods often overlook the essential properties of blood pressure as a physiological parameter. Blood pressure fluctuations are not random but follow strict physical laws and anatomical constraints of the circulatory system, such as the relative range of systolic and diastolic blood pressure and the continuity of signal gradients. If transfer learning is performed solely from a data-driven perspective without physiological prior guidance, the model is highly susceptible to getting trapped in local optima due to noise interference during online adaptation in the testing phase, leading to predictions that deviate from the reasonable physiological range. Therefore, how to organically combine the powerful self-learning capabilities of deep learning with physiological constraints and achieve online adaptive adjustment for individual target subjects has become a core technical challenge that urgently needs to be solved in the field of non-destructive continuous blood pressure monitoring. Summary of the Invention
[0007] The purpose of this invention is to provide a cross-subject blood pressure estimation method based on physiological constraints and test source adaptation. It first pre-trains a deep neural network model using a large-scale photoplethysmography (PPG) dataset, constructing a source domain baseline model capable of capturing general cardiovascular dynamics characteristics through end-to-end learning. Subsequently, for an unlabeled millimeter-wave radar target domain dataset, a test source adaptation (TTA) mechanism based on physiological constraints is introduced during the model inference stage, achieving complete cross-subject transfer from contact signals to non-contact signals. This method retains the advantages of deep learning in modeling complex nonlinear relationships while guiding the model to adaptively adjust through physiological constraints. It can stably and accurately estimate systolic and diastolic blood pressure under unknown subjects and different measurement scenarios, achieving high precision and high reliability in continuous non-invasive blood pressure monitoring.
[0008] The present invention discloses a cross-subject blood pressure estimation method based on physiological constraints and test source adaptation, comprising the following steps:
[0009] Step 1: Collect the publicly available PPG signal dataset – PulseDB. Each raw record in this dataset is segmented into 10-second segments, and the signal sampling frequency is uniformly set to 125Hz. To meet experimental requirements, this invention selects 70,000 data points from PulseDB to initially construct the experimental dataset. Each data point is then subjected to a 0.5-10Hz bandpass filter to obtain the final PPG sequence signal.
[0010] Step 2: Acquire neck pulse wave signals using millimeter-wave radar to create a target dataset. First, perform range-dimensional FFT on the original two-dimensional radar data to locate the neck range cells (bins) and extract the original phase signals. Then, use a circular dynamic tracking algorithm to eliminate DC offset. Finally, combine the extended differential cross-multiplication (DACM) algorithm to solve the phase entanglement and ambiguity problems and extract high-quality neck pulse wave signals.
[0011] Step 3: Obtain a pre-trained model based on the publicly available PPG dataset. This includes the following steps:
[0012] Step 3 (a): Perform Z-score normalization on each PPG sequence signal obtained by bandpass filtering in Step 1.
[0013] Step 3 (ii): Based on the PPG sequence signals preprocessed in Step 3 (i), construct a complete experimental dataset; then randomly divide the dataset into training set, validation set and test set in a ratio of 7:1.5:1.5.
[0014] Step 3 (III) The designed network is the Three-Stream Fusion Pulse Wave Blood Pressure Prediction Network (MSTF-Net), which adopts a three-stream parallel fusion architecture to extract deep information of PPG signals from three dimensions: local features, spatial attention, and temporal dependence.
[0015] Step 3 (4): Pre-train MSTF-Net based on the training set partitioned in Step 3 (2). The specific training process uses the Adam optimizer, with an initial learning rate set to... The training epochs were set to 30. The mean squared error (MSE) loss function was used. During training, an early stopping strategy was employed to prevent overfitting: training was terminated and the current optimal model parameters were saved when the validation set loss did not decrease for 10 consecutive epochs. Simultaneously, the optimal model parameters were tested on the test set.
[0016] Step 4: Construct a cross-signal transfer learning framework to achieve blood pressure estimation across modalities and subjects. This includes the following steps:
[0017] Step 4 (a): Preprocess the neck pulse wave signal acquired in Step 2: First, use 0.5-10Hz bandpass filtering to remove high-frequency noise and baseline drift, then perform Z-score normalization in the same way as in Step 3 (a) to obtain the standardized neck pulse wave sequence signal.
[0018] Step 4 (II): Load the optimal pre-trained model parameters saved in Step 3. To retain the general hemodynamic knowledge learned by the pre-trained model on the PPG large-scale data, and to give the model the ability to adapt to radar signals and new subjects, most of the convolutional layer parameters in the backbone network are frozen, and only the adapter layer, normalization layer, and feature fusion layer in the model are set to the learnable state.
[0019] Step 4 (3): Adaptive online test source based on physiological constraints is used for cross-subject experiments. During the inference phase, for subject samples not previously seen in the target domain, real-time parameter iteration is performed without labeling through the following sub-steps:
[0020] 1. Enhanced Physiological Perception Countermeasures: Adding gradient-based adaptive smoothing perturbations to the input radar signal. By minimizing the consistency loss between the original output and the adversarial example output, the robustness of the model to non-contact measurement noise is improved.
[0021] 2. Dynamic physiological anchoring: The pre-trained model generates initial predicted values for the current sample as "physiological anchors". The mean square error (MSE) between the predicted value and the anchor is calculated during the adaptation process. A weight coefficient that decays exponentially with the number of iterations is introduced to ensure that the model does not deviate from a reasonable physiological baseline while adapting to new subjects.
[0022] 3. Morphological Correlation Constraint: Pearson correlation coefficient is introduced to calculate the correlation between the predicted blood pressure output sequence and the reference anchor sequence. By minimizing the correlation loss, the model is forced to maintain the physiological coherence of systolic and diastolic blood pressure in the temporal fluctuation morphology during adaptive iteration, preventing non-physiological abrupt changes caused by instantaneous noise in radar signals.
[0023] Step 4 (iv): During the adaptive iteration process, extract the statistical features of the initial prediction sequence in real time, specifically including the mean. with standard deviation Based on the confidence interval theory of normal distribution, a dynamic physiological confidence interval is constructed. (in (This is the adjustment coefficient). A dynamic gain penalty term is introduced into the loss function. When the predicted value of the model deviates from this dynamic range during the iteration process, a nonlinear gradient backpropagation is generated, which guides the model parameters to regress in a direction that conforms to the individual physiological benchmark of the subject. Thus, the physiological rationality of the predicted value is achieved without relying on hard truncation.
[0024] Step 4 (5): Construct the overall adaptive loss function, which weights and fuses the consistency loss, mean squared error anchoring loss, correlation loss, and physiological boundary perception loss. For each batch of target domain subject data, use the AdamW optimizer to perform multi-step gradient descent iterations on the learnable parameters activated in Step 4 (2). After processing each batch, restore the model parameters to the initial pre-training state saved in Step 3 to ensure that the inference process between each subject sample is independent and to eliminate data contamination between subjects.
[0025] Step 5: Switch the model, which has undergone online adaptive adjustment in Step 4, to evaluation mode. Input the standardized neck pulse wave sequence signal into the model, extract the deep representation output by the feature fusion layer, and calculate the final continuous estimation sequence of the subject's systolic blood pressure (SBP) and diastolic blood pressure (DBP) through the prediction head.
[0026] Beneficial effects
[0027] This invention addresses the technical challenges of cross-modal generalization and significant individual physiological differences in millimeter-wave radar blood pressure estimation. First, it internalizes large-scale hemodynamic prior knowledge of photoplethysmography (PPG) signals through a pre-training mechanism, enabling the model to capture general cardiovascular characteristics even before encountering target radar data. Then, during the inference phase, it cleverly utilizes a test source adaptive architecture to achieve real-time dynamic calibration of lightweight adapter parameters based solely on the unlabeled raw signals of the current subject. This significantly reduces calibration costs and data labeling dependencies during practical deployment. Simultaneously, it introduces multi-dimensional physiological constraints and a dynamic statistical interval sensing mechanism, ensuring the model adheres to the physical logic of the human circulatory system during adversarial perturbation iterations. This effectively suppresses random noise interference in non-contact measurement environments, ultimately achieving high-precision, robust, and continuous monitoring of systolic and diastolic blood pressure across a completely cross-subject scenario. Attached Figure Description
[0028] Figure 1 This is a flowchart of the present invention;
[0029] Figure 2 This is a PPG signal;
[0030] Figure 3 This is the neck pulse wave signal;
[0031] Figure 4 A three-stream fusion pulse wave blood pressure prediction network. Detailed Implementation Plan
[0032] The present invention will be further described below with reference to the accompanying drawings:
[0033] like Figure 1The illustrated method for cross-subject blood pressure estimation based on physiological constraints and test source adaptation specifically includes the following steps:
[0034] Step 1: Collect the publicly available PPG signal dataset – PulseDB. Each raw record in this dataset is segmented into 10-second segments, with a uniform signal sampling frequency of 125Hz. To meet experimental requirements, this invention selects 70,000 data points from PulseDB to initially construct the experimental dataset. Each data point undergoes a 0.5-10Hz bandpass filter, ultimately yielding the PPG sequence signal as shown below. Figure 2 As shown, denoted as , where L=1250, 125Hz×10 seconds.
[0035] Step 2: Collect neck pulse wave signals using millimeter-wave radar to create a target dataset. Perform range-directed fast Fourier transform (FFT) on the radar's two-dimensional echo matrix, lock the neck range cell (Bin), and extract the original phase signal. Then, use a circular dynamic tracking algorithm based on least squares and gradient descent to eliminate DC offset.
[0036] ,
[0037] Simultaneously, an extended differential and cross-multiplication (DACM) algorithm is applied to resolve phase ambiguity and phase entanglement issues, yielding the final neck pulse wave signal, as shown below. Figure 3 As shown.
[0038] Step 3: Obtain a pre-trained model based on the publicly available PPG dataset. This includes the following steps:
[0039] Step 3 (a) Perform Z-score normalization on each PPG sequence signal obtained from Step 1 after bandpass filtering, i.e. ,in It is the average value of a single PPG signal. That is the standard deviation.
[0040] Step 3 (II): Based on the preprocessed PPG sequence signals from Step 3 (I), construct a complete experimental dataset. The dataset was then randomly divided into training and training sets in a ratio of 7:1.5:1.5. Validation set With test set .
[0041] Step 3 (III): The designed network is a three-stream fusion pulse wave blood pressure prediction network (MSTF-Net), as follows: Figure 4As shown, its core adopts a three-stream parallel fusion architecture, which extracts deep information of PPG signals from three dimensions: local features, spatial attention, and temporal dependence.
[0042] R-stream: This stream aims to capture the morphological features of the pulse wave at local time scales, such as peaks, troughs, and inflection points. It employs a one-dimensional convolutional architecture based on RepVGG, implementing multi-branch convolutional fusion during training and converting to a single-branch structure during inference, thus achieving efficient deployment while maintaining model expressiveness. The R-stream consists of multiple stages, progressively downsampling and expanding the number of channels, ultimately obtaining local feature vectors through global average pooling. The process can be represented as follows:
[0043] ,
[0044] in It consists of multiple RepVGG modules. Indicates global average pooling. This indicates a linear projection layer.
[0045] S-flow: The blood pressure-related information carried by the pulse wave varies significantly at different time points. For example, the rising edge of the pulse is usually more related to changes in arterial compliance, while morphological changes near the peak often reflect peripheral vascular resistance. Therefore, to enable the network to adaptively focus on these physiologically significant key signal segments, S-flow borrows the idea of convolutional block attention modules and introduces a spatial attention mechanism in the one-dimensional pulse wave scenario to achieve dynamic weighting of important time intervals. Specifically, given the input signal... First, average pooling and max pooling features are concatenated along the channel dimension, and then input into a one-dimensional convolution kernel of size [size missing]. The convolution operator is used to model its saliency distribution in the time dimension.
[0046] ,
[0047] in, This is the Sigmoid function. The enhanced signal is...
[0048] ,
[0049] Next, channel expansion is performed using 1×1 convolution, followed by batch normalization and activation function to obtain...
[0050] ,
[0051] T-flow: Blood pressure changes are reflected not only within the single-cycle morphology of the pulse wave but also in long-term dynamic patterns across multiple cardiac cycles, such as temporal variations in pulse wave conduction velocity, heart rate variability, and pulse wave cycle differences. Therefore, T-flow is responsible for modeling these physiologically significant temporal correlations of blood pressure over a long time span. In terms of structural design, T-flow employs a temporal convolutional network (TCN) based on dilated convolutions, combined with a multi-scale convolutional fusion module, to capture complex dynamic patterns across cycles. First, the input sequence... Local temporal features are extracted through a set of one-dimensional convolutional embedding modules.
[0052] ,
[0053] Subsequently, via Layer dilated convolution stacked, with each layer having a dilation rate of [missing value]. Therefore, the receptive field expands exponentially, enabling the model to cover multiple cardiac cycles:
[0054] ,
[0055] in, The implementation incorporates two layers of dilated convolutions, BatchNorm, residual connections, and a causal pruning structure to ensure the output meets the causal requirements of time series data. To further enhance cross-scale temporal representation, T-flows are applied to the final features. Perform multi-scale convolutional fusion:
[0056] ,
[0057] in, This represents a multi-scale one-dimensional convolution operation, specifically defined as:
[0058] ,
[0059] in, The kernel size is denoted by 1, 3, 5, or 7. Then, from... Extract global average pooling and global max pooling features to obtain , The two sequences are concatenated at the channel dimension and then input into a fusion network to generate the final T-stream temporal embedding:
[0060] ,
[0061] in, and These are the weight matrix and bias vector for the linear transformation, respectively.
[0062] Step 3 (4): Training set based on the partitioning in Step 3 (2) The three-stream fusion pulse wave blood pressure prediction network constructed in step three (iii) Pre-training is performed, in which This is a preprocessed PPG signal segment. For the corresponding systolic and diastolic blood pressure labels, These are the total network parameters. The initial learning rate of the model. Set as The training epochs are set to 30. The goal of the pre-training phase is to learn a universal mapping relationship from physiological signals to blood pressure values, and its loss function is... L1 loss is defined as the mean absolute error (MAE) between the fused output and the true label.
[0063] ,
[0064] in, This indicates the fused output branch of the network; The L1 norm of the vector represents the sum of the absolute errors in predicting systolic and diastolic blood pressure. This represents the total number of training set samples during the pre-training phase. The pre-training process employs a cosine annealing learning rate scheduler, with a learning rate... The variation with time step t is as follows:
[0065] ,
[0066] in, The initial learning rate, To minimize the learning rate, This represents the total number of training rounds. Simultaneously, an early stopping strategy is employed to prevent overfitting: if the loss on the validation set is continuous... If the performance does not decrease in the validation set, training will be terminated early, and the parameters of the best-performing model on the validation set will be saved. During this pre-training phase, the intermediate layer adapter module in the network is turned off to ensure that the network's backbone feature extraction capabilities are fully learned.
[0067] Step 4: Construct a cross-signal transfer learning framework to achieve blood pressure estimation across modalities and subjects. This includes the following steps:
[0068] Step 4 (a): Preprocess the neck pulse wave signal acquired in Step 2: First, use 0.5-10Hz bandpass filtering to remove high-frequency noise and baseline drift, then perform Z-score normalization in the same way as in Step 3 (a) to obtain the standardized neck pulse wave sequence signal.
[0069] Step 4 (II): Load the optimal pre-trained model parameters saved in Step 3. To retain the general hemodynamic knowledge learned by the pre-trained model on the large-scale PPG data, and to give the model the ability to adapt to radar signals and new subjects, freeze most of the convolutional layer feature extraction parameters in the network backbone. Only set the parameters of the adapter layer, normalization layer, and feature fusion layer inserted in the model to a learnable state. Let the set of learnable parameters activated at this time be denoted as . .
[0070] Step 4 (3): Adaptive online test source based on physiological constraints is used for cross-subject experiments. During the inference phase, for subject samples not previously seen in the target domain, real-time parameter iteration is performed without labeling through the following sub-steps:
[0071] 1. Enhanced physiological perception resistance: This occurs when inputting radar signals... A gradient-based adaptive smoothing perturbation is added to improve the model's robustness to non-contact measurement noise by minimizing the consistency loss between the original output and the adversarial example output. (Adversarial perturbation signal) The formula for generating it is:
[0072] ,
[0073] in, For symbolic functions, This represents the dynamic perturbation amplitude, adaptively calculated based on the signal's own standard deviation. Consistency loss. Defined as the sum of the mean square errors of the prediction results of the original signal and the perturbation signal:
[0074] ,
[0075] 2. Dynamic physiological anchoring: Utilizing a pre-trained model to generate initial predicted values for the current sample as "physiological anchor points". Calculate the current adaptive prediction value The mean squared error (MSE) between the anchor point and the target point. A dynamic weighting coefficient is introduced that decays exponentially with the predicted offset during the iteration process. This ensures that the model, while adapting to new subjects, does not deviate from reasonable physiological baselines. Dynamic anchoring loss. The formula is as follows:
[0076] ,
[0077] ,
[0078] 3. Morphological Correlation Constraint: Pearson correlation coefficient is introduced to calculate the correlation between the predicted blood pressure output sequence and the reference anchor sequence. By minimizing the correlation loss, the model is forced to maintain the physiological coherence of systolic and diastolic blood pressure in terms of temporal fluctuation morphology during adaptive iteration, preventing non-physiological abrupt changes caused by instantaneous noise in the radar signal. For any sequence, the Pearson correlation loss... Defined as
[0079] ,
[0080] Step 4 (iv): During the adaptive iteration process, extract the statistical features of the initial prediction sequence in real time, specifically including the mean. with standard deviation Based on the confidence interval theory of normal distribution, a dynamic physiological confidence interval is constructed. (in The adjustment coefficient is preferably 2.5. A dynamic gain penalty term (Range Loss) and a collapse-prevention variance constraint term (Variance Loss) are introduced into the loss function. When the model's predicted value deviates from this dynamic range during the iteration process, a nonlinear gradient backpropagation is generated, guiding the model parameters to regress towards a direction consistent with the individual physiological baseline of the subject. This achieves physiological rationality of the predicted values without relying on rigid numerical truncation. The penalty term formula is as follows:
[0081] ,
[0082] Step four (five): Construct the total adaptive loss function, which weights and fuses the consistency loss, mean squared error anchoring loss, correlation loss, and physiological boundary perception loss to obtain the total loss. The calculation formula is:
[0083] .
[0084] For each batch of target domain radar subject data, the AdamW optimizer is used to optimize the learnable parameters activated in step four (ii). Perform multi-step (e.g., 5-step) gradient descent iterations. After each batch is processed, restore the model parameters to the initial pre-training state saved in step three to ensure that the inference process between each subject sample is independent and completely eliminate data contamination between subjects.
[0085] Step 5: Switch the model, which has undergone online adaptive adjustment in Step 4, to the evaluation mode (Eval). Input the standardized neck pulse wave sequence signal into the model, extract the deep representation output by the feature fusion layer, and calculate the final continuous estimation sequence of the subject's systolic blood pressure (SBP) and diastolic blood pressure (DBP) through the prediction head, thus completing the accurate estimation of non-invasive blood pressure.
Claims
1. A method for cross-subject blood pressure estimation based on physiology constraints and test source adaptation, characterized in that Includes the following steps: Step 1: Collect the publicly available PPG signal dataset – PulseDB. Each raw record in this dataset is segmented into 10-second segments, and the signal sampling frequency is uniformly set to 125Hz. To meet experimental requirements, this invention selects 70,000 data points from PulseDB to initially construct the experimental dataset. Each data point undergoes a 0.5-10Hz bandpass filter, resulting in the PPG sequence signal shown in Figure 2, denoted as... , where L=1250, 125Hz×10 seconds; Step 2: Collect neck pulse wave signals using millimeter-wave radar to create a target dataset. Perform range-directed fast Fourier transform (FFT) on the radar's two-dimensional echo matrix, lock the neck range cell (Bin) and extract the original phase signal. Then, use a circular dynamic tracking algorithm based on least squares and gradient descent to eliminate DC offset. , Meanwhile, the extended differential and cross-multiplication (DACM) algorithm is applied to solve the phase ambiguity and phase entanglement problems, and the final neck pulse wave signal is shown in Figure 3. Step 3: Obtain a pre-trained model based on the publicly available PPG dataset; Step 4: Construct a cross-signal transfer learning framework to achieve blood pressure estimation across modalities and subjects; Step 5: Switch the model, after online adaptive adjustment in Step 4, to the evaluation mode (Eval). Input the standardized neck pulse wave sequence signal into the model, extract the deep representation output by the feature fusion layer, and calculate the final continuous estimation sequence of the subject's systolic blood pressure (SBP) and diastolic blood pressure (DBP) through the prediction head, thus completing the accurate estimation of non-invasive blood pressure; A cross-subject blood pressure estimation method based on physiological constraints and test source adaptation according to claim 1, characterized in that... Step three includes the following steps: Step 3: Obtain a pre-trained model based on the publicly available PPG dataset. This includes the following steps: Step 3 (a) Perform Z-score normalization on each PPG sequence signal obtained from Step 1 after bandpass filtering, i.e. ,in It is the average value of a single PPG signal. It is the standard deviation; Step 3 (II): Based on the preprocessed PPG sequence signals from Step 3 (I), construct a complete experimental dataset. The dataset was then randomly divided into training and training sets in a ratio of 7:1.5:1.
5. Validation set With test set ; Step 3 (III) The designed network is the three-stream fusion pulse wave blood pressure prediction network (MSTF-Net) as shown in Figure 4. Its core adopts a three-stream parallel fusion architecture to extract deep information of PPG signals from three dimensions: local features, spatial attention, and temporal dependence. R-stream: This stream aims to capture the morphological features of the pulse wave at local time scales, such as peaks, troughs, and inflection points. It employs a one-dimensional convolutional architecture based on RepVGG, implementing multi-branch convolutional fusion during training and converting to a single-branch structure during inference, thus achieving efficient deployment while maintaining model expressiveness. The R-stream consists of multiple stages, progressively downsampling and expanding the number of channels, ultimately obtaining local feature vectors through global average pooling. The process can be represented as follows: , in It consists of multiple RepVGG modules. Indicates global average pooling. Indicates a linear projection layer; S-flow: The blood pressure-related information carried by the pulse wave varies significantly at different time points. For example, the rising edge of the pulse is usually more related to changes in arterial compliance, while morphological changes near the peak often reflect peripheral vascular resistance. Therefore, to enable the network to adaptively focus on these physiologically significant key signal segments, S-flow borrows the idea of convolutional block attention modules and introduces a spatial attention mechanism in the one-dimensional pulse wave scenario to achieve dynamic weighting of important time intervals. Specifically, given the input signal... First, average pooling and max pooling features are concatenated along the channel dimension, and then input into a one-dimensional convolution kernel of size [size missing]. The convolution operator is used to model its saliency distribution in the time dimension; , in, This is the Sigmoid function. The enhanced signal is: , Next, channel expansion is performed using 1×1 convolution, followed by batch normalization and activation function to obtain... , T-flow: Blood pressure changes are reflected not only within the single-cycle morphology of the pulse wave but also in long-term dynamic patterns across multiple cardiac cycles, such as temporal variations in pulse wave conduction velocity, heart rate variability, and pulse wave cycle differences. Therefore, T-flow is responsible for modeling these physiologically significant temporal correlations of blood pressure over a long time span. In terms of structural design, T-flow employs a temporal convolutional network (TCN) based on dilated convolutions, combined with a multi-scale convolutional fusion module, to capture complex dynamic patterns across cycles. First, the input sequence... Local temporal features are extracted through a set of one-dimensional convolutional embedding modules. , Subsequently, via Layer dilated convolution stacked, with each layer having a dilation rate of [missing value]. Therefore, the receptive field expands exponentially, enabling the model to cover multiple cardiac cycles: , in, The implementation incorporates two layers of dilated convolutions, BatchNorm, residual connections, and a causal pruning structure to ensure the output meets the causal requirements of time series data. To further enhance cross-scale temporal representation, T-flows are applied to the final features. Perform multi-scale convolutional fusion: , in, This represents a multi-scale one-dimensional convolution operation, specifically defined as: , in, The kernel size is denoted by 1, 3, 5, or 7. Then, from... Extract global average pooling and global max pooling features to obtain , The two sequences are concatenated at the channel dimension and then input into a fusion network to generate the final T-stream temporal embedding: , in, and These are the weight matrix and bias vector for the linear transformation, respectively; Step 3 (4): Training set based on the partitioning in Step 3 (2) The three-stream fusion pulse wave blood pressure prediction network constructed in step three (iii) Pre-training is performed, in which This is a preprocessed PPG signal segment. For the corresponding systolic and diastolic blood pressure labels, These are the total network parameters. The initial learning rate of the model. Set as The training epochs are set to 30. The goal of the pre-training phase is to learn a universal mapping relationship from physiological signals to blood pressure values, and its loss function is... L1 loss is defined as the mean absolute error (MAE) between the fused output and the true label. , in, This indicates the fused output branch of the network; The L1 norm of the vector represents the sum of the absolute errors in predicting systolic and diastolic blood pressure. This represents the total number of training set samples during the pre-training phase. The pre-training process employs a cosine annealing learning rate scheduler, with a learning rate... The variation with time step t is as follows: , in, The initial learning rate, To minimize the learning rate, This represents the total number of training rounds. Simultaneously, an early stopping strategy is employed to prevent overfitting: if the loss on the validation set is continuous... If the performance does not decrease in the validation set, training will be terminated early, and the parameters of the best-performing model on the validation set will be saved. During this pre-training phase, the intermediate layer adapter modules in the network are turned off to ensure that the network's backbone feature extraction capabilities are fully learned. A cross-subject blood pressure estimation method based on physiological constraints and test source adaptation according to claim 1, characterized in that... Step four includes the following steps: Step 4: Construct a cross-signal transfer learning framework to achieve blood pressure estimation across modalities and subjects. This includes the following steps: Step 4 (a): Preprocess the neck pulse wave signal acquired in Step 2: First, use 0.5-10Hz bandpass filtering to remove high-frequency noise and baseline drift, then perform Z-score normalization, the same as in Step 3 (a), to obtain the standardized neck pulse wave sequence signal; Step 4 (II): Load the optimal pre-trained model parameters saved in Step 3. To retain the general hemodynamic knowledge learned by the pre-trained model on the large-scale PPG data, and to give the model the ability to adapt to radar signals and new subjects, most of the convolutional layer feature extraction parameters in the network backbone are frozen. Only the parameters of the adapter layer, normalization layer, and feature fusion layer inserted in the model are set to a learnable state. Let the set of learnable parameters activated at this time be denoted as . ; Step 4 (3): Adaptive online test source based on physiological constraints is used for cross-subject experiments. During the inference phase, for subject samples not previously seen in the target domain, real-time parameter iteration is performed without labeling through the following sub-steps:
1. Enhanced physiological perception resistance: This occurs when inputting radar signals... A gradient-based adaptive smoothing perturbation is added to improve the model's robustness to non-contact measurement noise by minimizing the consistency loss between the original output and the adversarial example output. (Adversarial perturbation signal) The formula for generating it is: , in, For symbolic functions, This represents the dynamic perturbation amplitude, adaptively calculated based on the signal's own standard deviation. Consistency loss. Defined as the sum of the mean square errors of the prediction results of the original signal and the perturbation signal: , 2. Dynamic physiological anchoring: Utilizing a pre-trained model to generate initial predicted values for the current sample as "physiological anchor points". Calculate the current adaptive prediction value The mean squared error (MSE) between the anchor point and the target point. A dynamic weighting coefficient is introduced that decays exponentially with the predicted offset during the iteration process. This ensures that the model, while adapting to new subjects, does not deviate from reasonable physiological baselines. Dynamic anchoring loss. The formula is as follows: , , 3. Morphological Correlation Constraint: Pearson correlation coefficient is introduced to calculate the correlation between the predicted blood pressure output sequence and the reference anchor sequence. By minimizing the correlation loss, the model is forced to maintain the physiological coherence of systolic and diastolic blood pressure in terms of temporal fluctuation morphology during adaptive iteration, preventing non-physiological abrupt changes caused by instantaneous noise in the radar signal. For any sequence, the Pearson correlation loss... Defined as , Step 4 (iv): During the adaptive iteration process, extract the statistical features of the initial prediction sequence in real time, specifically including the mean. with standard deviation Based on the confidence interval theory of normal distribution, a dynamic physiological confidence interval is constructed. (in The adjustment coefficient is preferably 2.
5. A dynamic gain penalty term (Range Loss) and a collapse-prevention variance constraint term (Variance Loss) are introduced into the loss function. When the model's predicted value deviates from this dynamic range during the iteration process, a nonlinear gradient backpropagation is generated, guiding the model parameters to regress towards a direction consistent with the individual physiological baseline of the subject. This achieves physiological rationality of the predicted values without relying on rigid numerical truncation. The penalty term formula is as follows: , Step four (five): Construct the total adaptive loss function, which weights and fuses the consistency loss, mean squared error anchoring loss, correlation loss, and physiological boundary perception loss to obtain the total loss. The calculation formula is: . For each batch of target domain radar subject data, the AdamW optimizer is used to optimize the learnable parameters activated in step four (ii). Perform multi-step (e.g., 5-step) gradient descent iterations. After each batch is processed, restore the model parameters to the initial pre-training state saved in step three to ensure that the inference process between each subject sample is independent and completely eliminate data contamination between subjects.