Sleeveless blood pressure continuous estimation method, device and system and storage medium
By combining a teacher-student distillation network with test batch normalization, the problems of signal quality instability and cross-domain distribution differences in cuffless blood pressure estimation are solved. This approach achieves high accuracy, interpretability, and deployability under single-channel conditions, making it suitable for various medical and health monitoring scenarios.
Patent Information
- Application Number
- CN202610016232.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-07
- Publication Date
- 2026-02-10
AI Technical Summary
Existing technologies struggle to achieve stable, deployable, and auditable cuffless continuous blood pressure estimation across multiple domains, populations, and long time periods, due to limitations such as unstable signal quality, insufficient model generalization ability, cross-domain distribution differences, and lack of interpretability.
We employ a combination of teacher-student distillation network and test batch normalization, use data quality index for signal quality gating, multi-scale convolution and lightweight UNet-1D network for feature extraction, and perform individual baseline calibration to achieve cross-domain adaptation and interpretability.
It achieves near-multi-channel estimation accuracy under single-channel conditions, reduces systematic bias across devices, scenarios, and populations, and improves the interpretability and reliability of the method, making it suitable for ICU monitoring, perioperative management, and daily wearable health monitoring.
Smart Images
Figure CN121489429A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of health monitoring technology, specifically relating to a cuffless blood pressure continuous estimation method, device, system, and storage medium. Background Technology
[0002] Hypertension is one of the major risk factors for cardiovascular disease, and both clinical diagnosis and treatment as well as outpatient management rely on long-term blood pressure monitoring. Currently, although invasive catheter artery measurement is considered the gold standard for blood pressure monitoring, it requires arterial puncture and catheterization, making it only suitable for intensive care or perioperative settings and difficult to promote in general wards and home environments. While cuff-based oscillometric measurement is non-invasive, the measurement process relies on periodic inflation and deflation, which has problems such as insufficient comfort and difficulty in achieving high temporal resolution continuous monitoring, and cannot continuously reflect risk information such as morning peaks, nocturnal blood pressure load, and short-term variability.
[0003] With the development of wearable devices, photoplethysmography (PPG) is considered a feasible approach for cuffless continuous blood pressure estimation due to its simple hardware structure, ease of integration into wristbands, bracelets, and support for high sampling rates. However, under actual wearing conditions, PPG signals are easily affected by motion artifacts, ambient light interference, changes in skin-sensor contact, and baseline drift, resulting in unstable signal quality and morphological feature distortion. This affects the effectiveness of model training and the stability of prediction results during deployment. Existing technologies often only remove low-quality segments after feature extraction or model prediction, failing to effectively couple quality assessment with model training and adaptation processes, thus making it difficult to suppress the impact of low-quality samples on model performance at the source.
[0004] In terms of modeling methods, traditional machine learning methods often rely on manually designed PPG morphology and time-frequency features, such as pulse propagation time, rising slope, notch amplitude, and systolic / diastolic area ratio. These methods offer some interpretability but are sensitive to population differences and cross-device variations, resulting in limited generalization ability. End-to-end deep learning methods based on convolutional neural networks, UNet, or Transformer can automatically extract multi-scale features and achieve high accuracy on single datasets. However, they typically require large model sizes and computational resources, making them unsuitable for deployment on resource-constrained wearable devices. Furthermore, the model's decision-making criteria are not easily correlated with clinical physiological mechanisms, affecting its reliability in medical scenarios.
[0005] In practical applications, a distributional discrepancy between the training and target domains is common. On the one hand, differences in sensor optical characteristics, acquisition processes, wearing methods, and environmental conditions can cause covariate drift in the input feature distribution. On the other hand, differences in blood pressure baselines and fluctuation ranges among different populations and individuals lead to label distribution shifts. Some existing domain adaptation methods can alleviate input distribution discrepancies to some extent, but without prior quality gating, they may introduce statistical information from noisy segments when aligning distribution statistics, resulting in negative transfer. On the other hand, schemes that rely solely on batch normalization to update statistics during testing are insufficient to simultaneously handle label distribution shifts and individual baseline differences.
[0006] Furthermore, existing technologies primarily focus on numerical error metrics on single datasets, lacking a quantitative assessment of "cross-domain inference consistency," i.e., an analysis of whether the model relies on the same physiological evidence to make predictions across different populations, devices, and scenarios. This results in a lack of quantifiable evidence to support existing methods in clinical applications and regulatory reviews, hindering subsequent version iterations and risk tracking. In summary, existing technologies have not yet formed a comprehensive technical chain from signal input, model representation, external domain adaptation to result auditing, making it difficult to achieve stable, deployable, and auditable cuffless continuous blood pressure estimation in multi-domain, multi-population, and long-term scenarios. Summary of the Invention
[0007] To address the problems existing in the prior art, this invention provides a cuffless blood pressure continuous estimation method, device, system, and storage medium. It can achieve continuous cuffless blood pressure estimation that meets medical error requirements under single-channel wearable PPG conditions, taking into account end-side deployability, cross-domain robustness, and interpretability. It is applicable to various scenarios such as ICU monitoring, perioperative management, and daily wearable health monitoring.
[0008] To achieve the above objectives, the present invention provides the following solution: A method for continuous estimation of cuffless blood pressure includes: Step S1: Acquire PPG signal; Step S2: Preprocess the PPG signal; Step S3: Calculate the Data Quality Index (DQI) and perform quality gating on the preprocessed PPG signal; Step S4: Construct a teacher-student distillation network; Step S5: DQI-weighted joint distillation and supervised training teacher-student distillation network; Step S6: Based on the trained teacher-student distillation network, perform unlabeled distribution adaptation on the unlabeled data in the target domain using time-batch normalized TTA-BN. Step S7: Perform individual baseline calibration based on affine / quantile mapping on the student network output after TTA-BN adaptation; Step S8: Online cuffless blood pressure estimation and quality feedback output.
[0009] Preferably, preprocessing of the PPG signal includes: resampling, bandpass filtering, detrending, and normalization.
[0010] Preferably, in step S4, a teacher network is constructed with three-channel pulse signals (PPG, VPG, and APG) as input. The teacher network adopts a one-dimensional convolution and encoder-decoder structure, and uses multi-scale convolution, downsampling, residual blocks, and skip connections to extract multi-scale pulse morphology features. It outputs a continuous blood pressure waveform aligned with invasive arterial pressure, as well as pulsatile systolic and diastolic blood pressure statistics. A channel elastic UNet-1D student network is constructed with single-channel PPG or few-channel pulse signals as input. The student network sets feature alignment positions in multiple intermediate layers of the encoder and simultaneously provides blood pressure waveform and SBP / DBP multiquantile estimates at the output.
[0011] The present invention also provides a cuffless continuous blood pressure estimation device, comprising: The first processing module is used to acquire PPG signals; The second processing module is used for preprocessing the PPG signal; The third processing module is used to calculate the Data Quality Index (DQI) and perform quality gating on the preprocessed PPG signal; The fourth processing module is used to construct the teacher-student distillation network; The fifth processing module is used for DQI-weighted joint distillation and supervised training of teacher-student distillation networks; The sixth processing module is used to perform unlabeled distribution adaptation on unlabeled data in the target domain using time-batch normalized TTA-BN based on the trained teacher-student distillation network. The seventh processing module is used to perform individual baseline calibration based on affine / quantile mapping on the student network output after TTA-BN adaptation; The eighth processing module is used for online cuffless blood pressure estimation and quality prompt output.
[0012] Preferably, the second processing module performs preprocessing on the PPG signal, including resampling, bandpass filtering, detrending, and normalization.
[0013] As a preferred approach, the fourth processing module constructs a teacher network with PPG, VPG, and APG three-channel pulse signals as input. The teacher network adopts a one-dimensional convolution and encoder-decoder structure, and uses multi-scale convolution, downsampling, residual blocks, and skip connections to extract multi-scale pulse morphology features. It outputs a continuous blood pressure waveform aligned with invasive arterial pressure, as well as pulsatile systolic and diastolic blood pressure statistics. A channel elastic UNet-1D student network is constructed with single-channel PPG or few-channel pulse signals as input. The student network sets feature alignment positions in multiple intermediate layers of the encoder and simultaneously provides blood pressure waveform and SBP / DBP multiquantile estimates at the output.
[0014] The present invention also provides a cuffless blood pressure continuous estimation system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a cuffless blood pressure continuous estimation method when executed by the processor.
[0015] The present invention also provides a storage medium storing a computer program that executes a cuffless blood pressure continuous estimation method when running.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention is an integrated cuffless blood pressure continuous estimation method based on quality gating and distillation adaptation. By introducing a differentiable data quality index at the signal inlet and using it collaboratively in model training, batch normalization adaptation during testing, and result confidence management, it effectively reduces the interference of low-quality pulse segments on model parameters and statistics. Through knowledge distillation of a multi-channel teacher network and a lightweight UNet-1D student network, it achieves estimation accuracy close to that of a multi-channel scheme using only a single-channel PPG, meeting the constraints of computing power and power consumption on wearable devices. By combining zero-gradient distribution adaptation with batch normalization during testing with individual baseline calibration based on affine / quantile→Affine, it completes scene-level and individual-level distribution alignment without large-scale annotation of the target domain, reducing system bias and scattering errors caused by cross-device, cross-scene, and cross-population issues. Furthermore, it is supplemented by cross-domain feature importance consistency analysis based on SHAP, improving the interpretability and acceptability of the method in clinical applications and regulatory reviews in terms of both numerical accuracy and reasoning basis. It is applicable to cuffless blood pressure continuous estimation in various application scenarios such as ICU monitoring, perioperative anesthesia management, and daily wearable health monitoring. Attached Figure Description
[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of the cuffless blood pressure continuous estimation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram comparing PPG signal waveforms of different qualities provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the teacher-student distillation network and adaptive calibration structure provided in an embodiment of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] Example 1 like Figure 1 As shown, the present invention provides a method for continuous estimation of blood pressure without a cuff, comprising: Step S1: PPG signal acquisition; Step S2: Preprocess the PPG signal; Step S3: Data Quality Index (DQI) calculation and quality gating; Step S4: Construction of the teacher-student distillation network; Step S5: DQI-weighted joint distillation and supervised training; Step S6: Adaptive testing based on zero gradient distribution normalized during test batch; Step S7: Individual baseline calibration based on affine / quantile mapping; Step S8: Online cuffless blood pressure estimation and quality feedback output.
[0022] As one embodiment of the present invention, in step S1, the wearable terminal 201 continuously collects the user's PPG signal, preferably with a sampling rate of 125 to 256 Hz.
[0023] In one embodiment of the present invention, in step S2, the photoplethysmography (PPG) signal is uniformly preprocessed on the terminal side or in the data processing unit. This includes: resampling the PPG signals acquired by different devices to unify the sampling rate to a target sampling rate; filtering the PPG signals using a zero-phase bandpass filter, preferably with a passband of 0.5–8 Hz, to remove baseline drift and high-frequency noise; detrending and amplitude normalizing each signal to reduce the impact of individual optical path differences and wearing status differences; and dividing the preprocessed PPG signal into multiple fixed-length time windows according to a preset window length (e.g., 6–8 s) and overlap rate. In scenarios with invasive arterial pressure or cuff-type blood pressure recording, the corresponding true blood pressure value can be mapped to each time window through time alignment, serving as a reference label for teacher network training and individual calibration.
[0024] As one embodiment of the present invention, in step S3, as follows: Figure 2 As shown, high-quality PPG waveforms differ significantly in morphology from low-quality PPG waveforms containing obvious baseline drift, saturation truncation, or motion artifacts. To explicitly introduce signal quality information during training and inference, this embodiment constructs multiple quality features for each time window w and synthesizes them into a data quality index DQI(w). Specifically, optionally, the statistics of the quality features can be calculated as follows: ; In the formula, This represents the number of sampling points within that time window. Indicates the signal mean. Indicates standard deviation, Indicates kurtosis. Indicates skewness, This represents the root mean square value. These statistical characteristics are used to characterize the quality features of the PPG waveform in terms of amplitude, symmetry, and energy distribution. Further, optionally, this embodiment can also calculate quality sub-indices such as template similarity, pulsatility stability, and spectral energy concentration, and uniformly denoted as... , , , After normalization, the specific calculation process of the data quality index DQI(w) can be expressed as follows: ; In the formula, q1(w) is the energy proportion of the 0.5-8Hz frequency band, q2(w) is the high-frequency noise energy proportion, q3(w) is the normalized cross-correlation index with the PPG pulsation of the population template, and q4(w) is the motion artifact index MI constructed based on the coherence of PPG and triaxial acceleration and / or sample entropy (MI is approximated by a weighted combination of the coherence Cxy(f0) of PPG and triaxial acceleration in the heart rate main frequency neighborhood and the zero-crossing rate zcr). These are learnable or preset weight coefficients. For bias terms; For the Sigmoid mapping function, such that The range of values is A higher value indicates higher signal quality within the window. Optionally, during the training phase, [the value can be adjusted]. As the weighting coefficient of the loss function for each time window, for the th Total loss for each window Perform weighted aggregation: ; Meanwhile, during the testing and adaptation phases, the DQI(w) can be used to determine whether to update the BatchNormalization statistic, and low-confidence labels can be added to the blood pressure estimation results of low-quality windows or they can be excluded from output, thus achieving quality gating.
[0025] As one embodiment of the present invention, in step S4, as follows: Figure 3 As shown, the teacher-student distillation network of the present invention includes a teacher network and a student network. Specifically, the teacher network takes PPG, VPG and APG three-channel signals as input and obtains a stable multi-scale pulse representation after training in the source domain; the student network adopts a channel-flexible UNet-1D structure, takes a single-channel PPG as the main input, and can be expanded to a three-channel input when a derivative channel exists, so as to take into account wearable terminal deployment and structural alignment with the teacher network. Specifically, a teacher network is constructed with three-channel pulse signals (PPG, VPG, and APG) as input. The teacher network adopts a one-dimensional convolution and encoder-decoder structure, and uses multi-scale convolution, downsampling, residual blocks, and skip connections to extract multi-scale pulse morphology features. It outputs a continuous blood pressure waveform aligned with invasive arterial pressure, as well as statistics such as pulsatile systolic and diastolic blood pressure. A channel-elastic UNet-1D student network is constructed with single-channel PPG or few-channel pulse signals as input. The student network sets feature alignment positions in multiple intermediate layers of the encoder and simultaneously provides blood pressure waveform and SBP / DBP multiquantile estimates at the output. This achieves a network structure suitable for wearable device deployment under limited parameter conditions.
[0026] As one embodiment of the present invention, in step S5, to introduce physiological priors without disrupting end-to-end modeling, the present invention calculates a set of morphological features for each time window and constructs a "morphological prior vector". The morphological features may include, but are not limited to, amplitude statistics, rising and falling times, full width at half maximum (FWHM), notch position, energy, and spectral centroid. After standardizing the vector on the training set, it is linearly transformed and mapped to the latent space, then expanded temporally into conditional channels with the same length as the input, and concatenated with the original PPG (and optional derivative channels) in the channel dimension. Furthermore, affine modulation similar to FiLM is introduced at the bottleneck layer of the student network: channel scaling factors and bias factors are generated from the prior vector through two layers of perceptrons, and intermediate features are scaled and offset channel by channel, thereby enhancing the network's utilization of morphological priors. When a prior branch significantly drifts or becomes unstable in a certain target domain, it can be reduced to zero through gating coefficients, causing the student network security to degenerate into a pure end-to-end structure.
[0027] On the output side, the student network sets up two types of task heads: one is a continuous blood pressure waveform regression head, used to output blood pressure waveforms aligned with invasive arterial pressure; the other is a quantile statistics head, used to output window-level or pulse-level SBP / DBP multiquantile estimates to support subsequent individual affine / quantile calibration. To effectively inherit the multi-channel capabilities of the teacher network while taking into account the real blood pressure labels, this invention constructs a joint loss function for each time window. Optionally, for the j-th time window, waveform distillation loss, quantile distillation loss, feature alignment loss, supervised regression loss, and morphological regularization loss can be introduced, and their calculation process can be represented as follows: ; In the formula, and These represent the blood pressure waveforms output by the student network and the teacher network, respectively. and For the corresponding quantile estimation, and These are intermediate features after linear projection. This is the reference blood pressure when a true value exists. Represents the time-domain total variation regularity. Represents the short-time Fourier transform amplitude spectrum. Assign weights to each loss term. This represents the data quality index corresponding to the time window. It is evident that high-quality windows have a higher weight in the overall loss, while the interference from low-quality windows is softly suppressed. Optionally, the training strategy can adopt a two-stage approach of "supervised warm-up + joint distillation": In the first stage, only supervised regression loss and morphological regularization loss are applied to the student network to enable it to obtain preliminary stable waveform modeling capabilities under single-channel input; in the second stage, waveform distillation, quantile distillation, and feature alignment loss are enabled and optimized together with the supervised loss, allowing the student network to maintain end-device deployability while inheriting as much discriminative representation as possible from the discriminative representation learned by the teacher network under multi-channel conditions. During training, random channel masking can also be applied to the derivative channels of the teacher network input, and the input morphology of the student network can be randomly mixed to enable the student to internalize the key information carried by the derivative channels even when only PPG input is available.
[0028] Furthermore, in the source domain training dataset, the teacher network is first supervised and trained using invasive arterial pressure or cuff blood pressure ground truth values to achieve stable blood pressure estimation capabilities under multi-channel input. Based on this, the teacher network parameters are fixed, and the intermediate features and outputs of the teacher network are used as soft objectives. Waveform distillation loss and statistical distillation loss are applied to the student network, and supervised regression loss is superimposed on the window containing the blood pressure ground truth values. Furthermore, morphological regularization losses such as total variation and second-order curvature are combined to constrain the smoothness and physiological rationality of the student network's output waveform. All of the above losses are aggregated using the DQI(w) corresponding to the window as a weighting coefficient, thereby enhancing the weight of high-quality samples in parameter updates and suppressing interference from low-quality samples. During training, random channel masking can be applied to the derivative channels VPG and / or APG of the teacher network, blocking the derivative channel input with a certain probability, guiding the student network to learn and internalize the morphological information carried by the derivative channels when only PPG input is available.
[0029] As one embodiment of the present invention, in step S6, for the unlabeled data in the target domain, batch normalization is used to adapt the unlabeled distribution. The statistics of each BatchNormalization layer in the student network are updated in batches to adapt to the differences in input distribution under different devices and scenarios.
[0030] Specifically, the statistical update process can be expressed as follows: ; In the formula, The first one at the end of training Mean and variance of layer BatchNormalization The mean and variance are estimated on the current batch in the target domain. for The momentum coefficient between them. Optionally, based on the average quality index of all windows within the current batch. Perform gating: when The above statistics are updated if the value is not lower than the preset threshold; otherwise, they remain unchanged. The values remain unchanged to avoid the adverse effects of low-quality signals on the statistics. The entire process does not involve backpropagation of network weights and is a zero-gradient adaptive process.
[0031] In one embodiment of the present invention, in step S7, the individual baseline calibration module 2024 performs individualized calibration on the student network output after TTA-BN adaptation for each subject, using a small amount of cuff reference blood pressure data. Specifically, optionally, the individual affine calibration can be implemented according to the following relationship: ; In the formula, The results are the blood pressure estimation results after TTA-BN adaptation. For the calibrated blood pressure estimation results, For individual gain parameters, For individual offset parameters, For the subject in Reference blood pressure values measured via cuff at specific time points. Solving using least squares or robust regression with regularization terms. This allows for the alignment of the subject's baseline blood pressure with its fluctuation range. When a limited number of reference samples are available, quantile-based quantile analysis can optionally be used. Affine (Q) A) Calibration: Using the errors between multiple quantile estimates and the corresponding quantile reference blood pressure as constraints, a similar objective function is constructed. Robust estimation is performed, but this invention does not limit this.
[0032] As one embodiment of the present invention, in step S8, in clinical or daily application scenarios, the wearable device continuously collects the user's PPG signal and performs preprocessing and window division according to steps S1 and 2. The DQI(w) of each time window is calculated according to step S2. The time windows are sequentially input into the student network after teacher-student distillation training, test batch normalization adaptation, and individual affine / quantile calibration, and the corresponding systolic and diastolic blood pressure estimation results are output in real time. For windows with significantly low DQI, it is possible to choose not to output the blood pressure estimation value, or to output the estimation value with a low confidence label to indicate that the current signal quality is insufficient, thereby realizing continuous, cuffless blood pressure monitoring with quality indication.
[0033] Optionally, to further monitor the inference stability of the method of the present invention across different devices, populations, and application scenarios, the present invention may also introduce a feature importance analysis module based on SHAP. Specifically, the data processing unit can calculate the SHAP contribution of each of the PPG morphological and statistical features used for blood pressure estimation on a preset background sample set, and statistically obtain the global feature importance vector under different data domains; then, based on the similarity of feature importance between different data domains, the model's "which features to use for decision-making" in each domain is quantitatively evaluated. When a significant deviation is found between the feature importance distribution of a target domain and the source domain, a prompt can be issued indicating that the model needs to be recalibrated or re-evaluated, so as to improve the interpretability and safety of the method of the present invention in clinical and long-term monitoring from both the perspectives of numerical error and inference basis.
[0034] Example 2 The present invention also provides a cuffless continuous blood pressure estimation device, comprising: The first processing module is used to acquire PPG signals; The second processing module is used for preprocessing the PPG signal; The third processing module is used to calculate the Data Quality Index (DQI) and perform quality gating on the preprocessed PPG signal; The fourth processing module is used to construct the teacher-student distillation network; The fifth processing module is used for DQI-weighted joint distillation and supervised training of teacher-student distillation networks; The sixth processing module is used to perform unlabeled distribution adaptation on unlabeled data in the target domain using time-batch normalized TTA-BN based on the trained teacher-student distillation network. The seventh processing module is used to perform individual baseline calibration based on affine / quantile mapping on the student network output after TTA-BN adaptation; The eighth processing module is used for online cuffless blood pressure estimation and quality prompt output.
[0035] As one embodiment of the present invention, the second processing module preprocesses the PPG signal by resampling, bandpass filtering, detrending, and normalization.
[0036] As one embodiment of the present invention, the fourth processing module constructs a teacher network with three-channel pulse signals (PPG, VPG, and APG) as input. The teacher network adopts a one-dimensional convolution and encoder-decoder structure, and uses multi-scale convolution, downsampling, residual blocks, and skip connections to extract multi-scale pulse morphology features, outputting a continuous blood pressure waveform aligned with invasive arterial pressure, as well as pulsatile systolic and diastolic blood pressure statistics. A channel elastic UNet-1D student network is constructed with single-channel PPG or few-channel pulse signals as input. The student network sets feature alignment positions in multiple intermediate layers of the encoder, and simultaneously provides blood pressure waveform and SBP / DBP multiquantile estimates at the output.
[0037] Example 3 The present invention also provides a cuffless blood pressure continuous estimation system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a cuffless blood pressure continuous estimation method when executed by the processor.
[0038] Example 4 The present invention also provides a storage medium storing a computer program that executes a cuffless blood pressure continuous estimation method when running.
[0039] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for continuous estimation of blood pressure without a cuff, characterized in that, include: Step S1: Acquire PPG signal; Step S2: Preprocess the PPG signal; Step S3: Calculate the Data Quality Index (DQI) and perform quality gating on the preprocessed PPG signal; Step S4: Construct a teacher-student distillation network; Step S5: DQI-weighted joint distillation and supervised training teacher-student distillation network; Step S6: Based on the trained teacher-student distillation network, perform unlabeled distribution adaptation on the unlabeled data in the target domain using time-batch normalized TTA-BN. Step S7: Perform individual baseline calibration based on affine / quantile mapping on the student network output after TTA-BN adaptation; Step S8: Online cuffless blood pressure estimation and quality feedback output.
2. The cuffless blood pressure continuous estimation method as described in claim 1, characterized in that, Preprocessing of PPG signals includes: resampling, bandpass filtering, detrending, and normalization.
3. The cuffless blood pressure continuous estimation method as described in claim 2, characterized in that, In step S4, a teacher network is constructed with three-channel pulse signals (PPG, VPG, and APG) as input. The teacher network adopts a one-dimensional convolution and encoder-decoder structure, and uses multi-scale convolution, downsampling, residual blocks, and skip connections to extract multi-scale pulse morphology features. It outputs a continuous blood pressure waveform aligned with invasive arterial pressure, as well as pulsatile systolic and diastolic blood pressure statistics. A channel elastic UNet-1D student network is constructed with single-channel PPG or few-channel pulse signals as input. The student network sets feature alignment positions in multiple intermediate layers of the encoder and simultaneously provides blood pressure waveform and SBP / DBP multiquantile estimates at the output.
4. A cuffless continuous blood pressure estimation device, characterized in that, include: The first processing module is used to acquire PPG signals; The second processing module is used for preprocessing the PPG signal; The third processing module is used to calculate the Data Quality Index (DQI) and perform quality gating on the preprocessed PPG signal; The fourth processing module is used to construct the teacher-student distillation network; The fifth processing module is used for DQI-weighted joint distillation and supervised training of teacher-student distillation networks; The sixth processing module is used to perform unlabeled distribution adaptation on unlabeled data in the target domain using time-batch normalized TTA-BN based on the trained teacher-student distillation network. The seventh processing module is used to perform individual baseline calibration based on affine / quantile mapping on the student network output after TTA-BN adaptation; The eighth processing module is used for online cuffless blood pressure estimation and quality prompt output.
5. The cuffless continuous blood pressure estimation device as described in claim 4, characterized in that, The second processing module preprocesses the PPG signal, including resampling, bandpass filtering, detrending, and normalization.
6. The cuffless continuous blood pressure estimation device as described in claim 2, characterized in that, The fourth processing module constructs a teacher network with PPG, VPG, and APG three-channel pulse signals as input. The teacher network adopts a one-dimensional convolution and encoder-decoder structure, and uses multi-scale convolution, downsampling, residual blocks, and skip connections to extract multi-scale pulse morphology features. It outputs a continuous blood pressure waveform aligned with invasive arterial pressure, as well as pulsatile systolic and diastolic blood pressure statistics. A channel elastic UNet-1D student network is constructed with single-channel PPG or few-channel pulse signals as input. The student network sets feature alignment positions in multiple intermediate layers of the encoder and simultaneously provides blood pressure waveform and SBP / DBP multiquantile estimates at the output.
7. A cuffless continuous blood pressure estimation system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the cuffless blood pressure continuous estimation method as described in any one of claims 1-3 when executed by the processor.
8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the cuffless blood pressure continuous estimation method as described in any one of claims 1-3.