Sleeveless blood pressure estimation method and device based on wearable equipment
Patent Information
- Application Number
- CN202512002076.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-29
- Publication Date
- 2026-03-17
Smart Images
Figure CN121667655A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of health monitoring technology, specifically relating to a cuffless blood pressure estimation method and device based on wearable devices, applicable to real-time health monitoring scenarios of low-power wearable devices such as smart bracelets and smartwatches. Background Technology
[0002] In recent years, with the development of sensor technology, wearable device technology, and artificial intelligence algorithms, health monitoring technology has gradually extended from traditional clinical medical environments to personal health management, and the types of health monitoring devices have become increasingly diverse. From professional-grade medical monitoring equipment to lightweight wearable devices, various products have their own characteristics in terms of monitoring accuracy, stability, battery life, and ease of use, and play an important role in cardiovascular monitoring, chronic disease management, and daily exercise and health.
[0003] Professional medical-grade devices, such as widely used electrocardiogram (ECG) monitors, can provide high-precision, high-sampling-rate, multi-lead cardiac electrical signal acquisition, enabling the monitoring of serious diseases such as arrhythmias and myocardial ischemia, and providing reliable evidence for clinical diagnosis. However, these devices are usually bulky, rely on multiple electrodes, and are expensive, making them unsuitable for long-term wear in everyday home environments or mobile settings. Furthermore, prolonged contact between electrodes and skin can lead to discomfort, skin irritation, or even electrode detachment, thus affecting signal quality and data stability.
[0004] Pulse oximeters, another common monitoring device, estimate blood oxygen saturation (SpO2) and pulse rate based on differences in light absorption. They are widely used in respiratory disease detection, sleep monitoring, and postoperative management. While small and easy to use, they have limited functionality and cannot cover more important cardiovascular risk indicators such as blood pressure and heart rate variability (HRV). Furthermore, pulse oximeters require precise finger placement, making true "continuous wear" and long-term real-time monitoring difficult.
[0005] Smart bracelets and smartwatches, as the fastest-growing type of health monitoring devices in recent years, generally incorporate accelerometers, optical heart rate sensors (PPG), and temperature sensors to record basic physiological indicators such as steps, exercise intensity, heart rate, and sleep quality. PPG technology, due to its low cost, small size, and ease of integration, has become the primary source of physiological signals in wearable devices. However, PPG is easily affected by factors such as ambient light disturbances, loose skin contact, and motion artifacts, resulting in low signal-to-noise ratios and significant waveform distortion, thus limiting its reliable application in higher-precision health indicators (such as cuffless blood pressure estimation).
[0006] Smart chest straps offer high wearing stability and ECG signal quality, and are commonly used for athlete training monitoring, providing more accurate heart rate and respiration information than wristbands. However, chest straps are uncomfortable to wear, unsuitable for long-term wear or daily use, and difficult to integrate into the general consumer market. Smart head-mounted devices (such as ear clip-on and glasses-style monitoring devices) can collect complex signals such as EEG and EOG, but their wearing methods are more invasive, consume more power, and their stability depends on wearing posture. Therefore, they are still unlikely to replace the widespread use of wrist-worn devices in daily health monitoring.
[0007] While the aforementioned devices can provide physiological data across various dimensions, most still rely on traditional cuff measurements for blood pressure, a key cardiovascular health indicator, failing to achieve truly continuous, interference-free, and low-power monitoring. Cuffless blood pressure estimation techniques have received significant attention in recent years, primarily based on PPG signal waveform characteristics, pulse transit time (PTT) between ECG and PPG, or by using deep learning models to directly regress blood pressure from the original waveform. However, existing methods still suffer from the following prominent problems: (1) Strong dependence on feature points: Many methods rely on the R peak of ECG, the waveform starting point of PPG, second derivative features, etc., but the signal quality fluctuates greatly in wearable scenarios, and feature points are difficult to extract stably, resulting in unstable blood pressure estimation.
[0008] (2) Large computational load of models: Deep neural networks (CNN, ResNet, Transformer, etc.) can achieve high accuracy on the server side, but the number of parameters is huge and the computational load is high, which is not suitable for deployment in the low-power hardware environment of smart bracelets.
[0009] (3) Weak cross-population generalization ability: There are significant differences in vascular elasticity, skin color, skin thickness and other factors among different individuals, making it difficult for a single model to maintain consistent performance across different users.
[0010] (4) Not robust to noise: PPG signals are significantly deformed in motion and other scenarios, and traditional algorithms are difficult to cope with strong noise conditions.
[0011] (5) Unstable continuous output: The predicted sequence may have instantaneous jumps, which affects the user experience and makes it difficult to support health trend analysis.
[0012] In addition, cuffless blood pressure estimation still faces practical engineering challenges such as poor continuous monitoring experience and high real-time requirements for low-power devices. Summary of the Invention
[0013] To address the problems existing in the prior art, the present invention provides a cuffless blood pressure estimation method and device based on wearable devices.
[0014] To achieve the above objectives, the present invention provides the following solution: A cuffless blood pressure estimation method based on wearable devices includes: Acquire photoplethysmography (PPG) data and electrocardiogram (ECG) data collected by the user's wearable device; The PPG and ECG data are preprocessed; Based on the preprocessed PPG data and filtered ECG data, multiple time segments are obtained by dividing the data into preset window lengths; For each time segment, perform multi-level wavelet decomposition and extract wavelet statistical features to obtain statistical features; The statistical features are combined with the user's vital signs data to construct a feature vector, and the feature vector is input into a lightweight multi-resolution neural network model to output systolic and diastolic blood pressure data for the corresponding time segment, thereby obtaining the user's continuous blood pressure estimation results.
[0015] Preferably, the preprocessing of the PPG data and ECG data includes: removing the mean, removing the trend, and performing bandpass filtering on the PPG data and ECG data.
[0016] As a preferred approach, multi-level wavelet decomposition is performed on each time segment, and the statistical characteristics of the energy, standard deviation, and mean absolute value of the detail coefficients at each level are calculated.
[0017] As a preferred option, the lightweight multi-resolution neural network model includes: a fully connected regression module and a MultiRes convolution module.
[0018] The present invention also provides a cuffless blood pressure estimation device based on a wearable device, comprising: The first processing module is used to acquire photoplethysmography (PPG) data and electrocardiogram (ECG) data collected by the user's wearable device. The second processing module is used to preprocess the PPG data and ECG data; The third processing module is used to divide the preprocessed PPG data and filtered ECG data into multiple time segments according to a preset window length; The fourth processing module is used to perform multi-level wavelet decomposition on each time segment and extract wavelet statistical features to obtain statistical features; The fifth processing module is used to combine the statistical features with the user's vital signs data to construct a feature vector, and input the feature vector into a lightweight multi-resolution neural network model to output the systolic and diastolic blood pressure data for the corresponding time segment, thereby obtaining the user's continuous blood pressure estimation results.
[0019] Preferably, the second processing module preprocesses the PPG data and ECG data by performing mean removal, trend removal, and bandpass filtering on the PPG data and ECG data.
[0020] As a preferred embodiment, the fourth processing module performs multi-level wavelet decomposition on each time segment and calculates the statistical characteristics of the energy, standard deviation, and mean absolute value of the detail coefficients at each level.
[0021] As a preferred option, the lightweight multi-resolution neural network model includes: a fully connected regression module and a MultiRes convolution module.
[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes multimodal physiological data collected by wearable devices to achieve continuous prediction of systolic blood pressure (SBP) and diastolic blood pressure (DBP) through wavelet domain feature extraction and lightweight deep neural networks. This invention offers advantages such as low computational complexity, high real-time performance, and applicability to low-power devices, meeting the needs of smartwatches for continuous and non-invasive blood pressure assessment in daily health monitoring. Attached Figure Description
[0023] 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.
[0024] Figure 1 This is a flowchart of a cuffless blood pressure estimation method based on wearable devices according to an embodiment of the present invention; Figure 2 A schematic diagram of the results of a smartwatch-based ECG / PPG non-invasive blood pressure estimation system. Detailed Implementation
[0025] 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.
[0026] 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.
[0027] Example 1 like Figure 1As shown, this invention provides a cuffless blood pressure estimation method based on wearable devices, enabling continuous calculation and real-time estimation of user blood pressure changes. This provides low-power, non-invasive, and long-term usable blood pressure assessment capabilities for health monitoring on smartwatches, including: Step S1: Obtain the user's wrist physiological monitoring data, wherein the wrist physiological monitoring data includes PPG (Photoplethysmography) data, ECG (Electrocardiogram) data, and accelerometer data; Step S2: Preprocess the ECG and PPG data to remove trend terms and environmental noise, and obtain stable heart rate-related waveforms through bandpass filtering; Step S3: Based on the wrist signal, divide the continuous data into time segments in 10-second windows; Step S4: Extract multi-scale detail coefficients from each segment of ECG and PPG signal using the multi-level wavelet decomposition method; Step S5: Calculate statistical characteristics such as energy, standard deviation, and mean absolute value based on detail coefficients; Step S6: Construct a feature vector by combining the user's height, weight, BMI, and other physical characteristics; Step S7: Standardize the feature vectors; Step S8: After completing the feature standardization process, input the feature vector into the lightweight multi-resolution neural network model to obtain the blood pressure prediction result for the corresponding window. Step S9: Perform 3σ anomaly detection and moving average smoothing based on the predicted result sequence; Step S10: Obtain the final systolic blood pressure SBP and diastolic blood pressure DBP assessment data; In one embodiment of the present invention, in step S2, the input ECG signal is set to... PPG signal is The filter employs a fourth-order Butterworth bandpass structure, with a bandwidth range of [missing information]. Hz is used to preserve the main frequency components related to heartbeat. The filtering operation can be expressed as: ; in, These are the filter coefficients; The present invention further performs a detrending operation on the signal to make it meet the stationarity requirements of the subsequent wavelet transform: ; The ECG and PPG signals preprocessed as described above can significantly reduce low-frequency drift and high-frequency noise, thereby improving the reliability of feature extraction.
[0028] In one embodiment of the present invention, in step S3, a fixed-duration sliding window is used to segment the signal, with each segment being 10 seconds long. Let the sampling rate be... Window duration is Then the length of each signal sample is: ; By using a sliding window approach, this invention can output a blood pressure prediction every 10 seconds, enabling the system to perform continuous monitoring. Furthermore, each signal segment corresponds to a blood pressure label (from synchronous acquisition or an external file), facilitating model training through supervised learning.
[0029] In one embodiment of the present invention, in step S4, the present invention performs multi-level wavelet decomposition on the ECG and PPG signals within each time window. This embodiment preferably uses a seven-level decomposition and the sym4 wavelet basis to effectively express the detailed changes of the signal at different scales. For any signal... The The layer detail factor can be expressed as: ; in, For input signal, For wavelet basis functions, the number of decomposition layers is... .
[0030] As one embodiment of the present invention, in step S5, the present invention extracts three types of statistical features from each layer of detail coefficients, namely energy. Standard deviation with absolute mean The specific definition is as follows: ; ; ; ; All layers , , The sequenced concatenation forms a multi-scale statistical feature vector. .
[0031] ECG and PPG are generated respectively. These features form a multi-resolution sequence of physiological characteristics. Since ECG and PPG reflect cardiac electrical activity and peripheral vascular pulsation, respectively, their multi-scale details can comprehensively describe factors such as heart rate strength, waveform steepness, and hemodynamic feedback, which helps deep models learn the subtle relationships of blood pressure changes.
[0032] As one embodiment of the present invention, in step S6, the present invention combines the above-mentioned wavelet statistical features with user vital signs parameters (including height). ,weight Together with Body Mass Index (BMI), they form a complete feature vector, where BMI is calculated using the following formula: ; This vector contains both static individual differences and dynamic signal characteristics.
[0033] In one embodiment of the present invention, step S7 involves standardizing all features to convert them to a uniform scale, thereby preventing differences in scale from affecting model convergence. The standardization method is as follows: ; In one embodiment of the present invention, in step S8, the standardized feature vector is first input into the fully connected regression module for scaling transformation, and then enters the MultiRes convolution module, specifically as follows: First, the fully connected regression module processes the standardized feature vectors... Zero-mean normalization is performed using the following formula: ; In the formula, and These are the mean and standard deviation of each feature dimension in the training set, respectively; the normalized features are input into the MultiRes convolutional module, and the output of the first fully connected layer is: ; in, , For trainable parameters, This is the activation function.
[0034] The MultiRes convolution module for Two levels of convolution extraction at different scales are performed, namely: ; ; in This represents a one-dimensional convolution operation. These correspond to different receptive field convolutional kernels in the MultiRes structure; then the convolutional output is input into the second fully connected layer: ; Finally, the predicted systolic and diastolic blood pressure for the corresponding time segment are obtained through the output layer: ; in, ; Furthermore, the lightweight multi-resolution neural network model is trained using a K-fold cross-validation strategy, and feature normalization and parameter updates are performed based on the training set distribution. The specific steps include: cross-dividing all samples, and at the k-fold... As a compromise, the training and test sets are as follows: ; ; in, and These are the training and testing index sets, respectively. Before each fold of training begins, the input features are standardized based on the training set, using the following normalization formula: ; in, and The first The mean and standard deviation of each feature in the training set are calculated. Batch training is used, and the prediction output for each training batch is: ; And using the mean squared error as the loss function, it is expressed as: ; in, The number of samples in the training set; For the first The actual blood pressure values of each sample.
[0035] Model parameters The update is performed using the Adam optimization algorithm, whose gradient descent steps are as follows: ; ; ; ; in, For learning rate, These are the exponential decay coefficients of the first and second moments. To prevent division by zero of constants.
[0036] After training, the test set features are standardized according to the statistics of the training set: ; And input the data into the model for inference: ; In one embodiment of the present invention, in step S9, to improve the smoothness and robustness of the continuous blood pressure sequence, abnormal prediction points are removed using a threshold of three times the standard deviation. The detection criteria are as follows: ; in, and These are the mean and standard deviation of the predicted sequence, respectively; a moving average is used to compensate for outlier samples, defined as: ; By employing a lightweight multi-resolution convolutional structure, fully connected regression units, and an outlier smoothing strategy, this invention achieves real-time estimation of continuous blood pressure in a low-power environment, improving the model's generalization ability and output stability.
[0037] In one embodiment of the present invention, in step S10, error indices are calculated based on the actual blood pressure and the predicted blood pressure, including mean absolute error (MAE), root mean square error (RMSE), and bias, expressed as follows: ; ; ; At this point, the lightweight blood pressure estimation process driven by ECG / PPG wavelet features described in this invention has been fully implemented, achieving continuous, real-time, and non-invasive blood pressure prediction capabilities suitable for smartwatch scenarios through multi-scale feature representation and lightweight model structure.
[0038] This invention includes signal preprocessing, segmented sliding windowing, wavelet feature extraction, individual vital sign fusion, deep learning modeling based on MultiRes convolutional modules, K-fold cross-validation training and testing, and outlier smoothing mechanisms. Its core idea is to analyze the dynamic physiological characteristics of PPG and ECG and construct a feature system adapted to wearable devices without relying on traditional pneumatic cuffs, enabling the deep model to effectively infer blood pressure trends from minute changes in physiological signals. This invention can achieve real-time estimation of blood pressure trends using low-power devices such as smartwatches; it is cuff-free, suitable for continuous monitoring scenarios, and has the advantages of small parameter count and high computational efficiency. Due to the use of multi-scale feature and vital sign fusion, it can output smooth and stable blood pressure prediction results under normal conditions, demonstrating good engineering feasibility.
[0039] Figure 2This diagram illustrates the continuous blood pressure prediction results obtained in this embodiment based on ECG / PPG multi-scale wavelet features and a lightweight MultiRes network model. Based on a dataset constructed from 29 subjects, this invention systematically evaluated the proposed non-invasive blood pressure estimation method based on ECG / PPG wavelet features and a lightweight MultiRes network using leave-one-out cross-validation.
[0040] In the experiment, one subject was used as the test set for each fold, while the remaining subjects were used for training to simulate cross-population generalization ability in real-world usage scenarios. Results showed that the predictive performance of each fold was generally stable. The mean absolute error of systolic blood pressure (SBP) was mostly concentrated between 3–6 mmHg, while diastolic blood pressure (DBP) typically remained within the range of 2.5–4.0 mmHg, indicating that the model could effectively extract blood pressure-related information from the wavelet statistical features of wrist ECG and PPG. In several folds with higher signal quality, the SBP prediction error could decrease to around 3 mmHg, while the DBP prediction error reached a minimum of 2.5 mmHg, indicating that the model could achieve a high level of fitting accuracy on some subjects. Regarding bias indicators, the prediction biases of SBP and DBP for most folds were within ±2 mmHg, with a bias close to 0, indicating that the model as a whole did not have a systematic tendency to overestimate or underestimate, meaning that the predicted values were statistically highly consistent with the actual blood pressure. A few folds showed a 3–5 mmHg shift in SBP, presumably related to individual differences, motion state, or signal noise, but this did not significantly affect the overall performance.
[0041] Statistical summarization of the 29-fold results revealed that the overall average MAE of the model was 4.977 mmHg for SBP and 3.424 mmHg for DBP, while the average RMSE was 6.575 mmHg for SBP and 4.343 mmHg for DBP. The biases were 0.732 mmHg for SBP and 0.028 mmHg for DBP. The error distribution shows that the MAE and RMSE values for SBP and DBP are quite close, indicating no significant extreme value errors during prediction. Combined with 3σ anomaly detection and moving average smoothing strategies, the blood pressure sequence output by the model maintained good continuity, stability, and noise resistance. Since the error of DBP was generally better than that of SBP, this is consistent with the physiological mechanism that PPG waveforms are more stable in diastole and more sensitive to changes in vascular compliance, further validating the effectiveness of multi-scale wavelet features in pulse wave analysis. The lightweight MultiRes network structure proposed in this invention achieves non-invasive blood pressure prediction accuracy with engineering application value despite having a much smaller parameter scale than deep convolutional networks or recurrent neural networks, and also possesses low power consumption characteristics suitable for deployment in smartwatches. All experimental results demonstrate that this invention maintains stable generalization ability in cross-subject prediction tasks, with low error levels, small bias, and stable output, meeting the real-time and accuracy requirements of continuous blood pressure monitoring devices, and providing a practical non-invasive blood pressure estimation solution for wearable health monitoring.
[0042] Example 2 The present invention also provides a cuffless blood pressure estimation device based on a wearable device, comprising: The first processing module is used to acquire photoplethysmography (PPG) data and electrocardiogram (ECG) data collected by the user's wearable device. The second processing module is used to preprocess the PPG data and ECG data; The third processing module is used to divide the preprocessed PPG data and filtered ECG data into multiple time segments according to a preset window length; The fourth processing module is used to perform multi-level wavelet decomposition on each time segment and extract wavelet statistical features to obtain statistical features; The fifth processing module is used to combine the statistical features with the user's vital signs data to construct a feature vector, and input the feature vector into a lightweight multi-resolution neural network model to output the systolic and diastolic blood pressure data for the corresponding time segment, thereby obtaining the user's continuous blood pressure estimation results.
[0043] As one embodiment of the present invention, the preprocessing of the PPG data and ECG data by the second processing module includes: removing the mean, removing the trend, and performing bandpass filtering on the PPG data and ECG data.
[0044] As one embodiment of the present invention, the fourth processing module performs multi-level wavelet decomposition on each time segment and calculates the statistical characteristics of the energy, standard deviation and mean absolute value of the detail coefficients of each level.
[0045] As one embodiment of the present invention, the lightweight multi-resolution neural network model includes: a fully connected regression module and a MultiRes convolution module.
[0046] 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 to the technical solutions of the present invention by those skilled in the art 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 cuffless blood pressure estimation method based on wearable devices, characterized in that, include: Acquire photoplethysmography (PPG) data and electrocardiogram (ECG) data collected by the user's wearable device; The PPG and ECG data are preprocessed; Based on the preprocessed PPG data and filtered ECG data, multiple time segments are obtained by dividing the data into preset window lengths; For each time segment, perform multi-level wavelet decomposition and extract wavelet statistical features to obtain statistical features; The statistical features are combined with the user's vital signs data to construct a feature vector, and the feature vector is input into a lightweight multi-resolution neural network model to output systolic and diastolic blood pressure data for the corresponding time segment, thereby obtaining the user's continuous blood pressure estimation results.
2. The cuffless blood pressure estimation method based on wearable devices as described in claim 1, characterized in that, Preprocessing the PPG and ECG data includes removing the mean, detrending, and bandpass filtering from the PPG and ECG data.
3. The cuffless blood pressure estimation method based on wearable devices as described in claim 2, characterized in that, For each time segment, perform multi-level wavelet decomposition and calculate the statistical characteristics of the energy, standard deviation, and mean absolute value of the detail coefficients at each level.
4. The cuffless blood pressure estimation method based on wearable devices as described in claim 3, characterized in that, The lightweight multi-resolution neural network model includes a fully connected regression module and a MultiRes convolution module.
5. A cuffless blood pressure estimation device based on a wearable device, characterized in that, include: The first processing module is used to acquire photoplethysmography (PPG) data and electrocardiogram (ECG) data collected by the user's wearable device. The second processing module is used to preprocess the PPG data and ECG data; The third processing module is used to divide the preprocessed PPG data and filtered ECG data into multiple time segments according to a preset window length; The fourth processing module is used to perform multi-level wavelet decomposition on each time segment and extract wavelet statistical features to obtain statistical features; The fifth processing module is used to combine the statistical features with the user's vital signs data to construct a feature vector, and input the feature vector into a lightweight multi-resolution neural network model to output the systolic and diastolic blood pressure data for the corresponding time segment, thereby obtaining the user's continuous blood pressure estimation results.
6. The cuffless blood pressure estimation device based on a wearable device as described in claim 5, characterized in that, The second processing module preprocesses the PPG and ECG data by performing mean removal, trend removal, and bandpass filtering on the PPG and ECG data.
7. The cuffless blood pressure estimation device based on a wearable device as described in claim 6, characterized in that, The fourth processing module performs multi-level wavelet decomposition on each time segment and calculates the statistical characteristics of the energy, standard deviation, and mean absolute value of the detail coefficients at each level.
8. The cuffless blood pressure estimation device based on a wearable device as described in claim 7, characterized in that, The lightweight multi-resolution neural network model includes a fully connected regression module and a MultiRes convolution module.