Dual-channel ballistocardiogram non-invasive heart rate and blood pressure estimation device and method

Abstract

This invention discloses a non-invasive heart rate and blood pressure estimation device and method based on dual-channel cardiac impact signal, including a measuring chair, a back BCG sensor, a hip BCG sensor, a signal acquisition module, a blood pressure estimation processing module, and a communication module. The measuring chair includes a seat and a backrest, the angle between the backrest and the seat being adjustable and fixed. The back BCG sensor is disposed in the backrest for acquiring back BCG signals. The hip BCG sensor is disposed in the seat or cushion for acquiring hip BCG signals. By combining an adjustable-angle seat, simultaneous acquisition of back-hip dual-channel BCG signals, window-level quality assessment, dual-channel phase difference / delay feature extraction, and a blood pressure estimation model, this invention enables cuffless estimation of systolic and diastolic blood pressure in scenarios such as sitting or during exercise recovery.

Description

Technical Field

[0001] This invention designs a medical auxiliary device, specifically a dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device and method. Background Technology

[0002] Blood pressure is an important physiological indicator reflecting cardiovascular function. Traditional cuff-type blood pressure monitors mainly rely on oscillometric or Korotkoff sound methods, obtaining systolic and diastolic blood pressure by inflating and deflating the cuff to compress the upper arm artery. While these methods are technically mature and widely used clinically, they typically only provide blood pressure values ​​at discrete time points. The measurement process requires wearing a cuff and applying pressure, which can be uncomfortable and makes long-term, low-interference, and continuous monitoring difficult. Therefore, in recent years, various cuffless or low-burden blood pressure measurement technologies have emerged, including pulse arrival time methods based on electrocardiograms and photoplethysmography (PPG), blood pressure estimation methods based on impact cardiography signals, non-invasive physiological signal acquisition methods based on mattresses or chairs, and health monitoring devices that integrate blood pressure monitors into chairs or mattresses.

[0003] Among existing cuffless blood pressure monitoring solutions, the most typical is the chair-based system that calculates pulse arrival time (PAT) based on electrocardiogram (ECG) and photoplethysmography (PPG) signals. These systems typically place ECG electrodes and PPG sensors in the chair armrests. Users, while seated, need to contact the electrodes with both hands and place their fingers or thumbs on the PPG sensors. Blood pressure is estimated by the time difference between the ECG R wave and the characteristic points of the PPG pulse wave. The chair-based cuffless blood pressure monitoring system proposed by Tang et al. calculates PAT using ECG and PPG signals and estimates blood pressure through a single individualized calibration. This system places the ECG electrodes and PPG sensors in the armrest area, requiring the user's hands to remain in contact with the electrodes and PPG sensors, and still requires individualized calibration before measurement. While this method reduces cuff pressure, it still relies on ECG electrodes and PPG sensors, and suffers from issues such as fixed contact points, hand posture, and PPG being affected by peripheral blood flow and contact pressure. Furthermore, PAT-based methods usually require individualized calibration, and the calibration process and parameter stability can affect long-term usability.

[0004] Existing technologies also employ cameras or image-based photoplethysmography (PPG) to achieve non-contact blood pressure measurement. For example, CN116433614A discloses a non-contact blood pressure measurement system that acquires facial images, extracts regions of interest (ROIs) from the face, and uses methods such as optimization variational mode decomposition and Euler image amplification to obtain imaging PPG signals. These signals are then further used to obtain non-contact cardiac impact signals and pulse transit time to estimate blood pressure. While this method eliminates the need for human contact with the sensor, its signal quality is easily affected by lighting conditions, facial movement, camera position, skin area selection, and environmental interference, placing high demands on the measurement environment and user cooperation. For home or rehabilitation scenarios in a seated position, relying solely on facial video signals is insufficient to reliably obtain high-quality body motion signals directly related to cardiac mechanical activity, and the impact of the contact coupling between the human body and the seat on the stability of blood pressure estimation remains unresolved.

[0005] Some existing technologies employ cuffless continuous blood pressure measurement methods, simultaneously acquiring BCG and PPG signals and estimating blood pressure through signal quality assessment, feature point detection, PTT features, BCG features, PPG features, and individual statistical data. For example, the method disclosed in CN117064354A requires simultaneous acquisition of BCG and PPG signals, extracting features such as PTT, BCG, and PPG from signals that meet quality requirements, and then combining this with calibrated blood pressure to calculate the blood pressure value. Although this type of method introduces BCG signals and signal quality assessment, it still requires the participation of the PPG channel, resulting in high requirements for system structure and simultaneous acquisition, and necessitates stable feature point detection. When PPG contact is unstable, peripheral blood flow changes significantly, or feature point identification is incorrect, the accuracy of blood pressure estimation can easily decrease. In contrast, non-contact blood pressure estimation based solely on the seat-embedded dual-channel BCG signal can reduce reliance on finger PPG, ECG electrodes, or additional wearable sensors.

[0006] Regarding BCG-based blood pressure estimation, Seok et al. proposed a chair-based dual-channel BCG and convolutional neural network-based blood pressure monitoring system. This system uses PVDF thin-film sensors on the backrest and seat surface to collect dual-channel BCG signals from the back and thighs. Empirical Mode Decomposition (EMD) and Hilbert Transform (HLD) are used to obtain phase information, and a convolutional neural network regression model is further used to estimate blood pressure. This study demonstrates the feasibility of using dual-channel BCG signals for recovery-phase blood pressure estimation. However, their approach primarily focuses on the dual-channel BCG phase signals and deep learning model within a fixed chair structure. It does not optimize the mechanical coupling quality between the human body and the sensor under different backrest angles, nor does it consider factors such as backrest angle, dual-channel signal quality, rhythm consistency, and cross-channel delay stability as applicability criteria before blood pressure measurement. For practical chair-based blood pressure measurement, different backrest angles alter the contact pressure and mechanical transmission path between the back, buttocks, and sensor, thus affecting BCG signal quality and blood pressure estimation stability—a factor that current solutions fail to adequately address.

[0007] There are also many existing mattress-type or supine blood pressure monitoring devices. For example, mattress-type flexible sensors used for dynamic continuous blood pressure monitoring typically collect ECG and BCG signals from the user in a supine position through multiple layers of flexible fabric, electrode layers, and buffer layers for subsequent continuous blood pressure monitoring. This type of solution is suitable for sleep or bed rest and has the advantage of good comfort. However, its measurement scenario is mainly supine, and the distribution of human body force, sensor placement, and signal coupling method are significantly different from those in a sitting chair scenario. Therefore, it cannot directly solve the problems of optimal measurement angle, signal quality gating, and blood pressure estimation for dual-channel BCG of the back and buttocks in a sitting position.

[0008] Furthermore, existing blood pressure measurement chairs or multi-functional health monitoring chairs mostly integrate traditional blood pressure monitors, cuffs, air pumps, pressure sensors, displays, or storage structures into the chair. For example, CN202489576U discloses a multi-functional chair with blood pressure measurement function, which has weight, body temperature, and blood pressure measuring devices installed on the chair body. The blood pressure measuring device still includes an inflatable cuff or wristband, a pressurizing air pump, an exhaust valve, and a pressure sensor. Such solutions can make the measurement posture more standardized, but they are still essentially cuff-type or wristband-type blood pressure measurement. They do not achieve cuffless, low-interference blood pressure estimation based on BCG signals from the seat contact interface, nor do they involve back-hip dual-channel signal fusion, backrest angle optimization, or blood pressure estimation algorithms.

[0009] There are also some publicly available solutions such as "blood pressure measurement chairs," "standardized blood pressure measurement chairs," "reclining chairs that can detect weight, heart rate, and blood pressure," and "hypertension prevention chairs." These solutions mainly focus on mounting the blood pressure monitor on a chair, standardizing arm height, and helping users maintain a standard measurement posture, or combining the detection of multiple parameters such as weight, heart rate, and body temperature. These solutions help improve the ease of use of traditional blood pressure monitors to some extent, but most still rely on cuffs, wristbands, or external blood pressure modules. The measurement process still involves obvious operating steps and a sense of pressure, and cannot achieve continuous or quasi-continuous, imperceptible data collection while the user is naturally seated.

[0010] Existing fiber optic sensing-based vital sign acquisition devices or methods typically utilize fiber optic sensors to collect physiological signals such as heartbeat, respiration, and body movement, offering advantages such as resistance to electromagnetic interference, high sensitivity, and suitability for long-term monitoring. For example, some fiber optic vital sign monitoring systems can extract vital sign information through fiber optic sensing modules and signal processing modules. This type of technology provides a good foundation for non-invasive physiological monitoring; however, most solutions primarily focus on vital sign acquisition or sleep monitoring, failing to establish a complete closed loop for seated blood pressure measurement, encompassing "seat structure—optimal backrest angle—dual-channel BCG quality evaluation—blood pressure estimation model." Furthermore, they do not clearly address the differences in analyzability of dual-channel BCG signals and the stability of blood pressure measurement under different backrest angles.

[0011] In summary, existing technologies have the following main shortcomings: First, most traditional blood pressure measurement chairs still rely on cuffs or wristbands, requiring users to actively wear them and wait for inflation and deflation, making it difficult to achieve non-intrusive, low-interference measurement. Second, some cuffless blood pressure methods require ECG, PPG, cameras, or multiple heterogeneous sensors to work together, resulting in a complex system structure that is easily affected by hand contact pressure, light, motion artifacts, or peripheral blood flow. Third, existing BCG or dual-channel BCG blood pressure estimation schemes focus primarily on signal processing or model prediction itself, lacking specific optimization for seat back angle, human-sensor mechanical coupling, dual-channel signal quality, and cross-channel delay stability. Fourth, mattress-type or reclining schemes cannot be directly applied to seated chair measurement scenarios. Fifth, existing methods rarely establish signal quality gating and angle applicability judgment mechanisms before blood pressure estimation, making it easy to produce unstable estimates under low-quality windows or unsuitable sitting postures.

[0012] Therefore, it is necessary to propose a non-intrusive blood pressure measurement chair and method for seated scenarios, which combines an adjustable backrest structure, a dual-channel fiber optic sensor for the back and hips, backrest angle optimization, dual-channel BCG signal quality evaluation, cross-channel time delay / phase feature extraction, and a small-sample blood pressure estimation model. This allows users to complete low-burden physiological signal acquisition and blood pressure trend estimation in a natural sitting position, and improves signal stability and blood pressure estimation reliability through optimal backrest angle and quality gating. Summary of the Invention

[0013] To address the problems existing in the prior art, this invention provides a non-invasive heart rate and blood pressure estimation device and method based on dual-channel cardiac impact signals. It combines an adjustable-angle seat, simultaneous acquisition of back-hip dual-channel BCG, window-level quality assessment, dual-channel phase difference / delay feature extraction, and a blood pressure estimation model to estimate systolic and diastolic blood pressure without cuffs in scenarios such as sitting or during exercise recovery.

[0014] The technical solution of the present invention is: a dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device, including a measuring chair, a back BCG sensor, a buttock BCG sensor, a signal acquisition module, a blood pressure estimation and processing module, and a display and communication module;

[0015] The measuring chair includes a seat and a backrest, and the angle between the backrest and the seat can be adjusted and fixed arbitrarily.

[0016] The back BCG sensor is disposed in the backrest and is used to collect back BCG signals; the back BCG sensor is connected to the signal acquisition module and is used to transmit the collected back BCG signals to the signal acquisition module.

[0017] The hip BCG sensor is disposed in the seat or cushion of the chair and is used to collect hip BCG signals; the hip BCG sensor is connected to the signal acquisition module and is used to transmit the collected hip BCG signals to the signal acquisition module.

[0018] The signal acquisition module synchronously samples and receives the back BCG signal acquired by the back BCG sensor and the buttock BCG signal acquired by the buttock BCG sensor, and transmits the received signals to the blood pressure estimation processing module.

[0019] The blood pressure estimation processing module receives the back BCG signal and buttock BCG signal transmitted from the signal acquisition module and performs blood pressure estimation.

[0020] The display and communication module is used to display the blood pressure estimation output results, blood pressure change trends, and signal quality status data, and can transmit the data to an external display.

[0021] Furthermore, the blood pressure estimation processing module receives the back BCG signal and the buttock BCG signal transmitted from the signal acquisition module to perform blood pressure estimation;

[0022] The specific steps are as follows:

[0023] Step 1: Preprocess the BCG signals from the back and buttocks respectively;

[0024] Step 2: For each reference blood pressure measurement point, there are back BCG signal segments and buttock BCG signal segments within a preset time range or nearby. The back BCG signal segments and buttock BCG signal segments are segmented into multiple short-term sub-windows according to the preset short window length. The signal quality of each short-term sub-window is evaluated, and the back channel quality score, buttock channel quality score, and dual-channel combined quality score of each short-term sub-window are calculated respectively.

[0025] The dual-channel joint quality score is obtained by weighting a single-channel baseline quality term, a dual-channel rhythm consistency term, and a cross-channel delay stability term; the single-channel baseline quality term is represented by the geometric mean of the back channel quality score and the buttock channel quality score; the dual-channel rhythm consistency term is represented by the heartbeat peak matching ratio of the two channels, the correlation of the heartbeat interval sequence, and the stability of the heartbeat interval difference; the cross-channel delay stability term is represented by the standard deviation of the cross-channel peak delay and the proportion of delay anomalies.

[0026] Step 3: Employ a quality-aware adaptive feature extraction strategy. For short-term sub-windows that simultaneously meet the requirements of back channel quality score, buttock channel quality score, and combined dual-channel quality score, extract dual-channel BCG features of the back BCG signal and buttock BCG signal. For short-term sub-windows where one channel has a low quality score but the other channel still has a usable quality score, retain the single-channel BCG signal with better quality and extract the corresponding single-channel BCG auxiliary features. For short-term sub-windows where both the back channel quality score and the buttock channel quality score are low, remove the short-term sub-window or assign it a lower weight.

[0027] Step 4: Input the features extracted in Step 3 into the blood pressure estimation model. The blood pressure estimation model includes a systolic blood pressure estimation model, a diastolic blood pressure estimation model, a systolic blood pressure change estimation model relative to the individual's resting baseline, and a diastolic blood pressure change estimation model relative to the individual's resting baseline, thereby outputting the estimated systolic blood pressure value, estimated diastolic blood pressure value, systolic blood pressure change, diastolic blood pressure change, and short-term blood pressure trend curve for the corresponding reference blood pressure measurement point.

[0028] Furthermore, the dual-channel BCG features mentioned in step three include, but are not limited to, dual-channel event delay features, dual-channel phase difference features, dual-channel correlation and transmission features, dual-channel amplitude ratio features, dual-channel energy ratio features, and dual-channel image fusion features. The dual-channel event delay features include the beat-by-beat matching ratio of the back channel peak and the buttock channel peak, the mean cross-channel peak delay, the median delay, the standard deviation of the delay, the interquartile range of the delay, the proportion of delay anomalies, the slope of the delay change, and the consistency of the two-channel heartbeat interval sequence. The dual-channel phase difference features include the Hilbert instantaneous phase difference between the back channel and the buttock channel, the circular mean of the phase difference, the circular standard deviation of the phase difference, phase synchronization stability, the cross-spectral dominant frequency phase difference, the cross-spectral weighted phase difference, and the cross-channel delay converted from the phase difference. The dual-channel correlation and transmission features include the cross-correlation peak value, the cross-correlation lag, the heartbeat frequency band coherence, the inter-channel transmission gain, the inter-channel phase difference, and the inter-channel energy distribution difference.

[0029] The single-channel BCG auxiliary features mentioned in step three include heart rate, heartbeat interval, heartbeat interval variability, peak amplitude, amplitude variability, root mean square of signal, frequency domain energy, heartbeat frequency band energy, respiratory frequency band energy, envelope features, diastolic decay features, and cycle template morphology features. The diastolic decay features include the descent slope after the J peak, the area after the J peak, the amplitude of the subsequent trough, the peak-to-trough difference, the half-peak width, the energy proportion of the second half of the cycle, and the time interval from the main peak to the subsequent feature point. The cycle template morphology features include the average template amplitude, template area, template width, template skewness, template kurtosis, template stability, and the correlation or distance with the reference template after heartbeat peak alignment.

[0030] Furthermore, the features extracted in step three also include image-based features derived from the BCG signal to extract nonlinear morphology and channel coupling information; the image-based features include single-channel Gram angle field features, dual-channel cross-recursive graph features, and dual-channel fusion graph features; the dual-channel fusion graph includes back channel amplitude, hip channel amplitude, difference between the two channels, product of the two channels, and absolute difference between the two channels, and further calculates image entropy, texture statistics, row and column direction change features, diagonal features, and frequency domain texture features.

[0031] Furthermore, the features extracted in step three also include contextual features introduced in the application scenario of exercise recovery, to help the model characterize the dynamic changes in blood pressure recovery after exercise; the contextual features include, but are not limited to, the percentage of target exercise load, recovery time, blood pressure measurement sequence number, measurement window sequence number, heart rate change trend, and time position relative to the resting state.

[0032] Furthermore, in the blood pressure estimation model described in step four, during the model training phase, the system does not perform linear interpolation on adjacent cuff blood pressure measurements, but instead uses each real reference blood pressure measurement point as a supervision label; specifically as follows: each reference blood pressure measurement point is paired with its corresponding dual-channel BCG signal segment within its time range to form a blood pressure point-level training sample; the system extracts features from multiple short-term sub-windows within this time range, and performs statistical aggregation or weighted aggregation based on the window quality score, thereby forming the blood pressure estimation feature vector corresponding to the reference blood pressure measurement point.

[0033] Furthermore, in step two, short-term sub-windows with a combined dual-channel quality score less than 0.40 are designated as low-quality windows, while short-term sub-windows with a combined dual-channel quality score greater than or equal to 0.70 are designated as high-quality windows. For each blood pressure estimation time period, the quality statistics of multiple short-term sub-windows within that time period are calculated, including the mean, median, minimum quality score, proportion of high-quality windows, proportion of low-quality windows, proportion of effective dual-channel windows, proportion of effective back channel windows, proportion of effective buttock channel windows, and the best single-channel quality score. When the median quality score within the blood pressure estimation time period is greater than or equal to 0.60, and the proportion of low-quality windows is less than or equal to 0.30, the blood pressure estimation time period can be determined to meet the blood pressure estimation quality requirements. When the quality is low or the proportion of effective windows is insufficient, the system can reduce the reliability of the blood pressure estimation results at that time point or prompt for re-collection of data.

[0034] Furthermore, the BCG signals from the back and buttocks described in step one are preprocessed separately, and the specific methods are as follows:

[0035] S11. Perform 1-10Hz bandpass filtering on the acquired back BCG signal and buttock BCG signal respectively;

[0036] S12. Detect motion artifacts based on the moving root mean square amplitude. Calculate the local RMS of the two signals after filtering in step S11, and take the larger value to construct a joint artifact mask.

[0037] S13. Perform interpolation to complete the artifact region;

[0038] S14. The completed signal is denoised using empirical mode decomposition, ensemble empirical mode decomposition, wavelet denoising, adaptive filtering, or a combination thereof.

[0039] S15. Peak detection is performed on the denoised and cleaned signal to obtain the back channel peak sequence and the buttock channel peak sequence, and false peaks in the artifact region are removed.

[0040] Furthermore, the angle between the backrest and the seat can be adjusted to 100±10°, 120±10°, 140±10°, 160±10°, or 180° to accommodate people of different heights and ensure effective sensing of the back and buttocks by the seat sensors. More preferably, the angle between the backrest and the seat can be adjusted to 140°.

[0041] This invention also provides a non-invasive method for estimating heart rate and blood pressure using dual-channel cardiac impulse signals, the specific steps of which are as follows:

[0042] Step 1: After the user sits on the measuring chair, the system adjusts the backrest angle to the preset measuring angle;

[0043] Step 2: Use a back BCG sensor to collect back BCG signals and transmit the collected back BCG signals to the signal acquisition module; use a buttock BCG sensor to collect buttock BCG signals and transmit the collected buttock BCG signals to the signal acquisition module.

[0044] Step 3: The signal acquisition module synchronously samples and receives the back BCG signal acquired by the back BCG sensor and the buttock BCG signal acquired by the buttock BCG sensor, and transmits the received signals to the blood pressure estimation processing module.

[0045] Step 4: The blood pressure estimation processing module receives the back BCG signal and buttock BCG signal transmitted from the signal acquisition module and performs blood pressure estimation.

[0046] Step 5: Use the display and communication module to display the blood pressure estimation output, blood pressure change trend and signal quality status, and be able to transmit the data to an external display.

[0047] The beneficial effects of this invention are:

[0048] 1. Compared with traditional cuff blood pressure monitors, this invention does not require inflation of the user's arm with air each time it is measured. It can achieve weak contact or imperceptible blood pressure trend estimation in the user's natural sitting posture, reducing the discomfort caused by repeated cuff inflation. It is suitable for home health management, rehabilitation training, monitoring in office settings, and observation of blood pressure changes during exercise recovery.

[0049] 2. This invention simultaneously acquires BCG signals from two locations, the back and buttocks, enabling the simultaneous acquisition of cardiac mechanical vibration information from different body contact interfaces and force transmission paths. Compared to single-channel BCG signals, the dual-channel acquisition method provides redundant and complementary information. When the quality of one channel deteriorates due to body movement, changes in sitting posture, or poor local contact, the effective signal from the other channel can still be used for compensation, thereby improving the stability and usability of the blood pressure estimation process.

[0050] 3. By setting an adjustable backrest angle, preferably around 140°, this invention improves the mechanical coupling between the human back and buttocks and the sensor, thereby enhancing the basic availability, heart rhythm consistency, and cross-channel delay stability of the dual-channel BCG signal. This angle setting facilitates increased detectability of cardiac mechanical vibrations in the dual-channel sensor, providing a more reliable signal basis for subsequent peak delay, phase difference, and inter-channel coupling feature extraction.

[0051] 4. This invention introduces a window-level signal quality scoring and a blood pressure segment-level quality gating mechanism, which can evaluate the quality of each short-term BCG sub-window and determine the usability of the current blood pressure estimation segment based on indicators such as the dual-channel joint quality score, the proportion of low-quality windows, and the proportion of effective windows. Through this mechanism, the system can automatically filter low-quality segments or reduce the weight of low-quality windows in feature aggregation and model estimation, thereby reducing the interference of body motion artifacts, false peaks, and poor channel contact on the blood pressure estimation model.

[0052] 5. This invention further proposes a quality-aware adaptive feature extraction strategy. For windows where both the back and buttock channels meet quality requirements, the system extracts dual-channel time delay, phase difference, correlation, inter-channel amplitude ratio, and image fusion features. For windows with only one high-quality channel, the system extracts the heartbeat morphology, amplitude, envelope, frequency domain, and respiratory-related features of that high-quality single channel as compensation. For windows with both channels being low-quality, the system can reduce their weight or remove them. This strategy can fully utilize the complementarity of dual-channel BCG, avoid completely discarding corresponding blood pressure samples due to local low-quality windows, and improve the effective sample utilization rate in small sample and complex acquisition scenarios.

[0053] 6. The blood pressure estimation features extracted in this invention not only include traditional single-channel amplitude, heart rate, and frequency domain energy features, but also dual-channel peak delay, delay stability, cardiac interval consistency, Hilbert phase difference, cross-spectral phase difference, inter-channel correlation, inter-channel transmission features, and dual-channel image fusion features. These features can characterize blood pressure changes from multiple perspectives, including cardiac mechanical activity, mechanical transmission under sitting posture, cross-channel propagation delay, and dual-channel synchronization relationships, providing a richer physiological and mechanical interpretation compared to relying solely on single-channel amplitude or heart rate features.

[0054] 7. This invention employs a real reference blood pressure point-level modeling approach. Instead of linearly interpolating adjacent cuff blood pressure measurements, it directly pairs each real blood pressure measurement point with its corresponding BCG signal features within its time range. This method avoids the additional uncertainties introduced by interpolation labels, allowing the model to directly learn the correspondence between BCG features and real systolic and diastolic blood pressure or their variations, thus improving the reliability of training sample labels.

[0055] 8. This invention combines statistical and quality information from multiple short-term windows during feature aggregation. It can perform mean, median, standard deviation, slope of change, and quality-weighted aggregation of features from multiple 15-second sub-windows corresponding to the same blood pressure measurement point, thereby forming a more stable blood pressure estimation feature vector. This method can reduce the impact of random anomalies in a single short window on blood pressure estimation results and improve the model's ability to represent short-term blood pressure trends.

[0056] 9. This invention allows contextual information such as recovery time, target heart rate percentage, and blood pressure measurement sequence number to be input into a blood pressure estimation model along with BCG features to characterize the dynamic changes in blood pressure during the recovery process after exercise load. This approach is suitable for estimating blood pressure trends during the exercise recovery period and is helpful for rehabilitation training monitoring, exercise risk management, chronic disease follow-up, and personalized health assessment.

[0057] 10. This invention can output estimated systolic blood pressure, estimated diastolic blood pressure, blood pressure change relative to resting baseline, and short-term blood pressure trend curves. It can also provide a reliability score for the estimation results by combining BCG signal quality score, effective window ratio, dual-channel peak matching degree, cross-channel delay stability, and optimal single-channel quality score. When the signal quality is low or the effective window ratio is insufficient, the system can reduce the reliability of the blood pressure estimation results at that time point, or prompt the user to adjust their posture, improve sensor contact, or re-acquire data. Attached Figure Description

[0058] Figure 1 This is a schematic diagram showing representative angles of the measuring chair;

[0059] Figure 2 This is a system block diagram of a dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device.

[0060] Figure 3 This is a flowchart of a non-invasive heart rate and blood pressure estimation method using dual-channel cardiac impulse signals. Detailed Implementation

[0061] The present invention will now be further described with reference to the accompanying drawings.

[0062] like Figure 1-3 As shown, this invention provides a non-invasive heart rate and blood pressure estimation device and method based on dual-channel cardiac impact signals. This approach combines an adjustable-angle seat, simultaneous back-hip dual-channel BCG acquisition, window-level quality assessment, dual-channel phase difference / delay feature extraction, and a blood pressure estimation model to estimate systolic and diastolic blood pressure without a cuff in scenarios such as sitting or during exercise recovery.

[0063] (I) Device Structure

[0064] like Figure 1 ,2 As shown, the measuring chair includes a seat, a backrest, a frame, and an angle adjustment / locking mechanism. The angle adjustment / locking mechanism is used to adjust the angle between the backrest and the seat, with an adjustment range of 100° to 180°, preferably 130° to 150°, and more preferably about 140°; discrete settings such as 100°, 120°, 140°, 160°, and 180° can also be provided.

[0065] The sensing module includes a back BCG sensor and a hip BCG sensor. A back BCG sensor is located at the backrest to collect back BCG signals; a hip BCG sensor is located at the seat or cushion to collect hip BCG signals. Both sensors are connected to the signal acquisition module and are sampled synchronously using a unified clock.

[0066] The signal acquisition module includes a sensor excitation / demodulation unit, an analog front-end, an analog-to-digital conversion unit, and a data transmission unit. The sensor can be a fiber optic Mach-Zehnder interferometer, a piezoelectric thin film, an accelerometer, a pressure sensor, a force-sensitive thin film, or a combination thereof.

[0067] The blood pressure estimation processing module includes a processor, memory, and a model calculation unit for performing blood pressure estimation. The processing module can be located inside the chair or connected to an external computer, mobile phone, tablet, or cloud server.

[0068] The display and communication module is used to display the blood pressure estimation output results, blood pressure change trends, and signal quality status data, and can transmit the data to external terminals, cloud platforms, or health management systems for display.

[0069] (II) Signal Processing Flow

[0070] S1. Acquire BCG signals from the back and buttocks. The sampling rate is preferably 1000Hz, but other sampling rates in the range of 100Hz to 2000Hz are also acceptable.

[0071] S2. Bandpass filtering is performed on the two channels, preferably 1–10Hz, to preserve components related to cardiac mechanical activity and suppress baseline drift and high-frequency noise.

[0072] S3. Based on the sliding root mean square amplitude, detect volumetric artifacts. Calculate the local RMS for each of the two channels and take the larger value to construct a joint artifact mask.

[0073] S4. Complete the artifact region using piecewise cubic Hermite interpolation (PCHIP) or other interpolation methods.

[0074] S5. Denoise the completed signal using Empirical Mode Decomposition (EMD), ensemble empirical mode decomposition, wavelet denoising, adaptive filtering, or a combination thereof.

[0075] S6. Perform peak detection on the cleaned signal to obtain the back channel peak sequence and the buttock channel peak sequence, and remove false peaks in the artifact region.

[0076] (III) Posture Angle Optimization and Quality Gating

[0077] The system adjusts the backrest to a preset measurement angle before or during measurement. Based on the preliminary dual-channel BCG signal quality assessment results, the preferred backrest angle is approximately 130° to 150°.

[0078] The system evaluates the quality of synchronized BCG signals from two channels using a fixed duration window, with an optimal window length of 15 seconds. Single-channel quality scores can be obtained based on characteristics such as morphological consistency, artifact ratio, amplitude stability, number of effective peaks, and heartbeat interval stability. Dual-channel quality scores can further include baseline availability, dual-channel heartbeat interval consistency, and cross-channel delay stability.

[0079] When the dual-channel quality score is lower than the preset threshold, the system can discard the window, reduce the weight of the window in the blood pressure model, or prompt the user to adjust their sitting posture, back support, or backrest angle.

[0080] The specific plan is as follows:

[0081] A manual annotation tool for dual-channel BCG data was developed. To better meet the practical application needs of continuous monitoring and online analysis scenarios, a fixed 15-second window was used as the quality evaluation unit. The annotation interface displays the waveform of the current channel after detrending and normalization processing, the position of candidate peaks, and the suspected artifact interval (mask) obtained by pre-motion detection. Referring to the methods in existing BCG signal quality evaluation research that are based on manual visual inspection, combined with the identifiability of the main peak, waveform morphology characteristics, and amplitude / width constraints for quality judgment, the annotation criteria were expanded and reconstructed.

[0082] Specifically, this application classifies window-level single-channel quality into three levels: Label 2 represents a high-quality window, where the morphology of most cardiac cycles within the window is clear, the main peak is stably identifiable, the peak sequence consistency is good, and artifact coverage is almost zero; Label 1 represents a medium-quality window, where most cardiac beats within the window are still identifiable, but there is a certain proportion of morphological distortion, increased amplitude fluctuations, or local artifact interference; Label 0 represents a low-quality window, where the morphology of cardiac beats within the window is difficult to stably identify, or the signal is clearly dominated by artifacts, making it unsuitable as a reliable input for subsequent physiological parameter extraction. This classification method retains the operability of manual observation while providing a clear and hierarchical labeling basis for subsequent supervised learning modeling.

[0083] (1) Single-channel feature extraction and classification

[0084] To achieve interpretable modeling of window-level BCG signal quality, this application constructs 30-dimensional candidate quality features within each 15-second window, covering multiple aspects such as morphological statistics, rhythm and periodicity statistics, artifact proportion and continuity statistics, and amplitude stability statistics. To reduce redundancy and improve model interpretability, this application sequentially performs data quality filtering, Kruskal-Wallis univariate screening, and Spearman correlation-based redundancy removal. Considering that the above steps are still insufficient to determine the optimal feature subset, this application further employs Sequential Forward Selection (SFS), using the validation set macro-F1 as the objective function, and combines a Bagged Trees classifier to evaluate the performance of different feature combinations. The results show that the classification performance reaches its highest point when the sixth feature is added; further additions of features do not bring gains but instead slightly decrease performance. Therefore, this application ultimately selects 6-dimensional features as the single-channel supervised classification input, covering four quality dimensions: morphological consistency, amplitude stability, rhythm stability, and artifact proportion.

[0085] In the classification modeling stage, this application establishes independent supervised classification models for the back and hip channels. This design is based on the fact that the two channels exhibit systematic differences in human coupling path, contact pressure, signal amplitude, and waveform performance. Existing research indicates that the BCG wave groups emphasized at the seat cushion and backrest measurement locations are not identical. Therefore, directly mixing the samples from the two channels for modeling may force the classifier to learn a compromise boundary, thereby weakening the stability and interpretability of single-channel quality judgment. The classifier employs a Bagging Tree ensemble model. This method generates multiple training subsets through bootstrapping, trains decision trees on each subset, and finally provides the classification result through ensemble analysis, thereby reducing model variance and improving robustness to noise and nonlinear boundaries.

[0086] Considering the imbalance in the label distribution of both the back and hip channels, this application introduces a class weighting strategy based on the inverse of class frequency during single-channel model training to reduce the dominant influence of majority class samples on the training process. Training and evaluation employ a window-level hierarchical random splitting protocol, dividing all samples into training, validation, and test sets in an 80 / 10 / 10 ratio, and repeating this random splitting five times. Finally, model performance is reported in the form of mean ± standard deviation (SD) for the test set metrics.

[0087] Considering the imbalance in label distribution between the back and hip channels, this application employs a class weighting strategy based on the inverse of class frequency during single-channel classification model training to reduce the dominant influence of majority class samples on the training process. For a sample of class c in the training set, let its sample size be... The total number of samples is The number of categories is The weight of this type of sample is defined as:

[0088]

[0089] Subsequently, to avoid additional impacts on the training process caused by changes in the overall weight scale, this application further normalizes the sample weights to a mean of 1. This strategy can improve the effective contribution of minority class samples during training without changing the original sample distribution, thereby improving the model's balance in recognizing different quality levels.

[0090] (2) Continuous quality scoring and dual-channel fusion

[0091] After completing single-channel classification validation, this application further utilizes the class probabilities output by the classifier to construct a window-level continuous quality score. Existing research has used the Continuous Quality Index (SQI) or multi-index fusion to quantitatively evaluate BCG quality; based on this, and considering the natural ordering relationship between L0, L1, and L2, this application considers high-quality categories as the main effective contribution and medium-quality categories as partially effective contributions, thus mapping discrete labels to continuous quality scores. For the back and hip channels, their single-channel continuous quality scores are defined as follows:

[0092]

[0093]

[0094] Here, P(L2) represents the probability that the window belongs to the high-quality category, and P(L1) represents the probability that the window belongs to the medium-quality category. This definition uses the high-quality category as the main contributor and the medium-quality category as a partial contributor to describe the continuous transition state of window quality between "poor-medium-good" and provides a basis for subsequent dual-channel joint scoring.

[0095] At the dual-channel level, relying solely on the quality of a single channel is insufficient to determine whether a given window is suitable for dual-channel time-series analysis. Existing research has shown that multi-channel BCG can improve heart rate estimation performance through cross-channel information fusion, and chair-based two-channel BCG has also been used for non-sensory blood pressure correlation analysis. Furthermore, BCG signal quality and detection performance vary across different body positions; therefore, constructing a joint scoring framework to comprehensively evaluate dual-channel analyzability is reasonable. This framework consists of three primary indicators: single-channel baseline quality... Dual-channel JJ consistency item and cross-channel delay stability term .

[0096] (3) Basic quality items for single channel

[0097] The single-channel basic quality term is used to measure whether both channels simultaneously possess basic usability. Considering that low quality in either channel during dual-channel analysis could limit the reliability of subsequent time-series analyses, this application employs geometric mean fusion of the continuous quality scores from both channels:

[0098]

[0099] Compared to simple averaging, geometric mean has lower compensatory power and can better reflect the "weakest link effect" in dual-channel analysis, that is, when the quality of any channel is significantly lower, the combined basic quality will be significantly suppressed.

[0100] (4) Dual-channel JJ consistency item

[0101] The dual-channel JJ consistency term is used to measure the ability of the back and hip channels to jointly represent the same cardiac event sequence at the rhythm level. Previous studies have shown that multi-channel or multi-site BCG provides complementary information, and signal performance varies across different measurement locations; furthermore, the JJ interval extracted from BCG and the RR interval from ECG show good consistency in rhythm and HRV analysis. Therefore, this application constructs this index from three aspects: event matching, rhythm trend consistency, and rhythm difference stability, using beat-by-beat matching success rate as the metric. Correlation coefficient of two-channel JJ interval sequences and the standard deviation of the JJ difference sequence Measurement is performed. Among them, and As a positive indicator, This is a reverse indicator. To achieve unified integration, this application first normalizes each indicator and then uses a monotonic mapping function. The inverse index is mapped to a scoring format where "the larger the value, the better." Based on this, and drawing on the "mapping first, then fusion" approach from fuzzy comprehensive evaluation, and considering the lack of explicit prior weights, this application adopts an equal-weight fusion definition:

[0102]

[0103] in, This is used to avoid negative correlations making unreasonable contributions to the consistency score.

[0104] (5) Cross-channel delay stability term

[0105] The cross-channel time delay stability term is used to evaluate whether the relative arrival time of the same cardiac event is stable in the back and hip channels. This characteristic is a key prerequisite for subsequent cross-channel time series analysis and PTT-type index estimation. Previous studies have shown that the time delay characteristics based on BCG can be used for pulse propagation time correlation analysis and cuffless blood pressure estimation, and dual-channel chair-based BCG has also been used for continuous blood pressure estimation. Therefore, evaluating the analyzability of dual channels from the perspective of the stability of the relative arrival time across channels is reasonable. To characterize the stability of the time delay sequence, this application models it from two aspects: fluctuation amplitude and abnormal events, introducing the standard deviation of the beat-by-beat cross-channel time delay. and the proportion of time delay abnormal shooting As metrics, both of the above metrics are inverse metrics where "smaller values ​​are better," therefore, this application also first normalizes them and then applies a monotonic mapping function. Convert it to a scoring format where "the higher the value, the better". It should be noted that... and The specific definition pertains to the time delay stability characterization proposed in this paper for dual-channel tasks. Its design is based on the fact that artifacts, anomalous fluctuations, and systematic shifts in the time delay sequence significantly reduce the reliability of time delay indicators. Building upon this, this application continues to draw upon ideas from fuzzy comprehensive evaluation and comprehensive indicator construction, defining the cross-channel time delay stability term with equal weight fusion as:

[0106]

[0107] (6) Dual-channel joint quality analysis

[0108] The dual-channel joint quality score is used to comprehensively evaluate the overall suitability of the current window for dual-channel time series analysis tasks. Given... , and These correspond to three complementary quality dimensions with clear physiological significance: basic availability, rhythm consistency, and cross-channel propagation stability. However, there is currently a lack of sufficient prior or independently labeled evidence to support the assumption that any one dimension should be given higher importance. This application continues to draw on the concept of fuzzy comprehensive evaluation, regarding these three as complementary evaluation factors that jointly determine the joint quality of the two channels, and defines an equal-weighted fusion method. Dual-channel joint quality score:

[0109]

[0110] Based on the above definition, the dual-channel joint quality score no longer depends on the subjective priority of a certain module, but instead treats three quality dimensions with clear physiological significance as equally important sources of information to construct a comprehensive quality representation for dual-channel time series analysis tasks.

[0111] (iv) Characteristics of dual-channel blood pressure estimation

[0112] The system estimates blood pressure based on dual-channel BCG signals simultaneously acquired via the back and buttock channels. Each reference blood pressure measurement point corresponds to a segment of dual-channel BCG signal within a preset time range or its vicinity. The system further divides this blood pressure estimation time period into multiple short-term sub-windows. Preferably, each short-term sub-window is 15 seconds long; each blood pressure estimation time period may contain 2, 3, 10, or other numbers of short-term sub-windows. If the actual available acquisition time is less than the preset time, the actual effective duration is used; if the actual acquisition time exceeds the preset analysis length, data within the preset length is truncated for subsequent analysis. In the above embodiment, the short-term sub-window is preferably 15 seconds, but is not limited to 15 seconds. Depending on the sampling rate, heart rate range, and application scenario, the short-term sub-window length can be set to 5 seconds, 10 seconds, 20 seconds, 30 seconds, or other lengths. The blood pressure estimation time period is not less than 45 seconds, preferably 60 to 150 seconds. It can be adjusted according to the reference blood pressure measurement interval, exercise recovery speed, effective signal length, and real-time requirements.

[0113] To improve the reliability of blood pressure estimation features, the system first performs signal quality assessment for each short-term sub-window. Specifically, the system calculates the quality score for the back channel and the hip channel separately, and further combines the baseline quality of the two channels, the consistency of the heart rhythm between the two channels, and the stability of the cross-channel delay to obtain the joint quality score of the two channels. The joint quality score of the two channels can be obtained by weighting the baseline quality term, the rhythm consistency term, and the delay stability term. Preferably, the baseline quality term is represented by the geometric mean of the back channel quality score and the hip channel quality score; the rhythm consistency term is represented by the matching ratio of the heart peaks in the two channels, the correlation of the heart interval sequence, and the stability of the heart interval difference; the delay stability term is represented by the standard deviation of the cross-channel peak delay and the proportion of delay anomalies.

[0114] In a preferred embodiment, the system designates short-term sub-windows with a combined dual-channel quality score less than 0.40 as low-quality windows and short-term sub-windows with a combined dual-channel quality score greater than or equal to 0.70 as high-quality windows. For each blood pressure estimation time period, the system further calculates the quality statistics of multiple short-term sub-windows within that time period, including the mean, median, minimum quality score, proportion of high-quality windows, proportion of low-quality windows, proportion of effective dual-channel windows, proportion of effective back channel windows, proportion of effective buttock channel windows, and the optimal single-channel quality score. Preferably, when the median quality score within the blood pressure estimation time period is greater than or equal to 0.60 and the proportion of low-quality windows is less than or equal to 0.30, the blood pressure estimation time period can be determined to meet the blood pressure estimation quality requirements; when the quality is low or the proportion of effective windows is insufficient, the system can reduce the reliability of the blood pressure estimation result at that time point or prompt for re-collection of data.

[0115] During feature extraction, the system employs a quality-aware adaptive feature extraction strategy. For short-term sub-windows that simultaneously meet the quality requirements of the back channel, the buttock channel, and the combined quality of both channels, the system extracts dual-channel BCG features. For short-term sub-windows where one channel has lower quality but the other is still usable, the system retains the higher-quality single-channel BCG signal and extracts single-channel auxiliary features. For short-term sub-windows where both channels are of low quality, the system can either remove the window or assign it a lower weight. Thus, the system can utilize the dual-channel delay and coupling information when both channels are reliable, and compensate for the impact of body movement interference or poor contact on one channel by using the high-quality signal from the other channel, thereby improving the robustness of the blood pressure estimation process to local signal quality fluctuations.

[0116] The dual-channel BCG features extracted by the system include, but are not limited to, dual-channel event delay features, dual-channel phase difference features, dual-channel correlation features, dual-channel amplitude ratio features, dual-channel energy ratio features, and dual-channel image fusion features. The dual-channel event delay features include the beat-by-beat matching ratio of the back channel peak and the buttock channel peak, the mean cross-channel peak delay, the median delay, the standard deviation of the delay, the interquartile range of the delay, the proportion of delay anomalies, the slope of the delay change, and the consistency of the two-channel heartbeat interval sequence. The dual-channel phase difference features include the Hilbert instantaneous phase difference between the back channel and the buttock channel, the circular mean of the phase difference, the circular standard deviation of the phase difference, phase synchronization stability, the cross-spectral dominant frequency phase difference, the cross-spectral weighted phase difference, and the cross-channel delay converted from the phase difference. The dual-channel correlation and transmission features include the cross-correlation peak value, the cross-correlation hysteresis, the heartbeat frequency band coherence, the inter-channel transmission gain, the inter-channel phase difference, and the inter-channel energy distribution difference.

[0117] The system can also extract single-channel BCG auxiliary features. These single-channel features include heart rate, heartbeat interval, heartbeat interval variability, peak amplitude, amplitude variability, root mean square (RMS) signal, frequency domain energy, heartbeat frequency band energy, respiratory frequency band energy, envelope features, diastolic decay features, and cycle template morphology features. The diastolic decay features include the descent slope after the J-peak, the area after the J-peak, the amplitude of subsequent troughs, the peak-to-trough difference, half-peak width, the energy percentage of the second half of the cycle, and the time interval from the main peak to subsequent feature points. The cycle template morphology features include the average template amplitude, template area, template width, template skewness, template kurtosis, template stability, and correlation or distance with the reference template after heartbeat peak alignment. For single-channel compensation, the system can select the channel with higher quality within the current time period as the optimal single channel based on the quality scores of the back and hip channels, and extract the corresponding optimal single-channel features.

[0118] Furthermore, the system can convert the BCG signal into an image-based feature representation to extract nonlinear morphology and channel coupling information. These image-based features include single-channel Gram angle field features, dual-channel cross-recursive graph features, and dual-channel fusion graph features. The dual-channel fusion graph can be composed of back channel amplitude, hip channel amplitude, difference between the two channels, product of the two channels, and absolute difference between the two channels, and further calculates image entropy, texture statistics, row and column direction variation features, diagonal features, and frequency domain texture features. These image features, together with event delay features, phase difference features, single-channel morphological features, and signal quality features, constitute a blood pressure estimation feature vector.

[0119] In applications during the exercise recovery period, the system can also incorporate contextual features to assist the model in characterizing the dynamic changes in blood pressure during the post-exercise recovery process. These contextual features include, but are not limited to, the percentage of target exercise load, recovery time, blood pressure measurement sequence number, measurement window sequence number, heart rate change trend, and time position relative to the resting state. The system can input dual-channel BCG features, single-channel compensation features, image features, signal quality features, and contextual features into the blood pressure estimation model.

[0120] During the model training phase, the system does not perform linear interpolation on adjacent cuff blood pressure measurements. Instead, it uses each real reference blood pressure measurement point as a supervision label. Specifically, each reference blood pressure measurement point is paired with a dual-channel BCG signal segment within its corresponding time range to form a blood pressure point-level training sample. The system extracts features from multiple short-term sub-windows within this time range and performs statistical aggregation or weighted aggregation based on window quality scores. Aggregation methods include mean, median, standard deviation, maximum, minimum, interquartile range, effective window proportion, slope of change, and quality-weighted average, thereby forming the blood pressure estimation feature vector corresponding to the reference blood pressure measurement point.

[0121] The blood pressure estimation model can establish separate models for systolic and diastolic blood pressure, or it can establish models for the changes in systolic and diastolic blood pressure relative to an individual's resting baseline. Preferably, the system uses a partial least squares regression (PLR) model or a ridge regression model to estimate blood pressure. PLR can extract latent variables related to blood pressure labels under high-dimensional features and small sample conditions, and is suitable for multi-type BCG feature fusion modeling; ridge regression, by adding a regularization term to the least squares loss function, can reduce the risk of overfitting under small sample and high-dimensional feature conditions. During model training, the input consists of the BCG aggregate features, quality features, and optional context features corresponding to the actual blood pressure measurement points, and the output is the measured systolic blood pressure, diastolic blood pressure, or their relative changes at that time point.

[0122] In a preferred embodiment, the system employs a signal quality-based gating strategy for blood pressure estimation samples. For blood pressure estimation time periods that meet the quality gating conditions, the system utilizes its dual-channel BCG features, single-channel compensation features, and contextual features for blood pressure estimation. For time periods that do not meet the quality gating conditions but still have a single-channel effective window, the system can adopt a quality adaptive strategy, utilizing high-quality single-channel features for compensation estimation. For time periods where neither dual-channel nor single-channel features are available, the system can reduce the reliability of the estimation results at that time point or prompt for data re-collection.

[0123] In the exercise recovery phase example, after the subject reaches a preset target heart rate percentage, they enter a recovery state. The system simultaneously acquires dual-channel BCG signals and cuff reference blood pressure. Each cuff blood pressure measurement point is paired with its corresponding time range's BCG features to form real blood pressure point-level training samples. Recovery time, target heart rate percentage, and measurement sequence number can be used as auxiliary context variables input into the model to enhance the model's ability to characterize the post-exercise blood pressure recovery trend.

[0124] The model output includes point-level systolic and diastolic blood pressure estimates, and may also include changes in systolic and diastolic blood pressure, as well as short-term blood pressure trend curves. The system can repeatedly execute the above signal quality assessment, feature extraction, feature aggregation, and model estimation processes at preset time intervals to generate continuous or quasi-continuous short-term blood pressure change trends. Furthermore, the system can provide a reliability score for the blood pressure estimation results based on BCG signal quality score, effective window ratio, dual-channel peak matching degree, inter-channel delay stability, optimal single-channel quality, and model prediction error distribution.

[0125] In the small-sample validation of this embodiment, the use of real blood pressure point-level modeling avoids the additional uncertainties introduced by interpolation labels, allowing the model to directly learn the correspondence between BCG features and measured blood pressure. By combining signal quality gating, dual-channel BCG features, single-channel compensation features, and contextual features, the system can reduce the interference of low-quality BCG fragments on blood pressure estimation and improve the stability of blood pressure estimation during exercise recovery, demonstrating the preliminary feasibility of this method for estimating short-term blood pressure changes.

[0126] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device, characterized in that: It includes a measuring chair, a back BCG sensor, a hip BCG sensor, a signal acquisition module, a blood pressure estimation and processing module, and a display and communication module; The measuring chair includes a seat and a backrest, and the angle between the backrest and the seat can be adjusted and fixed arbitrarily. The back BCG sensor is disposed in the backrest and is used to collect back BCG signals; the back BCG sensor is connected to the signal acquisition module and is used to transmit the collected back BCG signals to the signal acquisition module. The hip BCG sensor is disposed in the seat or cushion of the chair and is used to collect hip BCG signals; the hip BCG sensor is connected to the signal acquisition module and is used to transmit the collected hip BCG signals to the signal acquisition module. The signal acquisition module synchronously samples and receives the back BCG signal acquired by the back BCG sensor and the buttock BCG signal acquired by the buttock BCG sensor, and transmits the received signals to the blood pressure estimation processing module. The blood pressure estimation processing module receives the back BCG signal and buttock BCG signal transmitted from the signal acquisition module and performs blood pressure estimation. The display and communication module is used to display the blood pressure estimation output results, blood pressure change trends, and signal quality status data, and can transmit the data to an external display.

2. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 1, characterized in that, The blood pressure estimation processing module receives the back BCG signal and buttock BCG signal transmitted from the signal acquisition module and performs blood pressure estimation. The specific steps are as follows: Step 1: Preprocess the BCG signals from the back and buttocks respectively; Step 2: For each reference blood pressure measurement point, there are back BCG signal segments and buttock BCG signal segments within a preset time range or nearby. The back BCG signal segments and buttock BCG signal segments are segmented into multiple short-term sub-windows according to the preset short window length. The signal quality of each short-term sub-window is evaluated, and the back channel quality score, buttock channel quality score and dual-channel combined quality score of each short-term sub-window are calculated respectively. The dual-channel joint quality score is obtained by weighting a single-channel baseline quality term, a dual-channel rhythm consistency term, and a cross-channel delay stability term; the single-channel baseline quality term is represented by the geometric mean of the back channel quality score and the buttock channel quality score; the dual-channel rhythm consistency term is represented by the heartbeat peak matching ratio of the two channels, the correlation of the heartbeat interval sequence, and the stability of the heartbeat interval difference; the cross-channel delay stability term is represented by the standard deviation of the cross-channel peak delay and the proportion of delay anomalies. Step 3: Employ a quality-aware adaptive feature extraction strategy. For short-term sub-windows that simultaneously meet the requirements of back channel quality score, buttock channel quality score, and combined dual-channel quality score, extract dual-channel BCG features of the back BCG signal and buttock BCG signal. For short-term sub-windows where one channel has a low quality score but the other channel still has a usable quality score, retain the single-channel BCG signal with better quality and extract the corresponding single-channel BCG auxiliary features. For short-term sub-windows where both the back channel quality score and the buttock channel quality score are low, remove the short-term sub-window or assign it a lower weight. Step 4: Input the features extracted in Step 3 into the blood pressure estimation model. The blood pressure estimation model includes a systolic blood pressure estimation model, a diastolic blood pressure estimation model, a systolic blood pressure change estimation model relative to the individual's resting baseline, and a diastolic blood pressure change estimation model relative to the individual's resting baseline, thereby outputting the estimated systolic blood pressure value, estimated diastolic blood pressure value, systolic blood pressure change, diastolic blood pressure change, and short-term blood pressure trend curve for the corresponding reference blood pressure measurement point.

3. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 2, characterized in that: The dual-channel BCG features mentioned in step three include, but are not limited to, dual-channel event delay features, dual-channel phase difference features, dual-channel correlation and transmission features, dual-channel amplitude ratio features, dual-channel energy ratio features, and dual-channel image fusion features. The dual-channel event delay features include the beat-by-beat matching ratio of the back channel peak and the buttock channel peak, the mean cross-channel peak delay, the median delay, the standard deviation of the delay, the interquartile range of the delay, the proportion of delay anomalies, the slope of the delay change, and the consistency of the two-channel heartbeat interval sequence. The dual-channel phase difference features include the Hilbert instantaneous phase difference between the back channel and the buttock channel, the circular mean of the phase difference, the circular standard deviation of the phase difference, phase synchronization stability, the cross-spectral dominant frequency phase difference, the cross-spectral weighted phase difference, and the cross-channel delay converted from the phase difference. The dual-channel correlation and transmission features include the cross-correlation peak value, the cross-correlation lag, the heartbeat frequency band coherence, the inter-channel transmission gain, the inter-channel phase difference, and the inter-channel energy distribution difference. The single-channel BCG auxiliary features mentioned in step three include heart rate, heartbeat interval, heartbeat interval variability, peak amplitude, amplitude variability, root mean square of signal, frequency domain energy, heartbeat frequency band energy, respiratory frequency band energy, envelope features, diastolic decay features, and cycle template morphology features. The diastolic decay features include the descent slope after the J peak, the area after the J peak, the amplitude of the subsequent trough, the peak-to-trough difference, the half-peak width, the energy proportion of the second half of the cycle, and the time interval from the main peak to the subsequent feature point. The cycle template morphology features include the average template amplitude, template area, template width, template skewness, template kurtosis, template stability, and the correlation or distance with the reference template after heartbeat peak alignment.

4. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 2, characterized in that: The features extracted in step three also include image features derived from the BCG signal to extract nonlinear morphology and channel coupling information; the image features include single-channel Gram angle field features, dual-channel cross-recursive graph features, and dual-channel fusion graph features; the dual-channel fusion graph includes back channel amplitude, hip channel amplitude, difference between the two channels, product of the two channels, and absolute difference between the two channels, and further calculates image entropy, texture statistics, row and column direction change features, diagonal features, and frequency domain texture features.

5. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 2, characterized in that: The features extracted in step three also include contextual features introduced in the application scenario of exercise recovery, to help the model characterize the dynamic changes in blood pressure during the post-exercise recovery process; the contextual features include, but are not limited to, the target exercise load percentage, recovery time, blood pressure measurement sequence number, measurement window sequence number, heart rate change trend, and time position relative to the resting state.

6. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 2, characterized in that: In the blood pressure estimation model described in step four, during the model training phase, the system does not perform linear interpolation on adjacent cuff blood pressure measurements. Instead, it uses each real reference blood pressure measurement point as a supervision label. Specifically, each reference blood pressure measurement point is paired with a dual-channel BCG signal segment within its corresponding time range to form a blood pressure point-level training sample. The system extracts features from multiple short-term sub-windows within this time range and performs statistical aggregation or weighted aggregation based on the window quality score to form the blood pressure estimation feature vector corresponding to the reference blood pressure measurement point.

7. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 2, characterized in that: In step two, short-term sub-windows with a combined dual-channel quality score less than 0.40 are designated as low-quality windows, and short-term sub-windows with a combined dual-channel quality score greater than or equal to 0.70 are designated as high-quality windows. For each blood pressure estimation time period, the quality statistics of multiple short-term sub-windows within that time period are calculated, including the mean, median, minimum, proportion of high-quality windows, proportion of low-quality windows, proportion of effective dual-channel windows, proportion of effective back channel windows, proportion of effective buttock channel windows, and the best single-channel quality score. When the median quality score within the blood pressure estimation period is greater than or equal to 0.60 and the proportion of low-quality windows is less than or equal to 0.30, the blood pressure estimation period can be determined to meet the blood pressure estimation quality requirements. When the quality is low or the proportion of effective windows is insufficient, the system can reduce the reliability of the blood pressure estimation results at that time point or prompt for data to be collected again.

8. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 2, characterized in that, Step one involves preprocessing the BCG signals from the back and buttocks, respectively. The specific methods are as follows: S11. Perform 1-10Hz bandpass filtering on the acquired back BCG signal and buttock BCG signal respectively; S12. Detect motion artifacts based on the moving root mean square amplitude. Calculate the local RMS of the two signals after filtering in step S11, and take the larger value to construct a joint artifact mask. S13. Perform interpolation to complete the artifact region; S14. The completed signal is denoised using empirical mode decomposition, ensemble empirical mode decomposition, wavelet denoising, adaptive filtering, or a combination thereof. S15. Peak detection is performed on the denoised and cleaned signal to obtain the back channel peak sequence and the buttock channel peak sequence, and false peaks in the artifact region are removed.

9. The dual-channel cardiac impact signal non-invasive heart rate and blood pressure estimation device according to claim 1, characterized in that: The angle between the backrest and the seat can be adjusted to 100±10°, 120±10°, 140±10°, 160±10° or 180°.

10. The method for non-invasive heart rate and blood pressure estimation using a dual-channel cardiac impulse signal based on the dual-channel cardiac impulse signal non-invasive heart rate and blood pressure estimation device according to claims 1-9, characterized in that, The specific steps are as follows: Step 1: After the user sits on the measuring chair, the system adjusts the backrest angle to the preset measuring angle; Step 2: Use a back BCG sensor to collect back BCG signals and transmit the collected back BCG signals to the signal acquisition module; use a buttock BCG sensor to collect buttock BCG signals and transmit the collected buttock BCG signals to the signal acquisition module. Step 3: The signal acquisition module synchronously samples and receives the back BCG signal acquired by the back BCG sensor and the buttock BCG signal acquired by the buttock BCG sensor, and transmits the received signals to the blood pressure estimation processing module. Step 4: The blood pressure estimation processing module receives the back BCG signal and buttock BCG signal transmitted from the signal acquisition module and performs blood pressure estimation. Step 5: Use the display and communication module to display the blood pressure estimation output, blood pressure change trend, and signal quality status data, and be able to transmit the data to an external display.

Images