A blood pressure measuring device and method with physiological morphology reconstruction function

Abstract

This invention relates to a blood pressure measurement device and method with physiological morphology reconstruction function, belonging to the fields of medical electronics technology and artificial intelligence biometric prediction. Addressing the problem that high-order features are easily damaged in existing blood pressure measurement technologies under strong physical interference environments, this method first simultaneously extracts and phase-aligns the original feature stream, first-order velocity feature stream, second-order excitation-state feature stream, and macroscopic morphological envelope; then, it independently performs frequency-band quantization quality assessment on each channel through a signal quality dynamic evaluation engine; when high-order feature channels are damaged, it uses low-order features as input priors and generates a virtual second-order compensated feature stream in real time through a pre-trained physiological dynamic reprojector; the routing is then re-aligned and reconstructed into a four-dimensional input feature tensor; blood pressure regression prediction is performed through a deep learning inference model, and output reliability is ensured through residual closed-loop verification. This invention can ensure signal integrity and full-frame coherence under severe body motion interference and is suitable for upper arm and wrist wearable devices.

Description

Technical Field

[0001] This invention relates to the fields of medical electronic technology and artificial intelligence biometric prediction. Specifically, it relates to a blood pressure measurement device and method with physiological morphology reconstruction function. It designs a non-destructive extraction methodology for cardiovascular pressure under strong physical interference conditions, which combines dynamic heterogeneous computation path selection and multi-flow signal cross-boundary reconstruction compensation functions. Background Technology

[0002] Blood pressure (BP, comprising systolic and diastolic pressure) depicts the pumping and pressure supply patterns of the life engine and the healthy damping of the entire vascular network. It is an undisputed and crucial threshold indicator in the system of cardiac circulatory physiology, and has long been regarded as the first sentinel for judging and preventing metabolic deterioration, the onset of chronic diseases, and damage to the cardiovascular system. Today, the gold standard for automatic blood pressure monitoring technology—the oscillometric method based on oscillation extraction—is widely available in millions of ordinary households due to its safety, non-destructive nature, lack of pollution, and affordability. The core operating principle of the traditional oscillometric method is as follows: using a flexible plastic inflatable cuff surrounding the limb (brachial or radial artery root segment) as a medium, a static pressure higher than the arterial hypertension is artificially introduced. Subsequently, the pressure is uniformly reduced by a valve at a specific slope, and the pulsating oscillation wave caused by the obstruction and release of blood flow is reproduced. Subsequently, the envelope was plotted using statistical laws, and the pressure reading was derived by estimating the empirical ratio algorithm (Maximum Amplitude Algorithm, MAA, etc.) using the characteristic amplitude coefficient.

[0003] While the ease of use of traditional empirical ratio algorithms has been widely validated, their theoretical foundation is based solely on the average central tendency of massive human statistical data. This leads to a significant decrease in estimation accuracy and potential for misjudgment when dealing with small, specialized populations—such as elderly individuals with severe aortic sclerosis, peripheral flow disturbances, or atrial fibrillation—and other marginally extreme subjects. With the rapid development of artificial intelligence (AI) algorithms, numerous domestic and international invention patents have pioneered end-to-end long short-term memory models, support vector machine structures, and deep multi-scale convolutional neural networks (MS-CNN) in an attempt to overcome the performance bottlenecks of traditional statistical models. Research indicates that by exploring weak, scattered local morphological features at waveform inflection points (such as the morphological features of the descending dicrotic notch trough and the subsequent rising dicrotic tidal wave peak), multi-scale deep network systems can significantly improve the accuracy of diastolic blood pressure prediction.

[0004] Despite performing well in ideal testing environments, in the complex daily health monitoring scenarios such as at home, walking, or working, even slight bodily movements of the user, scraping of the airway, irregular vibrations of the air pump motor, or muscle deformation caused by deep breathing can generate broadband artifact noise with significant frequency crossover with the original wave being detected. To deeply explore vascular physiological characteristics, cutting-edge neural network models generally extract "raw pressure pulse wave displacement data" and use numerical differentiation to transform it into first-order features (feature velocity) covering the slope of change and high-frequency sensitive second-order features (feature acceleration) as composite prediction inputs to expand the model's perceptual dimensions; however, according to signal processing principles, any numerical difference calculation inherently has the characteristic of amplifying high-frequency noise. Even the slightest, invisible background noise mixed into the pre-amplifier channel will be amplified dozens of times at the moment of second-order derivative calculation, thus dominating the original signal. This will instantly completely cover up the vascular conformal elastic inflection point shape that carries effective medical information and cause severe feature loss, resulting in severe distortion of the waveform profile. This phenomenon is usually manifested as a high-order "feature distortion" in academic and review standards.

[0005] Faced with these severe motion artifacts and high-frequency noise interference, current mainstream competitive technology proposals still adhere to narrow and traditional defensive strategies:

[0006] First, the widespread deployment of filtering algorithms such as low-pass filtering or wavelet denoising attempts to filter out high-frequency components. This results in the removal of the extremely high-frequency inflection point information of the blood vessel wall elasticity that was intended to be extracted, along with the high-order harmonics. Consequently, the effective high-frequency physiological features are lost synchronously, ultimately leading to a decrease in the feature extraction accuracy of deep networks.

[0007] Secondly, a simple signal gating mechanism is used to eliminate contamination windows. A common practice is to use amplitude as the sole criterion, directly marking periods with abnormal fluctuations as invalid data and discarding them. Over time, this can easily lead to excessive loss of continuous sampling segments, wasting valuable temporal baseline information that still has some analytical significance. Furthermore, it forces subsequent deep attention networks to suffer severe defocusing bias, resulting in a significantly increased statistical deviation (SD) parameter that exceeds the error tolerance specified by the standard when evaluating clinical validation testing standards (such as ISO 81060-2 and the international BHS protocol certification).

[0008] Based on the above analysis of conventional noise reduction methods, it can be seen that the key technical problem that needs to be solved in this specific application category is not simply relying on lossy feature processing methods such as discarding or filtering and suppressing. Instead, it is necessary to introduce innovative methods to reconstruct effective high-frequency feature envelopes in real time when the system determines that the physical signal feature flow is damaged or severely missing. In this way, the global noise reduction and security operation model can be reorganized by using a cross-order generation and correction architecture that takes into account both dynamic adaptability and high stability. Summary of the Invention

[0009] In view of this, the present invention provides a blood pressure measuring device and method with physiological morphology reconstruction function to solve the problem of overall network derivation collapse caused by contamination paralysis of high-frequency subdivision feature channels in the prior art.

[0010] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0011] One aspect of this invention is to provide a blood pressure measurement method to fill the technological gap in existing intelligent automatic blood pressure monitoring technologies, which are highly susceptible to severe interference and masking damage from irreversible mechanical and physiological random background noise in environments with variable motion disturbances and scenarios where high-order derivative physical characteristics are easily affected. The aim is to develop a blood pressure measurement computing hardware and software platform that encompasses multimodal concurrent detection, real-time high-precision domain-specific evaluation, and adaptive channel degradation and physiological feature reconstruction capabilities.

[0012] The blood pressure measurement method with physiological morphology reconstruction function described in this invention includes the following steps:

[0013] S1 Multi-order Feature Synchronous Extraction: Acquires the raw cuff pressure pulse signal and simultaneously extracts the four-dimensional temporal feature stream, including: the raw pressure pulse wave stream. First-order gradient feature flow Second-order gradient feature flow And the physiological macromorphic envelope flow obtained through the generalized Hilbert transform. A total of four phase-locked physical feature streams are constructed. Among them, phase alignment compensation is performed on the four-dimensional temporal feature streams through a digital buffer register with a fixed-length delay to achieve zero-timescale misalignment alignment and construct a high-fidelity spatiotemporal polymorphic feature tensor with dimension synchronization and phase consistency.

[0014] S2 Signal Quality Assessment: The phase-aligned four-dimensional time-series feature stream is input into the signal quality dynamic assessment and reconstruction engine. The engine independently calculates the statistical morphological consistency characteristics, local energy proportion characteristics in the time-frequency domain, and nonlinear dynamic signal entropy characteristics of each feature stream within the current window. The resulting fusion output is a high-dimensional local quality assessment vector used to characterize the signal-to-noise ratio safety margin in real time. ;in Indicates the original pressure pulse wave flow Signal quality metrics; Represents the first-order gradient feature flow Signal quality metrics; Represents the second-order gradient feature flow Signal quality metrics; Represents the physiological macromorphic envelope flow The signal quality index; each of the signal quality indices is used to characterize the reliability of the corresponding feature stream within the current time window, and its value is preferably a normalized scalar with a value range of 0 to 1, where the closer the value is to 1, the higher the signal quality, and the closer it is to 0, the higher the degree of interference to the signal.

[0015] S3 Quality Threshold Determination: Judgment Is it below the preset time series robustness safety lower limit? ,and Above the safety limit ;

[0016] If the result is "no", proceed to step S5 and execute the conventional parallel feature extraction route.

[0017] If the determination is "yes": proceed to step S4 to trigger the dynamic physiological remodeling mechanism;

[0018] S4 Physiological Feature Reconstruction: Isolating Contaminated Second-Order Gradient Feature Flow The first-order gradient feature flow that meets the quality standard and the original pulse flow As a priori reference benchmark, virtual second-order compensated feature flows are inverted, predicted, and synthesized in real time by loading a predefined nonlinear equivalent arterial transfer function operator. To replace the second-order gradient feature flow that is masked by environmental noise. flow;

[0019] S5. Route realignment step: Perform route selection and resynchronization for the characteristic channel; if determined as "No" by S3: select the original second-order excited-state characteristic flow. As an effective high-order feature channel; if determined as "yes" by S3, reconstructed by S4: select virtual second-order compensated feature flow. As an effective high-order feature channel, the selected effective high-order feature channel is re-routed and re-aligned with the original flow of the remaining physical branches that did not enter the reconstruction branch, and converged into a four-dimensional combined feature input tensor with spatial dimension and temporal residual synchronization.

[0020] S6 Blood Pressure Regression Prediction: Input the four-dimensional combined features after route realignment into the tensor and input it into the deep learning inference model to perform blood pressure regression prediction, and output systolic blood pressure SBP and diastolic blood pressure DBP;

[0021] S7 Closed-loop verification: Input the prediction result into the residual closed-loop verification module to determine the degree of deviation between the prediction result and the reconstructed oscilloscope envelope baseline. If the verification passes, the result is output; if it fails, a retest decision is triggered.

[0022] The phase alignment compensation in step S1 is achieved through a fixed-length delay digital buffer register, which compensates for the inherent phase shift and group delay of the discrete derivative algorithm, and realizes the zero-timescale misalignment alignment of physical feature flows of each order at the time origin.

[0023] The specific steps for obtaining the physical characteristic streams of four-channel phase-locked connections are as follows: A multi-order digital phase automatic compensator is used to precisely align and adjust each stream; the sampling interval is set to... The system's reference clock corresponding signal The input time. For those implemented in the order of The difference characteristic flow calculated by the moving average difference method Before feeding into the subsequent queue, a step size of 1 is forcibly performed. The shift register flush operation, in fact, follows the time-domain operation relationship. ,in For delay operators, This is a differential transfer function.

[0024] The frequency band quantization quality assessment in step S2 specifically includes:

[0025] The kurtosis coefficient, which is the deviation index of the normal distribution of the signal within the sliding window, is calculated. When the kurtosis coefficient deviates from the normal range, primary quality downweighting is triggered.

[0026] The fuzzy approximate entropy is calculated to measure the disorder and chaotic characteristics of the signal. When the score exceeds the critical threshold, the signal is marked as high-risk and distorted.

[0027] The frequency band energy concentration rate is calculated by short-time Fourier transform. When the proportion of high-frequency energy exceeds the baseline normal threshold, it is identified as a region with severe distortion of mechanical oscillation characteristics.

[0028] The specific steps for signal quality assessment are as follows:

[0029] For a duration of Calculating the normalized kurtosis coefficient of the data array within the sliding window This method measures the impact of broadband burst interference caused by sudden physical contact displacement on the system; it uses a fuzzy approximate entropy evaluation algorithm to measure the proportion of unpredictable, unsteady-state random fluctuations in the fluctuations; and it uses Fast Fourier Transform to statistically characterize specific frequency domain bands. Energy integral and extremely low frequency band The ratio of the internal energy integral If the kurtosis coefficient of a signal with a specific window length At the same time, the ratio of high-frequency and low-frequency structural components The rate of change increased sharply beyond the baseline deviation. If the physical branch is unreliable, the quality component evaluation score will be lower. .

[0030] The physiological dynamic reprojector in step S4 is obtained through pre-training in the following way: using a large amount of noise-free data from clinical subjects, high-dimensional mapping association rules between low-order displacement deformation state vectors and high-order momentum pulse change states in the same cardiac cycle are extracted through deep learning and solidified into nonlinear tensor mapping equations.

[0031] The expression for the nonlinear equivalent arterial transfer function operator is:

[0032]

[0033] in To parameterize and correct the nonlinear activation function layer, This represents the weight matrix from the input layer to the hidden layer. This represents the bias vector of the hidden layer. This represents the weight matrix from the hidden layer to the output layer. This represents the bias vector of the output layer; This represents the index of the current sampling time in the discrete-time series, corresponding to the [number]th [sample] time in the pressure pulse signal. One sampling point; The number of time taps indicating the duration of the playback; This is a dynamic gain adaptive adjustment factor; This represents the transpose operation of a vector or matrix.

[0034] The deep learning inference model described in step S6 is the MS-CNN inference engine, which includes heterogeneous parallel convolutional network branches: microscale sensitization branch: using a shallow network without downsampling pooling operation and a small-sized one-dimensional convolutional kernel array to extract high-frequency inflection point morphological features; mesoscale discrimination branch: extracting hydrodynamic features; macroscopic envelope branch: configuring a dilated convolution operator with a large hole rate and a large-scale average pooling layer to extract slow-wave pulse rhythm features.

[0035] After the MS-CNN inference engine performs feature extraction, a joint feature attention dynamic regression gating is introduced, based on the quality and safety evaluation vector. Perform adaptive weight allocation on the virtual second-order compensated feature flow. Apply a penalty decay factor.

[0036] The specific steps of feeding the multidimensional combined feature input tensor into a pre-trained multi-scale deep heterogeneous convolutional neural network include:

[0037] The input stream is processed by microscale sensitization branching to form second-order acceleration physical characteristics. It has a single-dimensional expansion operation array with a kernel size of 1×3 to 1×7, which is used to focus on characterizing the extremely small tidal wave signal at the millisecond level in the pulse phase.

[0038] The first-order velocity characteristic flow and the original pressure pulse wave flow are received by mesoscale temporal branch, and the dynamic trend line within a single heartbeat span is captured by the residual convolution block network kernel with different void ratios.

[0039] The slow leakage trend of cuff static pressure and the low-frequency spontaneous breathing drift and fluctuation pattern are extracted by macroscopic smoothing envelope branch. This branch is used in conjunction with a large average pooling array.

[0040] The specific operation of step S6 is as follows:

[0041] The evaluation vector is read in through a quality-perceived attention-gated network. The exclusive adjustment scalar weights for each major parallel path are calculated. For time window data that has experienced a strong reconstruction compensation mode, a confidence decay penalty function is introduced to cause the weight of the reconstructed branch to decrease exponentially, and conversely, the weight of the unreconstructed original waveform feature is increased to dominate the network decision.

[0042] The closed-loop verification in step S7 specifically includes:

[0043] Based on the estimated systolic and diastolic blood pressure, the theoretical smooth envelope template is generated by inverse deduction using the traditional oscilloscope amplitude statistical envelope function as the expected reference envelope baseline;

[0044] Extract the underlying morphological features of the original pulse wave envelope after primary low-pass filtering and smoothing, and align the two by area projection on the time axis;

[0045] The pointwise absolute residual integral between the theoretical template and the actual underlying envelope is calculated, and the relative residual deviation metric is calculated by normalization.

[0046] Set a rejection warning safety limit. When the residual measurement exceeds the safety limit, trigger a measurement rejection command and block the output display of the abnormal blood pressure calculation results.

[0047] Step S7 is performed as follows: The "expected noise-free morphological envelope baseline" is derived inversely using the Korotkoff tone-oscilloscope conversion statistical paradigm matrix. Compare the expected baseline with the actual underlying source envelope signal. If the residual integral between the two is compared with the metric parameter When this occurs, the system triggers a measurement rejection and blocking command, and links with the device's interactive terminal to issue a retest prompt that restricts motion remeasurement, creating a closed-loop interactive operation.

[0048] The method further includes configuring two sets of reconstruction mapping matrix operator pools with structural mirroring but independent weights for different anatomical acquisition positions of the upper arm monitor and the wrist wearable device.

[0049] The method further includes: collecting the user's continuous heart rate cycle feature matrix, determining whether the root mean square difference discrete variability (RMSSD) of adjacent cardiac cycles crosses the warning extreme value for atrial fibrillation; if the above arrhythmia exemption conditions are met and the output index of the external tremor noise auxiliary sensor does not significantly exceed the standard, then the high-order signal decay punishment scale is temporarily weakened and the reconstruction remediation action is suspended, and the original form is directly fed to the subsequent prediction network to ensure the retention of its pure natural original patient attribute fingerprint.

[0050] A blood pressure measurement device with physiological morphology reconstruction function includes: a multi-level feature synchronous extraction module configured to acquire the original cuff pressure pulse signal and simultaneously extract a four-dimensional temporal feature stream, including the original pressure pulse wave stream. First-order gradient feature flow Second-order gradient feature flow and physiological macromorphic envelope flow And perform phase alignment compensation;

[0051] The signal quality dynamic assessment and reconstruction judgment engine is configured to independently perform frequency band quantization quality assessment on each channel of the four-dimensional time-series feature stream, generate a quality safety assessment vector, and perform quality threshold judgment.

[0052] The physiological feature generation module is configured to isolate the contaminated channel when it is determined that the second-order excited-state feature channel is damaged, and generate a virtual second-order compensated feature flow in real time through a pre-trained physiological dynamic reprojector, using the original feature flow and the first-order velocity feature flow as input.

[0053] The routing realignment module is configured to perform routing selection and resynchronization of feature channels, select the original second-order excited-state feature stream or virtual second-order compensated feature stream as the effective high-order feature channel, and reassemble it to form a complete time-aligned four-dimensional input feature tensor.

[0054] The blood pressure regression prediction module is configured to input the four-dimensional input feature tensor after route realignment into the deep learning inference model to perform blood pressure regression prediction.

[0055] It also includes a residual closed-loop verification module, configured to perform inverse oscilloscope envelope verification, determine the degree of deviation between the prediction result and the baseline reconstruction, and trigger measurement rejection or output the final result.

[0056] The device further includes a quantization sensing deployment module configured to execute an asymmetric mixed-precision quantization strategy, wherein the high-frequency microscale channel maintains the INT16 or FP16 format, the macroscopic smoothing branch and attention control weights are quantized to the INT8 or INT4 format, and a low-level operator folding acceleration mechanism is introduced.

[0057] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the blood pressure measurement method.

[0058] An electronic device includes: a processor; and a memory for storing one or more programs; when the one or more programs are executed by the processor, the processor causes the processor to implement the blood pressure measurement method.

[0059] The beneficial effects obtained by this invention are:

[0060] (1) Significantly improved measurement accuracy under standard clinical conditions: Cross-channel morphological reconstruction effectively recovers and restores weak arterial elastic features at the aortic reflection echo level that were mistakenly rejected by traditional low-pass filtering or masked by background noise. These features are particularly sensitive and important in the usual extrapolation of diastolic pressure. This mechanism helps the deep computing fusion network obtain a high-fidelity full-dimensional physiological feature flow, which substantially constrains and reduces the random deviation of the overall measurement model's regressed pressure values; it also reduces the high-amplitude standard deviation constraint control previously caused by irregular environmental jitter to a more stable baseband (from the near-failure). The standard deviation value decreased to (between), so that it can stably meet the mandatory requirements of the internationally recognized ISO 81060-2 clinical independent subject measurement evaluation accuracy.

[0061] (2) Overcoming the limitations and continuity deficiencies of conventional discard-type noise reduction designs: Compared to existing adversarial strategies based on fragment "removal-rejection" (which are prone to long-term gaps leading to severe disorder of the attention mechanism and loss of representation points), this application scheme relies on the CSG computing pipeline to achieve "instantaneous in-situ equivalent substitution calculation of the instantaneously contaminated area". Even under severe body motion interference, it can still ensure that the three-dimensional tensor at the input analysis and prediction convolution operator maintains the integrity of the temporal slice structure and is coherent and undispersed throughout the frame, and the substitutes are all unbiased wave trains that conform to the characteristic wave train of cardiac biofluid dynamics mechanism. This step eliminates the large random extrapolation error introduced by "missing blank area data" from the source.

[0062] (3) Construction of a self-verifying, interlocking, fully closed-loop hard-core noise-resistant defense system: The phase lag alignment mechanism embedded in the backend offsets the inherent spatial and temporal differences caused by the step time shift in the original derivative calculation; further superimposed on the baseline regression verification rejection process of the post-execution judgment layer, a high-throughput evaluation barrier with interlocking layers of verification and high fault tolerance is established. This significantly improves the overall robustness of the miniature automatic pressure gauge product system against external interference. Attached Figure Description

[0063] Figure 1 This is a flowchart of a blood pressure measurement method based on reconstruction generation provided by an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of the time-domain adjustment mapping mechanism of the high-order digital phase automatic compensator provided in an embodiment of the present invention;

[0065] Figure 3 This invention provides a piecewise linear graph of attenuation determination of four independent quality assessment vectors under high-frequency strong noise and motion artifact interference scenarios.

[0066] Figure 4 This is the logical interconnection flow diagram of the core module for generating and reshaping virtual feature wave manifolds based on physiological dynamics provided in this embodiment of the invention;

[0067] Figure 5 This is a block diagram of multi-scale deep convolutional inference brainstem decomposition with an integrated quality-aware control weighting mechanism provided in this embodiment of the invention.

[0068] Figure 6 This is a hardware electronic communication layout and wiring diagram of the device hardware provided in this embodiment of the invention, including front-end sampling, edge fusion scheduling NPU and residual interception closed-loop structure. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. Unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other and derived as equivalents.

[0070] The core design logic of this invention is: "Following the law of conservation and correlation of cardiac circulatory dynamics, we perform cross-order physical morphological reconstruction compensation." This architecture overcomes the limitation of simply discarding damaged features in traditional methods. Because the waveform branching features of different orders within human arteries (displacement potential energy state, flow velocity kinetic energy state, and abrupt acceleration jump state) all originate from the same source—cardiac contraction and the wall rebound caused by the closure of the aortic semilunar valve—they exhibit a stable physiological synchronous correlation. Based on this physiological foundation, when sensing extreme body motion noise causes severe distortion of high-order features rich in minute details but susceptible to interference, the system can reverse-call the relatively low-frequency, noise-resistant first-order displacement deformation flow, and reconstruct and losslessly reproduce the high-order physical fingerprint inflection point features in the data-missing region based on a specific cardiovascular elastic mapping function.

[0071] The high-dimensional blood pressure measurement system proposed in this invention, which integrates physiologically consistent reconstructed defense features, adopts the following technical solution:

[0072] First, the system synchronously and in parallel acquires and aligns four-dimensional (original pressure pulse wave flow, first-order gradient feature flow, second-order gradient feature flow, and physiological macroscopic morphological envelope flow) temporal feature flows with time-locked matching and zero-delay phase latching. To ensure the signal accuracy of subsequent differential operations, the system uses a high-precision analog-to-digital converter with an effective resolution of 24-bit for multi-channel synchronous sampling, and strictly locks the sampling clock based on a hardware timer to maintain... A fixed high-frequency sampling rate outputs a high-resolution raw digital sequence. .

[0073] For feature extraction from the original digital sequence, the system's built-in hardware digital signal processor (DSP) performs three concurrent computations: the first path uses a polynomial difference fitting operator to generate a first-order gradient feature vector representing blood flow velocity. and the second-order gradient characteristic flow representing the acceleration of blood vessel wall motion. The second approach utilizes the short-window Hilbert transform to extract the physiological macroscopic morphological envelope flow. Used for baseline drift filtering.

[0074] The group delay caused by the inherent phase shift and time-domain smoothing operator in the discrete derivative algorithm can lead to a millisecond-level time base offset between the extracted higher-order derivative feature vector and the original signal at the same physiological pulsation starting point, thus causing systematic errors during multidimensional feature splicing.

[0075] To address this, phase alignment compensation is achieved by introducing a digital buffer register with a fixed-length delay into the signal pipeline. For example, if the total cumulative delay of the second-order acceleration feature stream compared to the original signal is... Before feature fusion, an equivalent delay is synchronously applied to the original signal branches that do not cause algorithmic delay. The delay window forces the four-dimensional temporal feature flow to achieve zero-timescale misalignment alignment at the time origin, thereby constructing a high-fidelity spatiotemporal polymorphic feature tensor with synchronized dimensions and consistent phase, which can be accurately parsed by the subsequent convolutional neural network.

[0076] As attached Figure 3 As shown, this embodiment employs a frequency-band quantization quality assessment mechanism. Unlike traditional coarse-grained rule-based judgments based on global or single-image values, this system performs refined path-by-path quantization evaluation of the signal interference level within the four time-series feature streams, generating independent evaluation column vectors. .

[0077] Second-order gradient characteristic flow rich in high-frequency components An example is used to illustrate its interference detection mechanism (e.g., transient broadband Gaussian white noise introduced by coughing or external vibration). The monitoring thread calculates its quality and safety factor. A series three-phase combination rating and screening method was designed and a comprehensive downgrade score was output:

[0078] (1) Calculate the normality deviation index, i.e., kurtosis coefficient, for the dataset within the sliding window (e.g., window length 2 seconds). When broadband smoothing artifact interference occurs, the peak distribution characteristics deviate; if determined... Or subject to extreme shocks This will trigger a primary quality downgrade.

[0079] (2) Simultaneously introduce nonlinear dynamic complexity assessment methods, namely, measure and evaluate the fuzzy approximate entropy of the disorder and chaotic characteristics of the signal. This measures the proportion of random burst components within the current analysis window. When unpredictable jitter caused by noise leads to... The score rose above the critical threshold. If this occurs, the channel is marked as facing high-risk signal distortion and degraded to poor quality.

[0080] (3) Activate the frequency domain energy ratio scanning interception threshold: Calculate the frequency band energy concentration rate using short-time Fourier transform to suppress sudden white noise. Evaluation Accumulated energy in the high-pass region The ratio of extremely low frequency core energy; if the proportion of this high frequency energy exceeds the baseline normal threshold... This indicates a region with severely distorted mechanical vibration characteristics.

[0081] When feature channels are determined to be damaged and disrupted anomalous data must be isolated to prevent prediction bias in the neural network, this invention provides an adaptive, fully automatic channel compensation mechanism to effectively restore the damaged network. This compensation mechanism relies on a low-power, lightweight pre-computed weight model operator deployed in processor memory. (i.e., Arterial Equivalent Project Operator). This model operator is obtained by pre-training with a large amount of noise-free data from clinical subjects. It solidifies and represents the high-dimensional mapping association rules between low-order displacement deformation state vectors and high-order momentum pulse change states in the same cardiac cycle through deep learning, and represents them as nonlinear tensor mapping equations.

[0082] When the corresponding Signal Quality Index (SQI) The reliability of the quantity Falling below the fault tolerance threshold Subsequently, the system will block the interfered input channel. At this time, the processor extracts the original pressure pulse wave flow that remains robust due to its anti-interference performance during the same period. and first-order gradient feature flow (The corresponding SQI values ​​are all above the safety threshold), and are used as input feature tensors to the reprojection generation unit for real-time inference calculation. The reprojection generation unit is a nonlinear mapping module used to perform calculations based on the original pressure pulse wave flow. and first-order gradient feature flow Reconstruct the corresponding second-order gradient feature flow .

[0083] A novel virtual second-order compensated feature flow is output through fast forward computation of the model. Based on the extrapolation fitting ability of first-order smooth characteristic curves, it reproduces and re-characterizes the temporal features of the dynamic tension representing arterial compliance that have been obscured by environmental noise (such as the diabetic microwave inflection point and the descending isthmus groove). See attached... Figure 6 As shown, the synthesis result achieves an equivalent restoration of physiological characteristic information within the damaged region.

[0084] Figure 5 The configuration of the AI ​​inference analytical model group used for the final multi-level regression prediction of blood pressure values ​​is shown. This embodiment abandons the traditional structure of a single stacked convolution operator and instead constructs a heterogeneous parallel analytical channel specifically for four data stream branches with different time-frequency characteristics.

[0085] Among them, the micro-determining fine-grained branch (A channel), specifically used for extracting high-frequency weak mutation characteristics in higher-order derivative features (covering both original second-order and reconstructed virtual second-order features), deploys a shallow network without downsampling pooling operations, employing only... to Small-sized one-dimensional convolutional kernel arrays are used to ensure accurate extraction and preservation of subtle morphological variables such as high-frequency pulse surge vertices and descending mid-slope grooves.

[0086] The macroscopic morphological envelope sensing branch (C channel) responsible for extracting the slow wave trend of the original blood pressure envelope and the baseline drift information of deep breathing is configured with a dilated convolution operator with a large hole rate and cascaded with a large-scale average pooling layer array to extract large-scale pulse rhythm and long-cycle decay trajectory of cuff static pressure across multiple cardiac cycles.

[0087] Before feature extraction from multiple branches at the front end is completed and the feature fusion prediction is performed in the fully connected decoding layer, the system introduces a quality-aware attention-based dynamic constraint gating mechanism (QAG gating module). This module reads the independent quality and safety assessment vectors of each channel generated by the system in the previous stage in real time. Based on this evaluation vector, a nonlinear normalized excitation decay model (e.g., with a penalty damping factor) is constructed. The mapping function redistributes the attention inference weights of each feature branch. Specifically, when a branch feature tensor is detected to contain a virtual feature fragment reconstructed and compensated by the CSG module, the system will apply a proportional influence penalty decay factor (e.g., configure a decay multiplier) based on the temporal proportion of the virtual fragment. Reduce the decision weight of this branch to that of the original signal. Correspondingly, the system will adaptively increase the weight of the original feature branches that are uncontaminated and have a high signal-to-noise ratio, thereby guiding the model to rely more on data branches with high safety in the final blood pressure regression calculation.

[0088] To ensure the clinical reliability of measurements, the system also incorporates a fault-tolerant closed-loop and baseline verification feedback mechanism; the specific baseline verification steps are as follows:

[0089] First, based on the predicted systolic blood pressure (SBP) and diastolic blood pressure (DBP) output by the deep model, the system performs inverse derivation by substituting them into the traditional oscilloscope amplitude statistical envelope function, generating a set of theoretical smooth envelope templates corresponding to the predicted blood pressure values, defined as the "expected reference envelope baseline". .

[0090] Subsequently, the system extracts the underlying morphological features of the original pulse wave envelope after primary low-pass filtering and smoothing. Then, the two are aligned by area projection on the time axis. Calculation theoretical template With the actual underlying envelope The relative residual deviation metric is calculated by integrating the point-by-point absolute residuals and normalizing them. The system sets strict rejection warning safety limits (e.g., when residual measurement...). (Time); Once the limit is exceeded, it indicates that there is a serious decrease in confidence of the current deep model inference or that unidentified continuous interference has occurred. The main control system will actively trigger a measurement rejection command, block the output display of the abnormal blood pressure calculation results, and simultaneously issue a safety operation warning of "Too much interference, it is recommended to retest" through the user interface of the device terminal.

[0091] In terms of clinical engineering device deployment, this pressure measurement system is compatible with both traditional upper arm ambulatory blood pressure monitors and novel wrist-worn wearable devices. Given the differences in the anatomical placement of these two device forms, which lead to fundamental physical distortions in the wave reflection mechanism, the CSG module in this embodiment of the invention specifically embeds two sets of structurally mirrored but weighted independent reconstruction mapping matrix operator pools.

[0092] For upper arm monitoring devices (targeting the brachial artery), the reconstructed unit scheduling loads a dedicated mapping matrix that focuses on characterizing peripheral reflexes that are less pronounced at the periphery and arterial trunk elasticity that is dominant. The reconstructed nonlinear waveform characteristics tend to characterize the smooth, wave-like dilatational waves that appear after the intima of the large artery is compressed and released.

[0093] For wrist-worn wearable devices (targeting the distal radial and ulnar arteries), mechanical friction noise is introduced to account for the increased susceptibility of multi-axis body movements at the wrist. CSG, on the other hand, uses a pre-simulation of the radial artery mapping matrix based on the microvascular distribution and distal pulse wave reflection delay impedance model. Furthermore, during feature evaluation and assembly, the quality lower limit of the system's QAG attention gating was set to be more stringent (i.e., This allows for more extreme isolation strategies to be implemented for channels with weak signal-to-noise ratios in order to ensure the stability of end-point blood pressure prediction.

[0094] In response to the audit requirements for wearable measurement traceability in medical certification regulations, in addition to deploying micro-pump valve drivers, sensor acquisition front-ends and power modules, the main control system of the device also has an additional hardware watchdog and a security operation log recording module consisting of built-in encrypted non-volatile flash memory.

[0095] During the real-time acquisition process, any severely disrupted abnormal waveform data that is significantly downweighted and isolated by the QAG gating module, or high-order virtual feature segments and their corresponding weight allocation vectors that have been repaired through CSG physiological reconstruction compensation mechanisms, will be stamped with a high-precision real-time synchronization timestamp (RTC) and permanently archived and backed up to the security operation log recording module. This operation completely makes the intermediate adaptive state of the nonlinear AI neural network transparent and logged, in order to meet the reliability traceability requirements of international medical device compliance reviews.

[0096] To achieve low-latency inference in portable wearable devices under conditions of limited battery power and micro-memory edge computing, this invention designs a quantization deployment framework for performing asymmetric dimensionality reduction on heterogeneous multi-branch networks: In a cloud computing center, relying on a large computing cluster, the mid-to-late stages of full-precision weight training for deep neural networks (global empirical loss rate) are completed. As the system approaches the convergence bottom edge and enters the fine-tuning window, it will initiate the quantization-aware training mechanism. This engine innovatively executes an asymmetric mixed-precision quantization policy based on the sensitivity attributes of different task metrics of the multi-branch parallel extraction architecture described above.

[0097] Given the high sensitivity of the high-frequency microscale channel (A channel), which is specifically designed to capture the characteristics of extremely small tidal waves and high-order harmonics in the descending strait, to micro-pressure gradients, and to avoid the truncation error during numerical fixed-point approximation erasing subtle physiological turning points, the system forces the convolution weight matrix and feature activation response maps within this channel to maintain a high lower bound on the data width during model solidification. Specifically, the channel is forcibly calibrated to either 16-bit integer (INT16) or the half-precision floating-point (FP16) format natively supported by the edge computing end.

[0098] For the macroscopically smoothed branch (i.e., the C channel) extracted from the long-period envelope trend, and the discrete attention control weights (Weight Activations) of the QAG gating module guided by the quality and safety coefficients of each path; due to its robust low-frequency gradual variation and insensitivity to numerical jitter, the system applies an extreme discrete layer compression ratio. When distributing the encapsulated model map to the edge, the fully connected parameter layers and convolutional kernel weights associated with such network modules are aggressively pruned and fixed-point quantized to 8 bits (INT8) or even a very low bit width of 4 bits (INT4) coarse-grained integer specification, in order to maximize the conservation of on-chip SRAM working cache and logic gate addressing area.

[0099] When loading the forward computation flow at the target edge microprocessor chip, the system introduces a low-level operator folding acceleration mechanism: at the processing logic level, dense linear convolutional filters and their batch normalization operation layer are mathematically absorbed and merged into a single instruction-level operation, and executed in a single clock cycle (Conv-BN Fusing); at the same time, the nonlinear mapping network layer (such as Sigmoid or Softmax) is subjected to an equivalent linear mapping pre-table operation, which is converted into a pre-stored hardware look-up table (LUT) with no dynamic floating-point operation overhead and only generating address index power consumption. Through the above dual-drive strategy of hybrid quantization and low-level convergence acceleration, the peak memory capacity and total computational power consumption of the blood pressure inference network per cycle show a dramatic drop in magnitude, ensuring that the high-dimensional, high-rank complex combination of disturbance rejection models can perfectly reside and run continuously on the miniature blood pressure monitoring wristband dial with a button-level power supply.

[0100] The offline tolerance confidence interval evaluation pipeline was designed in accordance with the Bland-Altman statistical standard for clinical evaluation in medicine. When the offline training metrics of the model reach the convergence bottleneck, a final error test is performed using a blind-box testing dataset of medical pulse diagnosis data that is completely independent of the training set.

[0101] At the end of the model evaluation pipeline, an automated consistency limit analysis model is introduced to extract the pressure divergence boundary between the reference mercury sphygmomanometer label and the calculated value. This is then statistically plotted... Consistency limit range (i.e.) This is used to prove that no matter how complex and severe the dynamic and generalized health scenario of the network is, the estimated mean error (ME) of the high and low voltage will never drift out. Furthermore, the discrete standard deviation (SD) is convergentally suppressed to the limiting threshold. Within the defined scope. The robust self-certification record generated by this mechanism ensures that this solution highly matches the latest international medical AAMI verification and ISO-81060 standard certification specifications from the software algorithm verification perspective.

[0102] To address the extremely challenging physical conditions requiring acoustic-optical-electric coupling (such as localized high-frequency hard reflections and unstable mechanical vibrations caused by conditions like excessive hair growth or keratosis in the arms of patients with specific illnesses), this system further incorporates an ultra-high-dimensional feature-level bypass adversarial filtering module based on a multidimensional wavelet manifold space. This branch focuses on eliminating sudden, broadband, non-stationary mechanical vibration noise, thereby preventing the overall temporal pool data from being overlaid by abnormal random burst noise and resulting in complete inference collapse.

[0103] This supplementary algorithm module bypasses traditional one-dimensional high-order difference operations in the discrete-time domain, employing a compactly supported mother wavelet family (such as the Morlet wavelet or Daubechies family) that supports scale scaling and time translation to perform a high-resolution continuous wavelet transform (CWT) on the damaged pulse wave across the entire time domain. This algorithm can extract broadband mechanical vibration abnormalities and physiologically smoothed pulse wave sources aliased in the time domain to a joint time-frequency mapping plane. Through second-order expansion operations, the one-dimensional aliased time series signal is decoupled and projected into a two-dimensional time-frequency thermogram array containing energy focusing, time-domain localization, and frequency band distribution characteristics. In this high-dimensional orthogonal mapping space, the underlying background wave of the long-period pulse envelope hidden at the lower level and the non-stationary high-frequency abrupt distortion artifacts superimposed on it can achieve feature-level spatial domain separation.

[0104] After generating the CWT 2D time-frequency heatmap, a custom-designed non-uniform spatial denoising convolutional residual network is cascaded into the system. This network performs feature isolation and removal within the pixel-level space of the 2D frequency band. Because the harsh mechanical vibrations transmitted to the system from cuff slippage or external impacts manifest as illogical truncated horizontal lines, independent high-potential spots, or non-gradual distribution patterns lacking smooth extension in the frequency domain on the CWT frequency plane, there is a significant spatial pattern difference compared to the physiological pulse spectrum cloud region formed by cardiovascular pumping, which exhibits continuous energy smooth transition, attenuation tails, and a naturally smooth halo topological phase. Image semantic segmentation and denoising processing via this topological convolutional network can accurately remove abrupt non-physiological artifact patches with extremely low signal loss.

[0105] After completing the deep spatial frequency domain denoising, the filtered, pure two-dimensional time-frequency feature blocks will directly undergo inverse operations (i.e., inverse continuous wavelet transform, ICWT), or be processed by nonlinear feature pooling to generate a set of bypass disaster-resistant generalization features that completely isolate destructive random noise. This high-order pure feature pool is synchronously bypassed and incorporated into the back-end fully connected inference and decision module, completing a multi-modal reconstruction and disaster-resistant protection closed loop that extends from the one-dimensional time domain to the two-dimensional frequency domain.

[0106] To further verify the actual technical progress effect of the "cross-flow physiological reconstruction" architecture described in this invention on improving measurement stability and accuracy, this embodiment uses an ablation test to quantitatively evaluate the contribution of each core module.

[0107] The verification experiment was conducted using four control implementation groups:

[0108] Group 1 (A1, benchmark comparison group): adopts conventional multi-scale convolutional architecture, only inputs raw pulse signal and pressure sequence, and does not include the reconstruction module of this invention;

[0109] Group 2 (A2, MDF Access Group): Based on Group 1, the Multidimensional Feature Flow (MDF) processing path is introduced, that is, the first-order and second-order physical flows are parsed simultaneously;

[0110] Group 3 (A3, SQI Access Group): Based on Group 2, the real-time evaluation mechanism of the Signal Quality Index (SQI) is further introduced;

[0111] Group 4 (A4, Full-Stack Implementation Group): Based on Group 3, introduce the virtual physical flow generated by the Physiological Reconstruction Engine (CSG) and activate the dual attention logic based on quality constraints.

[0112] The quantitative performance statistics of the measurement error distribution for each group on the validation set of subjects, which includes various physiological rhythm fluctuations and interference scenarios, are shown in the table below:

[0113]

[0114] A comparison of the characteristics of Group 1 and Group 2 shows that after introducing the phase-locked MDF architecture, the discrete standard deviation (SD) of the systolic blood pressure (SBP) converges significantly (reducing by about 3 mmHg). This proves that by constructing a multidimensional physical constraint flow, error divergence caused by random disturbances can be effectively suppressed.

[0115] Comparing the data from Group 3 (A3) and the full-stack implementation group (A4), it is evident that simply using Signal Quality Identification (SQI) for outlier gating / removal without implementing physiological reconstruction compensation leads to severe degradation of inference metrics when the model is contaminated due to discontinuous sampling windows (e.g., a significant rebound in the MAE dimension in Group A3). This technological contrast directly confirms that utilizing the cardiac circulation dynamics conservation correlation law for in-situ reconstruction and compensation of physiological flow not only eliminates the bias introduced by "black box feature loss," but also substantially recovers and significantly improves the overall measurement accuracy under interference conditions while ensuring signal integrity.

[0116] The aforementioned comparative experiments strongly support the advanced nature of the technical solution of this invention. Through the deep integration of MDF throughput and CSG reconstruction engine, this invention successfully solves the technical bottleneck of unusable high-frequency features in portable blood pressure measurement under real-world, high-interference scenarios. The experimental data clearly demonstrate the synergistic improvement in measurement consistency and accuracy, providing solid engineering verification support for achieving highly reliable, all-weather vital sign monitoring.

Claims

1. A blood pressure measurement method with physiological morphology reconstruction function, characterized in that, The method includes: S1 Multi-order Feature Synchronous Extraction: Acquires the raw cuff pressure pulse signal and simultaneously extracts the four-dimensional temporal feature stream, including: the raw pressure pulse wave stream. First-order gradient feature flow Second-order gradient feature flow And the physiological macromorphic envelope flow obtained through the generalized Hilbert transform. A total of four phase-locked physical feature streams are constructed. Among them, phase alignment compensation is performed on the four-dimensional temporal feature streams through a digital buffer register with a fixed-length delay to achieve zero-timescale misalignment alignment and construct a high-fidelity spatiotemporal polymorphic feature tensor with dimension synchronization and phase consistency. S2 Signal Quality Assessment: The phase-aligned four-dimensional time-series feature stream is input into the signal quality dynamic assessment and reconstruction engine. The engine independently calculates the statistical morphological consistency characteristics, local energy proportion characteristics in the time-frequency domain, and nonlinear dynamic signal entropy characteristics of each feature stream within the current window. The resulting fusion output is a high-dimensional local quality assessment vector used to characterize the signal-to-noise ratio safety margin in real time. ;in Indicates the original pressure pulse wave flow Signal quality metrics; Represents the first-order gradient feature flow Signal quality metrics; Represents the second-order gradient feature flow Signal quality metrics; Represents the physiological macromorphic envelope flow The signal quality index; each of the signal quality indices is used to characterize the reliability of the corresponding feature stream within the current time window, and its value is preferably a normalized scalar, with a value range of 0 to 1, where the closer the value is to 1, the higher the signal quality, and the closer the value is to 0, the higher the degree of interference to the signal. S3 Quality Threshold Determination: Judgment Is it below the preset time series robustness safety lower limit? ,and Above the safety limit ; If the result is "no", proceed to step S5 and execute the conventional parallel feature extraction route. If the result is "yes": proceed to step S4 to trigger the dynamic physiological remodeling mechanism; S4 Physiological Feature Reconstruction: Isolating Contaminated Second-Order Gradient Feature Flow The first-order gradient feature flow that meets the quality standard and the original pulse flow As a priori reference benchmark, virtual second-order compensated feature flows are inverted, predicted, and synthesized in real time by loading a predefined nonlinear equivalent arterial transfer function operator. To replace the second-order gradient feature flow that is masked by environmental noise. ; S5. Route realignment step: Perform route selection and resynchronization for the feature channel; if determined as "yes" by S3: select the second-order gradient feature flow. As an effective high-order feature channel; if after S4 reconstruction: select virtual second-order compensated feature flow. As an effective high-order feature channel; the selected effective high-order feature channel is combined with the original feature flow. First-order velocity characteristic flow Macroscopic morphological envelope Perform route reorganization to form a complete time-aligned four-dimensional input feature tensor; S6 Blood Pressure Regression Prediction: Input the four-dimensional input feature tensor after the route is re-aligned into the deep learning inference model, perform blood pressure regression prediction, and output systolic blood pressure SBP and diastolic blood pressure DBP; S7 Closed-loop verification: Input the prediction result into the residual closed-loop verification module to determine the degree of deviation between the prediction result and the reconstructed oscilloscope envelope baseline. If the verification passes, the result is output; if it fails, a retest decision is triggered.

2. The method according to claim 1, characterized in that, The frequency band quantization quality assessment in step S2 specifically includes: The kurtosis coefficient, which is the deviation index of the normal distribution of the signal within the sliding window, is calculated. When the kurtosis coefficient deviates from the normal range, primary quality downweighting is triggered. The fuzzy approximate entropy is calculated to measure the disorder and chaotic characteristics of the signal. When the score exceeds the critical threshold, the signal is marked as high-risk and distorted. The frequency band energy concentration rate is calculated by short-time Fourier transform. When the proportion of high-frequency energy exceeds the baseline normal threshold, it is identified as a region with severe distortion of mechanical oscillation characteristics.

3. The method according to claim 1, characterized in that, The physiological dynamic reprojector in step S4 is obtained through pre-training in the following way: using a large amount of noise-free data from clinical subjects, high-dimensional mapping association rules between low-order displacement deformation state vectors and high-order momentum pulse change states in the same cardiac cycle are extracted through deep learning and solidified into nonlinear tensor mapping equations.

4. The method according to claim 1, characterized in that, The deep learning inference model described in step S6 is the MS-CNN inference engine, which includes a heterogeneous parallel convolutional network branch: a microscale sensitization branch: a shallow network with no downsampling pooling operation and a small-sized one-dimensional convolutional kernel array to extract high-frequency inflection point morphological features; Mesoscale discrimination branch: extracts hydrodynamic features; macroscopic envelope branch: configures a large-void-rate dilated convolution operator and a large-scale average pooling layer to extract slow-wave pulse rhythm features.

5. The method according to claim 4, characterized in that, After the MS-CNN inference engine performs feature extraction, a joint feature attention dynamic regression gating is introduced, based on the quality and safety evaluation vector. Perform adaptive weight allocation on the virtual second-order compensated feature flow. Apply a penalty decay factor.

6. The method according to claim 1, characterized in that, The closed-loop verification in step S7 specifically includes: Based on the estimated systolic and diastolic blood pressure, the theoretical smooth envelope template is generated by inverse deduction using the traditional oscilloscope amplitude statistical envelope function as the expected reference envelope baseline; Extract the underlying morphological features of the original pulse wave envelope after primary low-pass filtering and smoothing, and align the two by area projection on the time axis; The pointwise absolute residual integral between the theoretical template and the actual underlying envelope is calculated, and the relative residual deviation metric is calculated by normalization. Set a rejection warning safety limit. When the residual measurement exceeds the safety limit, trigger a measurement rejection command and block the output display of the abnormal blood pressure calculation results.

7. The method according to claim 1, characterized in that, Also includes: For the different anatomical acquisition locations of upper arm monitors and wrist wearable devices, two sets of reconstruction mapping matrix operator pools with structural mirroring but independent weights are configured.

8. A blood pressure measuring device with physiological morphology reconstruction function, characterized in that, include: The multi-level feature synchronous extraction module is configured to acquire the raw cuff pressure pulse signal and simultaneously extract the four-dimensional temporal feature stream, including the raw pressure pulse wave stream. First-order gradient feature flow Second-order gradient feature flow and physiological macromorphic envelope flow And perform phase alignment compensation; The signal quality dynamic assessment and reconstruction judgment engine is configured to independently perform frequency band quantization quality assessment on each channel of the four-dimensional time-series feature stream, generate a quality safety assessment vector, and perform quality threshold judgment. The physiological feature generation module is configured to isolate the contaminated channel when it is determined that the second-order excited-state feature channel is damaged, and generate a virtual second-order compensated feature flow in real time through a pre-trained physiological dynamic reprojector, using the original feature flow and the first-order velocity feature flow as input. The routing realignment module is configured to perform routing selection and resynchronization of feature channels, select the original second-order excited-state feature stream or virtual second-order compensated feature stream as the effective high-order feature channel, and reassemble it to form a complete time-aligned four-dimensional input feature tensor. The blood pressure regression prediction module is configured to input the four-dimensional input feature tensor after route realignment into the deep learning inference model to perform blood pressure regression prediction. It also includes a residual closed-loop verification module, configured to perform inverse oscilloscope envelope verification, determine the degree of deviation between the prediction result and the baseline reconstruction, and trigger measurement rejection or output the final result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the blood pressure measurement method as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: processor; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the processor, the processor causes the processor to implement the blood pressure measurement method as described in any one of claims 1-7.

Images