A method and system for measuring arteriovenous fistula pressure based on intelligent sensors

Abstract

This invention discloses a method and system for measuring arteriovenous fistula pressure based on intelligent sensors, belonging to the field of non-invasive medical monitoring technology. The method involves deploying an intelligent sensor on the body surface at the proximal end of the arteriovenous fistula to filter out multimodal physiological noise and separate signal components from the acquired mixed signals, obtaining a high-fidelity fistula pressure waveform. The waveform is then automatically calibrated for feature points and averaged over multiple cycles to generate a standard periodic waveform template. Combined with a pre-set hemodynamic physical model, this template is used to perform inverse calculations to deduce a set of core mechanical parameters reflecting the internal state of the fistula. By comparing this parameter set with a historical baseline database, a dynamic evaluation matrix is ​​established, ultimately generating functional grading and risk warnings. This method achieves accurate separation of the target waveform from the body surface signal and inverse vectorization of the internal mechanical state of the fistula, improving the objectivity and early warning capabilities of non-invasive monitoring.

Description

Technical Field

[0001] This invention belongs to the field of non-invasive medical monitoring technology, specifically a method and system for measuring arteriovenous fistula pressure based on intelligent sensors. Background Technology

[0002] Non-invasive monitoring of arteriovenous fistula (AVF) function is a crucial aspect of hemodialysis treatment. Current technologies primarily rely on manual palpation, auscultation, or periodic ultrasound imaging by healthcare professionals for qualitative assessment. Recent advancements in surface sensor monitoring methods typically involve directly acquiring vibration signals from the skin surface and interpreting them as fistula pressure waveforms. However, human surface signals are a mixture of multiple physiological noise sources, including heartbeats, respiration, muscle activity, and disturbances from nearby blood vessels. Conventional electronic filtering techniques can only attenuate based on frequency and cannot distinguish and separate the pressure component purely originating from the target AVF from this multimodal mixed signal. This results in distorted waveforms, blurred features, and insufficient stability and reliability of the measurement results.

[0003] Even with relatively clear waveforms, existing analysis methods mostly stop at extracting and tracking intuitive morphological parameters such as waveform amplitude and period. These methods lack a quantitative physical model that links surface waveforms to the hemodynamic state within the fistula, and cannot deduce the core mechanical parameters reflecting vascular elasticity, blood flow resistance, and flow rate changes. This limits assessment to superficial observation, making it difficult to achieve early, objective, and quantitative diagnosis of internal pathological changes such as fistula stenosis and abnormal flow, and failing to meet the clinical needs for continuous dynamic monitoring and precise early warning. This invention aims to solve the technical problem of high-fidelity extraction of fistula pressure signals from mixed surface signals and inversely solving for core hemodynamic parameters from the waveform based on a physical model. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes a method for measuring arteriovenous fistula pressure based on a smart sensor, comprising:

[0006] A smart sensor is deployed at the corresponding position on the proximal end of the target arteriovenous fistula, and multimodal physiological noise filtering is performed to separate the signal components and generate noise-reduced fistula pressure waveform data.

[0007] The noise-reduced arteriovenous fistula pressure waveform data is subjected to automatic feature point calibration processing to identify the systolic initiation point, peak point, diastolic trough point and dicrotic notch point within the waveform cycle, and to construct a single-cycle waveform feature point sequence.

[0008] Based on the single-cycle waveform feature point sequence, multi-cycle waveform morphology alignment and averaging are performed to generate a standard periodic pressure waveform template that characterizes the local hemodynamic state of the arteriovenous fistula.

[0009] By combining the preset hemodynamic physical model with the standard periodic pressure waveform template, the waveform parameters are solved in reverse to obtain the set of core mechanical parameters reflecting the pressure changes inside the fistula.

[0010] Based on the comparative analysis of the core mechanical parameter set and the historical baseline parameter database, a dynamic evaluation matrix for arteriovenous fistula pressure status is established;

[0011] Based on the dynamic assessment matrix of arteriovenous fistula pressure status, a functional status classification description and pressure abnormality risk warning are generated for the target arteriovenous fistula.

[0012] Furthermore, the smart sensor is deployed at the corresponding position on the proximal end of the target arteriovenous fistula on the body surface, and performs multimodal physiological noise filtering processing to separate the signal components and generate noise-reduced fistula pressure waveform data, including:

[0013] Deploy a smart sensor with pressure sensing capability at the corresponding position on the proximal end of the target arteriovenous fistula, trigger the smart sensor to collect continuous raw pressure oscillation signals within the target time period, and obtain an initial set of physiological pressure signals;

[0014] Multimodal physiological noise filtering is performed on the initial physiological pressure signal set to separate the signal components originating from arterial pulsation, venous return and tissue background noise, and generate denoised arteriovenous fistula pressure waveform data;

[0015] The multimodal physiological noise filtering process performed on the initial physiological stress signal set includes:

[0016] The initial set of physiological pressure signals is input into an adaptive noise reference signal generator, which generates simulated arterial pulsation noise reference signals, venous return noise reference signals, and tissue motion noise reference signals simultaneously based on signal spectrum characteristics and a preset noise source model.

[0017] The simulated arterial pulsation noise reference signal, venous return noise reference signal, and tissue motion noise reference signal are respectively subjected to coherence analysis and adaptive weight adjustment with the initial physiological pressure signal set to construct a composite noise estimation signal that highly matches the real noise.

[0018] The composite noise estimation signal is subtracted from the initial physiological pressure signal set in the time domain to obtain the preliminary purification pressure signal;

[0019] The preliminary purified pressure signal is subjected to nonlinear wavelet threshold denoising to further suppress residual high-frequency random noise, and finally the denoised arteriovenous fistula pressure waveform data is output.

[0020] Furthermore, the automatic feature point calibration processing of the denoised arteriovenous fistula pressure waveform data includes:

[0021] Perform first-order and second-order differential calculations on the denoised arteriovenous fistula pressure waveform data to obtain the waveform change rate curve and acceleration curve;

[0022] Scan the waveform rate of change curve, identify the zero-crossing point where the rate changes from negative to positive and the acceleration is positive, and mark it as the starting point of the contraction period;

[0023] Within a preset time window following the start point of the contraction period, the global maximum value of the pressure amplitude is found and marked as the peak point.

[0024] After the peak point, search for the first point in the waveform rate of change curve that satisfies the preset negative rate threshold, and mark it as the diastolic valley point;

[0025] Between the peak point and the diastolic trough point, the local extreme points of the acceleration curve are analyzed, and the inflection point characteristics of the waveform are combined to determine and mark the dilatation notch.

[0026] Furthermore, the multi-cycle waveform morphology alignment and averaging process based on the single-cycle waveform feature point sequence includes:

[0027] Collect the sequence of single-cycle waveform feature points corresponding to multiple consecutive cycles, and use the starting point of the contraction period of each cycle as the time reference point to normalize and align the pressure waveform data in all cycles.

[0028] Calculate the mean and variance of the amplitude of all periodic pressure waveforms at each time point after alignment, and generate the average pressure waveform curve and the corresponding fluctuation range band.

[0029] The average pressure waveform curve is smoothed by interpolation to eliminate edge irregularities caused by slight differences in period length, forming a standard periodic pressure waveform template with regular shape and statistical representativeness.

[0030] Furthermore, the step of combining the preset hemodynamic physical model with the standard periodic pressure waveform template to perform inverse waveform parameter calculation includes:

[0031] The standard periodic pressure waveform template is input into a hemodynamic lumped parameter model that includes parameters of vascular elasticity, blood inertia, and peripheral resistance.

[0032] The gradient descent optimization algorithm is used to iteratively adjust the unknown mechanical parameters in the hemodynamic lumped parameter model so that the mean square error between the theoretical pressure waveform output by the model simulation and the standard periodic pressure waveform template is minimized.

[0033] When the mean square error converges to within a preset threshold, the iteration is terminated, and the corresponding vascular elasticity parameters, blood inertia parameters, and peripheral resistance parameters in the hemodynamic lumped parameter model at this time are exported.

[0034] Based on the theoretical relationship between the derived vascular elasticity parameters and pulse wave propagation, the pulse wave propagation velocity is calculated, and the corresponding proximal systolic pressure and mean pressure of the arteriovenous fistula are directly extracted from the simulation waveform, which together constitute the core set of mechanical parameters.

[0035] Furthermore, the establishment of a dynamic assessment matrix for arteriovenous fistula pressure status based on the comparative analysis of the core mechanical parameter set and the historical baseline parameter database includes:

[0036] The set of core mechanical parameters includes proximal systolic pressure, mean pressure, and pulse wave propagation velocity of the arteriovenous fistula.

[0037] From the historical baseline parameter database, retrieve the set of baseline core mechanical parameters of the target arteriovenous fistula in its historical healthy state, as well as typical parameter change patterns under different abnormal states;

[0038] Calculate the relative deviation of each item between the current set of core mechanical parameters and the baseline set of core mechanical parameters to form a parameter deviation vector;

[0039] The pattern matching degree is calculated by comparing the parameter deviation vector with each of the typical parameter change patterns to obtain a set of pattern matching scores.

[0040] The relative deviations and pattern matching scores are integrated into a multidimensional evaluation vector according to preset weighting coefficients, and the multidimensional evaluation vector is stored in a structured manner as the dynamic evaluation matrix of the fistula pressure status.

[0041] Further, the step of calculating the pattern matching degree between the parameter deviation vector and each of the typical parameter change patterns includes:

[0042] The parameter deviation vector is normalized to eliminate the influence of the difference in the dimensions of different mechanical parameters;

[0043] The cosine similarity between the normalized parameter deviation vector and the center vector of each of the typical parameter change patterns is calculated.

[0044] Simultaneously, the Mahalanobis distance between the normalized parameter deviation vector and the distribution of each of the typical parameter change patterns is calculated;

[0045] The cosine similarity and the Mahalanobis distance are weighted and fused to obtain the pattern matching score for the corresponding abnormal state.

[0046] Furthermore, the step of generating a functional status classification description and pressure abnormality risk warning for the target arteriovenous fistula based on the dynamic assessment matrix of the fistula pressure status includes:

[0047] The multidimensional evaluation vector in the dynamic evaluation matrix of the fistula pressure status is analyzed and input into a pre-trained classification decision tree model;

[0048] The classification decision tree model determines the multidimensional evaluation vector along the decision path according to the preset classification threshold, and finally falls into a specific leaf node;

[0049] Each leaf node is associated with a predefined functional status classification label and a set of pressure anomaly risk keywords. The functional status classification label and pressure anomaly risk keywords are output as the functional status classification description and pressure anomaly risk warning.

[0050] Furthermore, the pre-training process of the classification decision tree model includes:

[0051] Collect a large dataset containing historical arteriovenous fistula pressure data and their clinically validated end-state labels, including normal, early stenosis, significant stenosis, and thrombosis risk;

[0052] Extract the multidimensional evaluation vector corresponding to each data point from the dataset as a feature, and the final state label as the target;

[0053] The feature gain algorithm is used to select the optimal splitting feature and splitting point, and the decision tree is recursively constructed until the stopping condition is met;

[0054] The generated decision tree is pruned and optimized to prevent overfitting, and the optimized decision tree structure and its parameters are solidified into the pre-trained classification decision tree model.

[0055] Furthermore, the present invention also includes an arteriovenous fistula pressure measurement system based on a smart sensor, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the arteriovenous fistula pressure measurement method based on a smart sensor as described above.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] Smart sensors deployed on the body surface perform multimodal physiological noise filtering and employ signal separation technology to process mixed signals. This technology can extract the signal component that purely corresponds to the pressure change of the target arteriovenous fistula from the raw data containing various physiological interferences, based on the differences in signal source characteristics. This eliminates the contamination of the monitoring signal by irrelevant physiological activities and obtains a high-fidelity arteriovenous fistula pressure waveform. The key morphological features of the pressure waveform are clearly presented, and subtle features such as the systolic initiation point and dicrotic notch point within the waveform period can be stably and accurately marked by automatic algorithms, providing a reliable data foundation for subsequent analysis.

[0058] Based on a pre-defined hemodynamic physical model, a standard periodic pressure waveform template is generated as a boundary condition for inverse calculation. This process establishes a mathematical relationship between waveform morphology and vascular system physical parameters, thereby retrieving the core set of mechanical parameters within the arteriovenous fistula. These parameters quantitatively describe key states such as local blood flow impedance, vessel wall elasticity, and pressure gradient. This achieves a transformation from indirect waveform observation on the body surface to direct assessment of the internal mechanical state of the blood vessel, transforming the basis for functional assessment into objective and quantifiable physical indicators.

[0059] A dynamic assessment matrix is ​​established by comparing the core mechanical parameter set obtained through inverse engineering with historical baseline parameters. This matrix characterizes subtle changes in the hemodynamic status of the arteriovenous fistula by quantifying the multidimensional shift of current parameters relative to the individual baseline. Based on the pattern and degree of parameter shift, a graded description and risk warning are generated, evolving the status assessment from a single parameter threshold judgment to a comprehensive diagnosis based on multi-parameter collaborative analysis. This enhances the early identification capability of progressive changes in arteriovenous fistula function and the specificity of risk assessment. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the steps of the arteriovenous fistula pressure measurement method based on intelligent sensors described in this invention.

[0061] Figure 2 A flowchart for multimodal physiological noise filtering processing;

[0062] Figure 3 Flowchart for automatic feature point calibration processing;

[0063] Figure 4 Pattern matching score analysis of abnormal arteriovenous fistula states;

[0064] Figure 5 This is an analysis diagram of the characteristic points of a single-cycle pressure waveform in an arteriovenous fistula. Detailed Implementation

[0065] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] See Figure 1 The arteriovenous fistula pressure measurement method based on intelligent sensors is implemented through the following steps: Intelligent sensors are deployed at the corresponding position on the proximal end of the target arteriovenous fistula and subjected to multimodal physiological noise filtering to separate signal components and generate denoised fistula pressure waveform data. Automatic feature point calibration is performed on the denoised fistula pressure waveform data to identify the systolic initiation point, peak point, diastolic trough point, and dicrotic notch within the waveform cycle, constructing a single-cycle waveform feature point sequence. Based on the single-cycle waveform feature point sequence, multi-cycle waveform morphology alignment and averaging are performed to generate a standard periodic pressure waveform template characterizing the local hemodynamic state of the fistula. Combining a preset hemodynamic physical model with the standard periodic pressure waveform template, inverse waveform parameter calculation is performed to deduce the core mechanical parameter set reflecting the pressure changes within the fistula. A dynamic assessment matrix for the fistula pressure state is established based on a comparative analysis of the core mechanical parameter set and a historical baseline parameter database. Finally, a functional state classification description and pressure abnormality risk warning for the target arteriovenous fistula are generated based on the dynamic assessment matrix.

[0067] In one embodiment of the present invention, see [reference] Figure 2 In specific implementation, a smart sensor with pressure sensing capability is deployed at the proximal end of the target arteriovenous fistula on the body surface. The smart sensor is triggered to collect continuous raw pressure oscillation signals within a target time period to obtain an initial physiological pressure signal set. Multimodal physiological noise filtering is performed on the initial physiological pressure signal set to separate signal components originating from arterial pulsation, venous return, and tissue background noise, generating denoised fistula pressure waveform data. The multimodal physiological noise filtering process includes inputting the initial physiological pressure signal set into an adaptive noise reference signal generator. In specific implementation, the adaptive noise reference signal generator synchronously generates simulated arterial pulsation noise reference signals, venous return noise reference signals, and tissue motion noise reference signals based on signal spectrum characteristics and a preset noise source model. The simulated arterial pulsation noise reference signals, venous return noise reference signals, and tissue motion noise reference signals are then compared with the initial physiological pressure signal set through coherence analysis and adaptive weight adjustment to construct a composite noise estimation signal that highly matches the real noise. In some embodiments, the composite noise estimation signal is expressed by the following formula:

[0068]

[0069] in: This represents the composite noise estimation signal. This represents the simulated arterial pulsation noise reference signal. This represents the simulated venous return noise reference signal. This represents the simulated tissue motion noise reference signal. , and This represents the adaptive weighting coefficients dynamically adjusted through coherence analysis. In some embodiments, the composite noise estimation signal is subtracted from the initial physiological pressure signal set in the time domain to obtain the preliminary purified pressure signal. Optionally, the preliminary purified pressure signal is subjected to nonlinear wavelet threshold denoising to further suppress residual high-frequency random noise, ultimately outputting denoised arteriovenous fistula pressure waveform data. It is understood that the nonlinear wavelet threshold denoising process uses wavelet transform to decompose the preliminary purified pressure signal into multiple scale components. In a specific implementation, a preset threshold function is applied to each scale component to shrink the coefficients and filter out high-frequency random noise. Optionally, the processed scale components are synthesized into the final denoised arteriovenous fistula pressure waveform data through wavelet reconstruction. It is understood that the preset noise source model in the adaptive noise reference signal generator is constructed based on typical spectral templates of arterial pulsation, venous return, and tissue motion. In a specific implementation, coherence analysis calculates the cross-correlation function between the initial physiological pressure signal set and each simulated noise reference signal to evaluate signal similarity.

[0070] In one embodiment of the present invention, see [reference] Figure 3 In specific implementations, first-order and second-order differential calculations are performed on the denoised arteriovenous fistula pressure waveform data to obtain the waveform rate of change curve and acceleration curve. The zero-crossing point where the rate of change curve changes from negative to positive and the acceleration is positive is identified and calibrated as the systolic initiation point. In some embodiments, the first-order differential calculation uses the central difference method to process discrete denoised arteriovenous fistula pressure waveform data points to obtain the waveform rate of change curve. In specific implementations, the second-order differential calculation applies the central difference method again based on the waveform rate of change curve to obtain the acceleration curve. Optionally, the determination criterion for the systolic initiation point can be expressed as: for a point on the waveform rate of change curve... ,satisfy , And the corresponding acceleration Then the time point The systolic initiation point is designated as the peak point. This can be understood as finding the global maximum value of the pressure amplitude within a preset time window after the systolic initiation point and designating it as the peak point. In practice, the length of the preset time window is set based on historical data from typical cardiac cycles to avoid introducing interference from an excessively large search range.

[0071] In some embodiments, the first point in the waveform rate of change curve that satisfies a preset negative rate threshold after the peak point is identified as the diastolic trough point. It can be understood that the preset negative rate threshold... Given a constant less than zero, the search process starts from the time position corresponding to the peak point and checks the value of the waveform change rate curve point by point along the time axis. When the first value that satisfies this condition is encountered... When the time point is reached, the time point will be... The diastolic trough point is calibrated. In practice, the local extreme points of the acceleration curve between the peak point and the diastolic trough point are analyzed, and the inflection point characteristics of the waveform are combined to determine and calibrate the dicrotic notch. Optionally, this process is completed by locating the zero-crossing point of the acceleration curve where it changes from negative to positive within the corresponding interval, and verifying whether there is a significant inflection point in the pressure waveform before and after this point.

[0072] In specific implementation, a sequence of single-cycle waveform feature points corresponding to multiple consecutive cycles is collected, and the pressure waveform data within all cycles is normalized and aligned using the contraction start point of each cycle as the time reference point. It can be understood that time axis normalization and alignment linearly maps the original time axis of each cycle to a unified standardized time axis, aligning the contraction start points of all cycles to time zero. In some embodiments, the mean and variance of the amplitude of the pressure waveforms of all cycles at each time point after alignment are calculated to generate the average pressure waveform curve and the corresponding fluctuation range band. In specific implementation, for any point on the standardized time axis... Its average amplitude With variance By all in The fluctuation range is calculated based on the period of corresponding amplitude data at any given time. Optionally, the fluctuation range band is determined by... Definition. In specific implementation, the average pressure waveform curve undergoes smoothing interpolation to eliminate edge irregularities caused by minute differences in period length, forming a standard periodic pressure waveform template with a regular shape and statistical representativeness. It can be understood that the smoothing interpolation process uses cubic spline interpolation to resample the average pressure waveform curve to ensure a smooth and continuous waveform profile.

[0073] In the automatic feature point calibration process, the first and second derivative calculations are performed using the central difference method. This method processes discrete pressure data sequences acquired at fixed time intervals. The fixed sampling period of the smart sensor determines the time step between each adjacent data point. For internal data points in the sequence other than the first and last points, the central difference method takes the pressure value before the current point, the pressure value at the current point, and the pressure value after the current point for calculation. Specifically, the pressure value of the next data point is subtracted from the pressure value of the previous data point, and the difference is divided by twice the time step to obtain an approximate value of the pressure change rate at that internal point. This calculation process is continuously applied to every internal point in the entire pressure waveform data sequence to generate a waveform change rate curve that reflects the instantaneous rate of change. For the boundary points at the beginning and end of the pressure data sequence, the central difference method cannot be directly applied due to the lack of complete adjacent points. Therefore, a forward difference strategy is used at the beginning of the sequence, dividing the difference between the pressure value at the beginning point and the pressure value at the next point by the time step to calculate the rate of change. At the end of the sequence, a backward difference strategy is used, dividing the difference between the pressure value at the end point and the pressure value at the previous point by the time step to calculate the rate of change, thus ensuring that the entire curve has a complete definition. Based on the obtained waveform rate of change curve, the exact same central difference calculation process is applied again to perform discrete differentiation processing on the rate of change sequence, that is, calculating the difference between the preceding and following values ​​of each internal point in the rate of change sequence divided by twice the time step. Forward or backward difference is also used for the boundary points of the rate sequence. Finally, the acceleration curve describing the trend of the rate of change itself is output. These two curves provide direct mathematical criteria for subsequent scanning to identify the systolic start point, peak point, diastolic trough point, and dicrotic notch point.

[0074] In the smoothing interpolation step after multi-cycle waveform morphology alignment and averaging, cubic spline interpolation is used to process the average pressure waveform curve. This method constructs a globally smooth continuous curve based on each time point and its corresponding average pressure amplitude data. The implementation process first divides the entire time axis into multiple adjacent sub-intervals according to the known data points. Each sub-interval corresponds to two adjacent time points, and an independent cubic polynomial function is defined on each such sub-interval. Each polynomial contains four undetermined coefficients.

[0075] To ensure the final synthesized curve is smooth and passes through all original data points, these piecewise polynomials must satisfy strict continuity conditions, requiring adjacent polynomials to have equal function values, equal first-order derivative values, and equal second-order derivative values ​​at shared time points. In addition to the continuity of internal connection points, boundary conditions at both ends of the entire curve need to be specified to determine a unique solution. Natural boundary conditions are typically used, stipulating that the second-order derivative of the curve is zero at the start and end time points. Based on these conditions, a system of linear equations with all polynomial coefficients as unknowns can be constructed. The coefficient matrix of this system has a special tridiagonal structure. Solving this system of linear equations uses the chasing method, which consists of two main stages. The first stage is forward elimination, starting from the first equation and gradually eliminating elements on the lower diagonal of the coefficient matrix through algebraic operations, transforming the original system of equations into an upper triangular form. The second stage is backward substitution, solving in reverse order starting from the last equation, sequentially calculating all unknown polynomial coefficients. After obtaining all coefficients, for the interpolation calculation of any point on the time axis, first determine which sub-interval the point falls into, and then use the cubic polynomial and its coefficients corresponding to the sub-interval to calculate the pressure amplitude. By performing intensive sampling calculations in all sub-intervals, a seamless and smooth curve is finally generated. This curve eliminates the irregular fluctuations introduced near the endpoints of the average waveform due to the small difference in period length, forming a standard periodic pressure waveform template with regular shape and statistical representativeness.

[0076] In one embodiment of the present invention, a standard periodic pressure waveform template is input into a hemodynamic lumped parameter model that includes parameters of vascular elasticity, blood inertia, and peripheral resistance. A gradient descent optimization algorithm is used to iteratively adjust the unknown mechanical parameters in the hemodynamic lumped parameter model to minimize the mean square error between the theoretical pressure waveform output by the model simulation and the standard periodic pressure waveform template. In some embodiments, the hemodynamic lumped parameter model describes the pressure-flow relationship in the form of a set of differential equations, with the model input being a time series of the standard periodic pressure waveform template and the model output being the theoretical pressure waveform.

[0077] In practice, the gradient descent optimization algorithm reduces the loss function value by calculating the gradient of the loss function with respect to the unknown mechanical parameters and updating the parameter values ​​along the negative gradient direction. The loss function can be understood as the mean square error between the theoretical pressure waveform and the standard periodic pressure waveform template, and its calculation formula is:

[0078]

[0079] in: Represents the loss function. This represents a vector of unknown mechanical parameters, including vascular elasticity parameters, blood inertia parameters, and peripheral resistance parameters. This indicates that the hemodynamic lumped parameter model is in the parameter vector. Below the time point The theoretical pressure value output by the simulation. This indicates the standard periodic pressure waveform template at time point. Pressure value, This represents the total number of time points. Optionally, parameter updates are performed in a fixed step size along the negative direction of the loss function gradient. In specific implementations, the iterative adjustment process continues until the mean square error converges to within a preset threshold. At this point, the iteration terminates, and the corresponding vascular elasticity parameters, blood inertia parameters, and peripheral resistance parameters in the hemodynamic lumped parameter model are derived. In some embodiments, the preset threshold is set to a small positive number based on the measurement accuracy requirements. It can be understood that the pulse wave propagation velocity is calculated based on the theoretical relationship between the derived vascular elasticity parameters and pulse wave propagation. In specific implementations, the calculation of pulse wave propagation velocity is based on the relationship between vascular elasticity parameters, vascular geometry, and blood density. Optionally, the corresponding proximal fistula systolic pressure and mean pressure are directly extracted from the simulation waveform. The proximal fistula systolic pressure is taken as the maximum value of the simulation waveform within one cycle, and the mean pressure is taken as the integral average value of the simulation waveform within one cycle. In specific implementations, vascular elasticity parameters, blood inertia parameters, peripheral resistance parameters, pulse wave propagation velocity, proximal fistula systolic pressure, and mean pressure together constitute the core set of mechanical parameters.

[0080] In one embodiment of the present invention, the core mechanical parameter set includes the proximal systolic blood pressure, mean blood pressure, and pulse wave velocity of the arteriovenous fistula. The baseline core mechanical parameter set of the target arteriovenous fistula under historical healthy conditions and typical parameter change patterns under different abnormal conditions are retrieved from a historical baseline parameter database. In some embodiments, the historical baseline parameter database stores the core mechanical parameter set of each patient's arteriovenous fistula under functionally normal conditions in each test as the baseline core mechanical parameter set. In a specific implementation, the typical parameter change patterns under different abnormal conditions are derived by statistically analyzing a large number of clinically diagnosed abnormal case data. Each abnormal condition corresponds to a statistically representative parameter change pattern vector and its covariance matrix information. It can be understood that the parameter deviation vector is formed by calculating the item-by-item relative deviation between the current core mechanical parameter set and the baseline core mechanical parameter set. See Table 1.

[0081] Table 1: Typical Parameter Variation Patterns

[0082]

[0083] In specific implementations, a set of pattern matching scores is obtained by calculating the pattern matching degree between the parameter deviation vector and each typical parameter change pattern. In some embodiments, the pattern matching degree calculation includes normalizing the parameter deviation vector to eliminate the influence of differences in the dimensions of different mechanical parameters. It can be understood that the normalization process uses the z-score method, that is, subtracting the mean of the parameter in the baseline data from each element in the parameter deviation vector and then dividing by its standard deviation. Optionally, the cosine similarity is calculated between the normalized parameter deviation vector and the center vector of each typical parameter change pattern. Simultaneously, the Mahalanobis distance between the normalized parameter deviation vector and the distribution of each typical parameter change pattern is calculated. In specific implementations, the cosine similarity and Mahalanobis distance are weighted and fused to obtain the pattern matching score for the corresponding abnormal state. Pattern Matching Score The calculation formula is:

[0084]

[0085] in: This represents the pattern matching score for the k-th abnormal state. This represents the normalized parameter deviation vector. The center vector representing the typical parameter change pattern of the k-th abnormal state. This represents the covariance matrix of the typical parameter change pattern corresponding to the k-th abnormal state. Indicates the preset weighting coefficients and The dot product of vectors is represented by the symbol "·". and Let them represent the magnitudes of the vectors, respectively. This indicates the transpose operation. This represents the inverse operation of a matrix. It can be understood that the relative deviations and pattern matching scores are integrated into a multi-dimensional evaluation vector according to preset weighting coefficients, and this multi-dimensional evaluation vector is then structured and stored as a dynamic evaluation matrix of arteriovenous fistula pressure status.

[0086] See Figure 4 This is a pattern matching score analysis chart of abnormal arteriovenous fistula states, showing the pattern matching scores of 10 patients under three abnormal states. A higher score indicates a greater similarity between the current parameter deviation and the typical pattern of that abnormal state. Patient 10 has high matching scores under all three abnormal states, suggesting a complex condition and the possibility of multiple overlapping risks. Patient 6 has the lowest thrombosis risk matching score, indicating a significant difference between their current parameter deviation and the thrombosis risk pattern, and a relatively low risk. For patients with high matching scores in the early stages of stenosis, close follow-up monitoring and timely intervention are necessary to prevent disease progression. For patients with significant stenosis or high thrombosis risk matching scores, further examinations should be prioritized to clarify the diagnosis and develop a treatment plan.

[0087] In one embodiment of the present invention, in a specific implementation, the multidimensional evaluation vector in the dynamic evaluation matrix of arteriovenous fistula pressure status is parsed and input into a pre-trained classification decision tree model. The classification decision tree model determines the multidimensional evaluation vector along the decision path according to a preset grading threshold, ultimately placing it into a specific leaf node. In some embodiments, the multidimensional evaluation vector, as an input feature, is compared with the splitting rules stored in each internal node starting from the root node to determine whether to proceed along the path of the left or right child node. In a specific implementation, the splitting rule is typically in the form of "feature..." Is the value less than or equal to the threshold? ",in This represents the x-th dimension feature in the multidimensional evaluation vector. This represents the preset grading threshold. The judgment process can be understood as recursively proceeding until a leaf node that no longer contains child nodes is reached. Each leaf node is associated with a predefined functional state grading label and a set of pressure anomaly risk keywords. The functional state grading label and pressure anomaly risk keywords are output as a description of the functional state grading and a warning of pressure anomaly risks. Optionally, the functional state grading label includes "normal," "early stenosis," "significant stenosis," and "thrombosis risk," while the pressure anomaly risk keywords correspond to specific descriptions such as "abnormal flow velocity near the anastomosis," "high probability of luminal stenosis," and "increased risk of thrombosis."

[0088] The pre-training process of the classification decision tree model involves collecting a large dataset containing historical arteriovenous fistula pressure data and their clinically validated final state labels, including normal, early stenosis, significant stenosis, and thrombosis risk. In practice, a multidimensional evaluation vector corresponding to each case is extracted from the dataset as a feature, and the final state label serves as the target for model learning. In some embodiments, a feature gain algorithm is used to recursively construct the decision tree by selecting the optimal splitting features and split points until a stopping condition is met. The feature gain algorithm evaluates the discriminative power by calculating the information gain brought by each potential split point of a feature. (Information gain) The calculation is based on a subset of data. The specific formula for the entropy change before and after the split is:

[0089]

[0090] in: Indicated by features At the threshold Split data subset The information gain obtained Representing a subset of data entropy, and Let each represent a subset of data that satisfies and does not satisfy the conditions after the split. , and These represent the number of samples in the corresponding data subsets. In practice, the feature and split point that generate the maximum information gain are selected as the optimal splitting rule for the current node. This can be understood as recursively constructing the decision tree until a stopping condition is met, which includes the number of samples in a node falling below a preset minimum or all samples belonging to the same final state label. The generated decision tree is pruned to prevent overfitting, and the optimized decision tree structure and its parameters are solidified into a pre-trained classification decision tree model. Optionally, the pruning optimization uses a cost-complexity pruning method, removing redundant branches by balancing the complexity of the decision tree with the classification accuracy on the validation set.

[0091] See Figure 5 This is a single-cycle pressure waveform feature point analysis diagram of an arteriovenous fistula (AVF), showing the pressure change waveform of the AVF within a complete cardiac cycle and identifying four key physiological feature points. It is a core visualization tool for assessing the hemodynamic status of the fistula. From the overall waveform morphology, the ascending limb during systole is steep, and the descending limb during diastole is smooth, consistent with the hemodynamic characteristics of a normal AVF. These four feature points are the core of constructing the single-cycle waveform feature point sequence. Subsequent multi-cycle alignment and averaging can generate a standard cycle pressure waveform template. Based on this template, combined with a hemodynamic physical model, core mechanical parameters such as proximal systolic pressure, mean pressure, and pulse wave velocity can be inversely solved. An abnormally high peak pressure may indicate fistula stenosis; an excessively low diastolic trough pressure or abnormal waveform morphology may indicate a risk of thrombosis.

[0092] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for measuring arteriovenous fistula pressure based on a smart sensor, characterized in that, The method includes: A smart sensor is deployed at the corresponding position on the proximal end of the target arteriovenous fistula, and multimodal physiological noise filtering is performed to separate the signal components and generate noise-reduced fistula pressure waveform data. The noise-reduced arteriovenous fistula pressure waveform data is subjected to automatic feature point calibration processing to identify the systolic initiation point, peak point, diastolic trough point and dicrotic notch point within the waveform cycle, and to construct a single-cycle waveform feature point sequence. Based on the single-cycle waveform feature point sequence, multi-cycle waveform morphology alignment and averaging are performed to generate a standard periodic pressure waveform template that characterizes the local hemodynamic state of the arteriovenous fistula. By combining the preset hemodynamic physical model with the standard periodic pressure waveform template, the waveform parameters are solved in reverse to obtain the set of core mechanical parameters reflecting the pressure changes inside the fistula. Based on the comparative analysis of the core mechanical parameter set and the historical baseline parameter database, a dynamic evaluation matrix for arteriovenous fistula pressure status is established; Based on the dynamic assessment matrix of arteriovenous fistula pressure status, a functional status classification description and pressure abnormality risk warning are generated for the target arteriovenous fistula.

2. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 1, characterized in that, The smart sensor is deployed at the proximal end of the target arteriovenous fistula on the body surface, and performs multimodal physiological noise filtering to separate the signal components and generate denoised fistula pressure waveform data, including: Deploy a smart sensor with pressure sensing capability at the corresponding position on the proximal end of the target arteriovenous fistula, trigger the smart sensor to collect continuous raw pressure oscillation signals within the target time period, and obtain an initial set of physiological pressure signals; Multimodal physiological noise filtering is performed on the initial physiological pressure signal set to separate the signal components originating from arterial pulsation, venous return and tissue background noise, and generate denoised arteriovenous fistula pressure waveform data; The multimodal physiological noise filtering process performed on the initial physiological stress signal set includes: The initial set of physiological pressure signals is input into an adaptive noise reference signal generator, which generates simulated arterial pulsation noise reference signals, venous return noise reference signals, and tissue motion noise reference signals simultaneously based on signal spectrum characteristics and a preset noise source model. The simulated arterial pulsation noise reference signal, venous return noise reference signal, and tissue motion noise reference signal are respectively subjected to coherence analysis and adaptive weight adjustment with the initial physiological pressure signal set to construct a composite noise estimation signal that highly matches the real noise. The composite noise estimation signal is subtracted from the initial physiological pressure signal set in the time domain to obtain the preliminary purification pressure signal; The preliminary purified pressure signal is subjected to nonlinear wavelet threshold denoising to further suppress residual high-frequency random noise, and finally the denoised arteriovenous fistula pressure waveform data is output.

3. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 1, characterized in that, The automatic feature point calibration processing of the denoised arteriovenous fistula pressure waveform data includes: Perform first-order and second-order differential calculations on the denoised arteriovenous fistula pressure waveform data to obtain the waveform change rate curve and acceleration curve; Scan the waveform rate of change curve, identify the zero-crossing point where the rate changes from negative to positive and the acceleration is positive, and mark it as the starting point of the contraction period; Within a preset time window following the start point of the contraction period, the global maximum value of the pressure amplitude is found and marked as the peak point; After the peak point, search for the first point in the waveform rate of change curve that satisfies the preset negative rate threshold, and mark it as the diastolic valley point; Between the peak point and the diastolic trough point, the local extreme points of the acceleration curve are analyzed, and the inflection point characteristics of the waveform are combined to determine and mark the dilatation notch.

4. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 1, characterized in that, The step of performing multi-cycle waveform morphology alignment and averaging based on the single-cycle waveform feature point sequence includes: Collect the sequence of single-cycle waveform feature points corresponding to multiple consecutive cycles, and use the starting point of the contraction period of each cycle as the time reference point to normalize and align the pressure waveform data in all cycles. Calculate the mean and variance of the amplitude of all periodic pressure waveforms at each time point after alignment, and generate the average pressure waveform curve and the corresponding fluctuation range band; The average pressure waveform curve is smoothed by interpolation to eliminate edge irregularities caused by slight differences in period length, forming a standard periodic pressure waveform template with regular shape and statistical representativeness.

5. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 1, characterized in that, The process of combining a preset hemodynamic physical model with the standard periodic pressure waveform template to perform inverse waveform parameter calculation includes: The standard periodic pressure waveform template is input into a hemodynamic lumped parameter model that includes parameters of vascular elasticity, blood inertia, and peripheral resistance. The gradient descent optimization algorithm is used to iteratively adjust the unknown mechanical parameters in the hemodynamic lumped parameter model so that the mean square error between the theoretical pressure waveform output by the model simulation and the standard periodic pressure waveform template is minimized. When the mean square error converges to within a preset threshold, the iteration is terminated, and the corresponding vascular elasticity parameters, blood inertia parameters, and peripheral resistance parameters in the hemodynamic lumped parameter model at this time are exported. Based on the theoretical relationship between the derived vascular elasticity parameters and pulse wave propagation, the pulse wave propagation velocity is calculated, and the corresponding proximal systolic pressure and mean pressure of the arteriovenous fistula are directly extracted from the simulation waveform, which together constitute the core set of mechanical parameters.

6. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 1, characterized in that, The dynamic assessment matrix of arteriovenous fistula pressure status is established based on the comparative analysis of the core mechanical parameter set and the historical baseline parameter database, including: The set of core mechanical parameters includes proximal systolic pressure, mean pressure, and pulse wave propagation velocity of the arteriovenous fistula. From the historical baseline parameter database, retrieve the set of baseline core mechanical parameters of the target arteriovenous fistula in its historical healthy state, as well as typical parameter change patterns under different abnormal states; Calculate the relative deviation of each item between the current set of core mechanical parameters and the baseline set of core mechanical parameters to form a parameter deviation vector; The pattern matching degree is calculated by comparing the parameter deviation vector with each of the typical parameter change patterns to obtain a set of pattern matching scores. The relative deviations and pattern matching scores are integrated into a multidimensional evaluation vector according to preset weighting coefficients, and the multidimensional evaluation vector is stored in a structured manner as the dynamic evaluation matrix of the fistula pressure status.

7. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 6, characterized in that, The step of calculating the pattern matching degree between the parameter deviation vector and each of the typical parameter change patterns includes: The parameter deviation vector is normalized to eliminate the influence of the difference in the dimensions of different mechanical parameters; The cosine similarity between the normalized parameter deviation vector and the center vector of each of the typical parameter change patterns is calculated. Simultaneously, the Mahalanobis distance between the normalized parameter deviation vector and the distribution of each of the typical parameter change patterns is calculated; The cosine similarity and the Mahalanobis distance are weighted and fused to obtain the pattern matching score for the corresponding abnormal state.

8. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 1, characterized in that, The process of generating a functional status classification description and pressure abnormality risk warning for the target arteriovenous fistula based on the dynamic assessment matrix of the fistula pressure status includes: The multidimensional evaluation vector in the dynamic evaluation matrix of the fistula pressure status is analyzed and input into a pre-trained classification decision tree model; The classification decision tree model determines the multidimensional evaluation vector along the decision path according to the preset classification threshold, and finally falls into a specific leaf node; Each leaf node is associated with a predefined functional status classification label and a set of pressure anomaly risk keywords. The functional status classification label and pressure anomaly risk keywords are output as the functional status classification description and pressure anomaly risk warning.

9. The method for measuring arteriovenous fistula pressure based on a smart sensor according to claim 8, characterized in that, The pre-training process of the classification decision tree model includes: Collect a large dataset containing historical arteriovenous fistula pressure data and their clinically validated end-state labels, including normal, early stenosis, significant stenosis, and thrombosis risk; Extract the multidimensional evaluation vector corresponding to each data point from the dataset as a feature, and the final state label as the target; The feature gain algorithm is used to select the optimal splitting feature and splitting point, and the decision tree is recursively constructed until the stopping condition is met; The generated decision tree is pruned and optimized to prevent overfitting, and the optimized decision tree structure and its parameters are solidified into the pre-trained classification decision tree model.

10. A pressure measurement system for arteriovenous fistulas based on intelligent sensors, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the arteriovenous fistula pressure measurement method based on a smart sensor as described in any one of claims 1 to 9.

Images