Vascular hemodynamic abnormality monitoring system and pulse analysis method thereof

Abstract

The present application relates to the technical field of medical monitoring, in particular to a vascular hemodynamic abnormality monitoring system and a pulse wave analysis method thereof, the core of the present application lies in that a flexible piezoelectric sensor array is used to capture human vascular pulse wave signals, the signals are transmitted to a wearable monitoring device for preliminary preprocessing, and then transmitted to a data processing host through a wireless manner, a pulse wave analysis module in the host adopts a multi-level analysis method, including multi-scale signal decomposition and reconstruction, nonlinear feature extraction and pattern recognition, multi-parameter adaptive judgment standard, deep learning feature automatic extraction and multi-site information fusion decision, ensuring the accuracy and comprehensiveness of the analysis, finally, the processing result is transmitted to an application platform to generate a vascular health report, realizing non-invasive continuous monitoring, this innovation effectively fills the gap of the prior art in continuous monitoring, and has important significance for real-time monitoring of vascular health status and prevention of vascular diseases.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical monitoring, in particular to a blood vessel hemodynamic abnormality monitoring system and a pulse wave analysis method thereof, which are especially suitable for early monitoring and early warning of blood vessel elasticity abnormality and deep vein thrombosis. BACKGROUND

[0002] Blood vessel hemodynamic abnormality, especially deep vein thrombosis, is a common and dangerous disease in clinic. At present, the methods for detecting blood vessel hemodynamic abnormality in clinic mainly include ultrasonic Doppler, venography and D-dimer detection. These detection methods can reflect the blood vessel state to a certain extent, but all have obvious shortcomings.

[0003] Ultrasonic Doppler detection needs to be operated by a professional physician, and can only provide single-point and short-time detection results, and cannot realize continuous monitoring; venography is an invasive examination and is not suitable for frequent use; D-dimer detection is a post-confirmation method and does not have a warning function. At the same time, most of the existing detection means depend on large medical equipment, which is expensive and cannot meet the daily monitoring needs.

[0004] In addition, most of the existing wearable health monitoring devices focus on the monitoring of basic physiological parameters such as heart rate and blood pressure, and lack professional assessment ability for blood vessel health status. Especially for early warning of potential fatal diseases such as deep vein thrombosis, the early warning ability is insufficient, and it is difficult to provide effective risk warning before the occurrence of clinical symptoms.

[0005] The existing technology also has limitations in signal processing. The traditional pulse wave analysis method mainly uses simple time domain or frequency domain analysis, which cannot effectively capture the subtle characteristics of blood vessel state changes; at the same time, it lacks individualized evaluation standards, and it is difficult to adapt to the physiological differences between different individuals, resulting in a high false alarm rate.

[0006] Therefore, it is of great clinical significance and application value to develop a system that can non-invasively and continuously monitor the blood vessel hemodynamic state and has an early abnormality warning function. SUMMARY

[0007] In view of the problems in the prior art, the purpose of the present application is to provide a blood vessel hemodynamic abnormality monitoring system and a pulse wave analysis method thereof, which can realize non-invasive continuous monitoring of the blood vessel health status, especially early warning of deep vein thrombosis, so as to provide sufficient time window for clinical intervention.

[0008] The present application provides a blood vessel hemodynamic abnormality monitoring system and a pulse wave analysis method thereof, which comprises:

[0009] a flexible piezoelectric sensor array for collecting human blood vessel pulse wave signals;

[0010] a wearable monitoring device electrically connected with the flexible piezoelectric sensor array for receiving the pulse wave signal and performing preliminary preprocessing;

[0011] a data processing host connected with the wearable monitoring device through wireless communication for receiving and processing the pulse wave signal;

[0012] a pulse wave analysis module arranged in the data processing host for performing multi-level analysis on the pulse wave signal, the pulse wave analysis module comprising a multi-scale signal decomposition and reconstruction unit, a nonlinear feature extraction and pattern recognition unit, a multi-parameter adaptive judgment standard unit, a deep learning feature automatic extraction unit, and a multi-site information fusion decision unit connected in sequence; and

[0013] an application platform connected with the data processing host through a network for receiving the processing result and generating a blood vessel health report.

[0014] Preferably, the multi-scale signal decomposition and reconstruction unit comprises:

[0015] a signal decomposition subunit for receiving the pulse wave signal and decomposing the pulse wave signal into a plurality of sets of coefficients at multiple scales through a plurality of wavelet basis functions;

[0016] a denoising processing subunit for performing soft threshold shrinkage denoising on the sets of coefficients; and

[0017] a reconstruction subunit for dynamically adjusting the weights of the coefficients at different scales according to the signal energy distribution and generating reconstructed multi-scale feature coefficients.

[0018] Preferably, the nonlinear feature extraction and pattern recognition unit comprises:

[0019] an autocorrelation calculation subunit for calculating pulse wave autocorrelation sequences at different time windows;

[0020] a nonlinear fitting subunit for performing nonlinear fitting on the autocorrelation sequences using a Gaussian function;

[0021] a feature extraction subunit for extracting feature parameters reflecting blood vessel elasticity, blood flow resistance, cycle stability, amplitude variability, and waveform complexity from the autocorrelation sequences and fitting results; and

[0022] a feature optimization subunit for screening the most discriminative features from the initial feature set according to sample discriminability and stability scores.

[0023] Preferably, the multi-parameter adaptive judgment standard unit comprises:

[0024] a static threshold subunit configured to set a fixed threshold based on physiological common sense;

[0025] a group threshold subunit configured to set a group threshold according to factors such as age and gender;

[0026] a personalized threshold subunit configured to learn the normal state range of a user and establish a personalized model;

[0027] a state recognition subunit configured to distinguish between resting, light activity and strenuous exercise states; and

[0028] an abnormality classification subunit configured to classify abnormalities into five categories of vascular elasticity abnormality, blood flow resistance abnormality, vascular compliance abnormality, blood flow velocity abnormality and blood flow pattern abnormality.

[0029] Preferably, the deep learning feature automatic extraction unit comprises:

[0030] an input layer configured to receive a 32-channel pulse wave signal;

[0031] a feature extraction layer comprising a plurality of convolutional layers and pooling layers, configured to extract features from different perspectives;

[0032] an attention mechanism module configured to automatically identify and enhance the features of important sensor channels;

[0033] a residual connection module configured to solve the gradient vanishing problem in deep networks; and

[0034] an output layer configured to generate a vascular state classification result.

[0035] Preferably, the multi-site information fusion decision unit comprises:

[0036] a weight allocation subunit configured to allocate weights to the monitoring results of different sites;

[0037] a time series analysis subunit configured to analyze short-term trend changes, medium-term periodic changes and long-term evolution trends;

[0038] a risk assessment subunit configured to determine a risk level based on the comprehensive information; and

[0039] a report generation subunit configured to integrate information and generate a comprehensive assessment report containing vascular elasticity, blood flow resistance and thrombus risk.

[0040] Preferably, the flexible piezoelectric sensor array comprises 32 piezoelectric sensors made of medical-grade polyvinylidene fluoride piezoelectric film, arranged in a 4x8 matrix structure, with a single sensor size of 3mmx3mm and a thickness of 0.2mm.

[0041] Preferably, the wearable monitoring device comprises:

[0042] a data acquisition unit for multi-channel synchronous acquisition control;

[0043] a signal preprocessing unit for band-pass filtering, notch filtering, and outlier detection of the acquired signals;

[0044] a Bluetooth communication unit for transmitting the processed data to the data processing host, and a power management unit for providing stable power supply and implementing low-power consumption management.

[0045] Preferably, the application platform comprises:

[0046] a data receiving module for receiving the processing results sent by the data processing host;

[0047] a data storage module for storing historical monitoring data;

[0048] an analysis and display module for generating visual charts and trend analysis;

[0049] an early warning and reminder module for issuing an alarm when an anomaly is detected; and

[0050] a user interaction module for receiving user feedback and configuring system parameters.

[0051] A pulse wave analysis method applied to the blood vessel hemodynamic anomaly monitoring system, comprising the following steps:

[0052] acquiring a pulse wave signal and sending the acquired pulse wave signal to a data processing host;

[0053] performing multi-scale decomposition on the pulse wave signal to obtain a feature coefficient set at different scales;

[0054] extracting nonlinear features from the feature coefficient set, including vessel elasticity, blood flow resistance, cycle stability, amplitude variability, and waveform complexity features;

[0055] establishing a multi-level adaptive judgment standard according to individual differences and physiological states, and evaluating the nonlinear features;

[0056] automatically extracting deep features from the original signal using a deep learning network and performing vessel state classification;

[0057] fusing the evaluation results of multiple monitoring sites, performing time series pattern analysis and comprehensive risk assessment; and

[0058] generating a blood vessel health report, including abnormal type, risk level, and health recommendations.

[0059] The present application collects pulse wave signals through a flexible piezoelectric sensor array, combines multi-level signal processing and analysis technology, and realizes comprehensive evaluation of blood vessel elasticity, blood flow resistance and thrombus risk. The system adopts a bottom-to-top signal processing and analysis strategy, and through innovative technologies such as multi-scale signal decomposition, nonlinear feature extraction, adaptive threshold judgment, deep learning feature extraction and multi-site information fusion, the sensitivity and specificity of blood vessel anomaly detection are greatly improved.

[0060] The beneficial effects of the present application mainly include:

[0061] 1. Non-invasive continuous monitoring of blood vessel health status is realized, filling the gap in continuous monitoring of the prior art;

[0062] 2. Through multi-level signal processing and analysis technology, the detection sensitivity is significantly improved, and the thrombus risk rising trend can be found 48-72 hours before the clinical symptoms appear;

[0063] 3. The individual adaptive threshold judgment standard is adopted, which effectively solves the misjudgment problem caused by individual differences, and reduces the false positive rate from 25% of the traditional method to 8%;

[0064] 4. The multi-site monitoring results are innovatively combined to improve the overall reliability of the detection;

[0065] 5. The system has high integration degree, simple operation, is suitable for daily wearing and long-term monitoring, and improves the user compliance;

[0066] 6. The cost is significantly lower than that of traditional hospital examination equipment, and has wide popularization and application value. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 It is the overall structure schematic diagram of the blood vessel hemodynamic abnormality monitoring system of the present application;

[0068] Figure 2 It is the structure schematic diagram of the flexible piezoelectric sensor array of the present application;

[0069] Figure 3 It is the structure schematic diagram of the wearable monitoring device of the present application;

[0070] Figure 4 It is the structure schematic diagram of the pulse wave analysis module of the present application;

[0071] Figure 5 It is the signal processing flowchart of the multi-scale signal decomposition and reconstruction unit of the present application;

[0072] Figure 6 It is the function block diagram of the nonlinear feature extraction and pattern recognition unit of the present application;

[0073] Figure 7 This is a flowchart of the judgment process of the multi-parameter adaptive judgment standard unit of the present invention;

[0074] Figure 8 This is a network structure diagram of the deep learning feature extraction unit of the present invention;

[0075] Figure 9 This is a flowchart of the multi-part information fusion decision-making unit of the present invention;

[0076] Figure 10 This is a flowchart of the pulse wave analysis method of the present invention. Detailed Implementation

[0077] Please refer to the attached document. Figures 1-10 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are for illustrative purposes only and should not be construed as limiting the invention.

[0078] Reference Figure 1 The present invention provides a vascular hemodynamic abnormality monitoring system, including a flexible piezoelectric sensor array 1, a wearable monitoring device 2, a data processing host 3, a pulse wave analysis module 4, and an application platform 5.

[0079] A flexible piezoelectric sensor array 1 is used to acquire pulse wave signals from human blood vessels. In one embodiment of the invention, the flexible piezoelectric sensor array 1 senses changes in the pulsating pressure of subcutaneous blood vessels by contacting the surface of human skin, converting mechanical energy into electrical signals to achieve non-invasive acquisition of pulse wave signals. This array is made of flexible materials, suitable for prolonged attachment to the skin surface, improving user comfort. Preferably, when the pulsating pressure of the blood vessels is conducted through the skin to the sensor surface, the voltage signal generated by the flexible piezoelectric sensor is proportional to the change in intravascular pressure, thereby accurately capturing vascular dynamic characteristics.

[0080] The wearable monitoring device 2 is electrically connected to the flexible piezoelectric sensor array 1 to receive pulse wave signals and perform preliminary preprocessing. Preferably, the wearable monitoring device 2 adopts a watchband design, which can be worn on areas rich in blood vessels such as the wrist and ankle for convenient daily use. The wearable monitoring device 2 has a built-in microprocessor and wireless communication module, which can preprocess the collected signals and transmit them to the data processing host 3. In practical applications, for example, for monitoring deep vein thrombosis in long-term bedridden patients, the wearable monitoring device 2 can be worn on the lower limbs to continuously monitor venous return; while for patients who need to monitor carotid artery blood flow, the device can be worn around the neck to achieve real-time monitoring of cerebral blood supply.

[0081] The data processing host 3 is connected with the wearable monitoring device 2 through wireless communication, for receiving and processing the pulse wave signals. In an embodiment of the present application, the data processing host 3 can be a smart phone, a tablet computer or a dedicated processing device, responsible for running complex signal processing and analysis algorithms. Preferably, the data processing host 3 adopts a multi-thread parallel processing architecture, processing monitoring data from different parts in real time, ensuring the response speed when a large amount of data is transmitted at the same time. In addition, the data processing host 3 also has a cache function, which can temporarily store data when the wireless connection is interrupted, and automatically synchronize after the connection is restored, ensuring data integrity.

[0082] The pulse wave analysis module 4 is arranged in the data processing host 3, which is the core part of the system, for multi-level analysis of the pulse wave signals. The pulse wave analysis module 4 includes a multi-scale signal decomposition and reconstruction unit 41, a nonlinear feature extraction and pattern recognition unit 42, a multi-parameter adaptive judgment standard unit 43, a deep learning feature automatic extraction unit 44 and a multi-site information fusion decision unit 45 connected in sequence. These five functional units form a complete signal processing chain, gradually extracting valuable features from the original pulse wave signals, and finally realizing accurate evaluation of the vascular state. For example, in the early stage of detecting deep vein thrombosis of lower limbs, the system can identify the changes in pulse wave characteristics caused by increased venous return resistance and give an early warning 48-72 hours before the appearance of clinical symptoms, providing a valuable time window for clinical intervention.

[0083] The application platform 5 is connected with the data processing host 3 through a network, for receiving the processing results and generating vascular health reports. The application platform 5 displays the monitoring results to the user through a visual interface and gives an early warning prompt when an abnormality is detected. In addition, the application platform 5 can also store data in the cloud, supporting remote medical consultation and long-term health management. In clinical practice, medical staff can remotely view the vascular health status of patients through the application platform 5 and timely adjust the treatment plan; for patients recovering at home, the system can provide continuous monitoring and health guidance, reducing the frequency of hospital visits.

[0084] Reference Figure 2The flexible piezoelectric sensor array 1 of the present application includes 32 piezoelectric sensors arranged in a 4x8 matrix structure. Each piezoelectric sensor is made of medical-grade polyvinylidene fluoride piezoelectric film, with a single sensor size of 3mmx3mm and a thickness of 0.2mm. The spacing between the piezoelectric sensors is 2mm, ensuring that the array can cover a large enough area while maintaining overall flexibility. The design sensitivity of the flexible piezoelectric sensor array 1 is greater than 5mV / g, which can accurately capture weak pulse wave signals. This sensitivity design is based on clinical research, which shows that the pressure change generated by normal adult arterial pulsation is about 0.2-0.8g, while the pressure change caused by early vascular abnormalities is only 0.05-0.1g. Therefore, higher sensitivity is required to capture these small changes.

[0085] Preferably, the flexible piezoelectric sensor array 1 is made of flexible printed circuit board technology, which has good ductility and adhesion, and is suitable for long-term attachment to the skin surface without causing discomfort. At the same time, the sensor surface is coated with a biocompatible material to avoid skin allergic reactions caused by long-term contact. In clinical applications, for example, for obese patients, traditional rigid sensors often have difficulty adhering to the skin surface, resulting in a decrease in signal quality. The flexible design of the present application can well adapt to the needs of patients of different body types, ensuring signal acquisition quality.

[0086] Reference Figure 3 The wearable monitoring device 2 of the present application includes a data acquisition unit 21, a signal preprocessing unit 22, a Bluetooth communication unit 23, and a power management unit 24.

[0087] The data acquisition unit 21 is used for multi-channel synchronous acquisition control and can simultaneously read the signals of 32 piezoelectric sensors with a sampling frequency of 200Hz and a sampling accuracy of 12 bits. The data acquisition unit 21 uses a high-precision analog-to-digital converter to achieve accurate digitization of the signal. To ensure acquisition accuracy, the crosstalk suppression between channels of the data acquisition unit 21 is greater than 60dB, effectively avoiding signal interference between channels. In actual monitoring processes, for example, when a patient needs to monitor multiple sites (such as both lower extremities) simultaneously, the high-precision multi-channel synchronous acquisition capability ensures the time alignment of data from different parts, providing a basis for subsequent information fusion analysis.

[0088] The signal preprocessing unit 22 is used for band-pass filtering, notch filtering and outlier detection of the collected signal. Preferably, the passband range of the band-pass filtering is 0.5-40 Hz, which can effectively retain the effective components of the pulse wave signal while filtering out baseline drift and high-frequency noise; the center frequency of the notch filtering is 50 Hz (or 60 Hz), which is used to suppress power frequency interference; the outlier detection is based on the moving median algorithm, which can identify and correct abnormal data points in the collection process. The selection of the passband range is based on the spectral characteristics of the human pulse wave signal: the fundamental frequency (heart rate) is usually in the range of 0.8-2 Hz, the harmonic component can reach 10-15 Hz, and the subtle changes related to vascular elasticity and compliance are reflected in the frequency band up to 40 Hz. By selecting 0.5 Hz as the low-frequency cutoff point, baseline drift caused by respiration and body position changes can be filtered out; by selecting 40 Hz as the high-frequency cutoff point, the high-frequency components related to vascular status can be retained while filtering out high-frequency noise such as muscle electrical interference.

[0089] The Bluetooth communication unit 23 adopts Bluetooth 5.0 low-power technology, which is used to transmit the processed data to the data processing host 3. The transmission rate is greater than 2 Mbps, which meets the real-time data transmission requirements. The data packet size is 256 bytes per packet, and the CRC16 check algorithm is used to ensure the accuracy of transmission. In daily application scenarios, for example, when the user is active at home, the Bluetooth communication unit 23 ensures stable connection with the data processing host such as a smart phone; when the transmission quality decreases, the system automatically adjusts the sampling rate and compression ratio to prioritize real-time transmission of critical data.

[0090] The power management unit 24 adopts a rechargeable lithium polymer battery with a capacity of 200 mAh, which is used to provide stable power supply and realize low-power management. In the continuous monitoring mode, the device can work for more than 24 hours; in the intermittent monitoring mode, the working time can be extended to more than 72 hours. The power management unit 24 also supports wireless charging and USB Type-C charging, which is convenient for daily use. For chronic disease patients who need long-term monitoring, the power management unit 24 adopts an intelligent power consumption scheduling strategy, which dynamically adjusts the sampling frequency and processing depth according to the user's activity state and signal complexity, while ensuring the monitoring quality and maximizing the battery life.

[0091] Referring to Figure 4 The pulse wave analysis module 4 of the present application is the core of the system, which includes five functional units. The five units form a layer-by-layer progressive signal processing chain, the output of each unit is used as the input of the next unit, forming a complete information flow, and finally realizing accurate evaluation of the vascular status.

[0092] Referring to Figure 5 The multi-scale signal decomposition and reconstruction unit 41 includes a signal decomposition sub-unit 411, a denoising processing sub-unit 412 and a reconstruction sub-unit 413.

[0093] The signal decomposition subunit 411 is configured to receive the pulse wave signal and decompose the pulse wave signal into a plurality of sets of coefficients at different scales by a plurality of wavelet basis functions. In one embodiment of the present application, the signal decomposition subunit 411 constructs a library of wavelet basis functions with 8 different characteristics, each of which is optimized for different frequency components of the pulse wave signal. The pulse wave signal S(t) can be represented by a wavelet transform as follows:

[0094] C j,k =∫S(t)·ψ j，k (t)dt,

[0095] where C j，k is the wavelet coefficient, i represents the scale parameter, k represents the translation parameter, and ψ j，k (t) is the wavelet basis function, which is defined as:

[0096] ψ j，k (t) = 2 -j / 2 ·ψ(2 -j t-k),

[0097] where ψ(t) is the mother wavelet function; 2 -j / 2 is the normalization factor to ensure energy conservation at different scales; 2 -j is the scale factor to control the stretching of the wavelet; and k is the translation factor to control the translation of the wavelet. The present application preferably uses Daubechies wavelet (db4) as the mother wavelet, as it has good localization characteristics and energy concentration, and is particularly suitable for capturing transient changes in the pulse wave signal. In actual applications of vascular monitoring, such as detecting the early stage of deep vein thrombosis, this multi-scale decomposition can simultaneously capture changes in blood flow velocity (low frequency components) and changes in vascular wall stiffness (high frequency components), providing comprehensive evaluation of the vascular status.

[0098] The signal decomposition subunit 411 performs 8-layer progressive decomposition on the original pulse wave signal, with the first layer capturing features in the 20-40 Hz frequency band and the eighth layer capturing features in the 0.5-1 Hz frequency band. This multi-scale decomposition strategy can comprehensively capture the features of the pulse wave signal in different frequency bands, providing a rich source of information for subsequent analysis. In clinical applications, different frequency band features are related to different pathological states: the 0.5-1 Hz frequency band mainly reflects the basic frequency of the heart rate, and abnormalities may indicate arrhythmia; the 1-5 Hz frequency band reflects vascular compliance, and abnormalities may indicate arteriosclerosis; the 5-15 Hz frequency band reflects vascular resistance, and abnormalities may indicate thrombosis; and the 15-40 Hz frequency band may contain the characteristics of the micro-vibration of the vascular wall, and abnormalities may indicate vascular intima damage. Through full-band analysis, the system can provide a more comprehensive evaluation of vascular health.

[0099] The de-noising processing sub-unit 412 is configured to perform soft threshold shrinkage de-noising on the coefficient set. The soft threshold shrinkage function is defined as:

[0100]

[0101] wherein δ λ (x) is the coefficient value after soft threshold shrinkage; λ is the threshold value, which determines the de-noising strength; x is the original wavelet coefficient; sgn(x) is the sign function, which takes the value of 1 when x>0, -1 when x<0, and 0 when x=0; |x| represents the absolute value of x. The threshold value λ is usually set as σ is the noise standard deviation, and N is the signal length. In practical applications, σ is usually estimated by the median absolute deviation of the minimum scale (highest frequency) sub-band of the wavelet coefficient: σ = median(|C 1，k |) / 0.6745, wherein 0.6745 is the standardization factor of the median absolute deviation of the standard normal distribution. The de-noising processing sub-unit 412 applies different threshold values to the coefficients of each scale, i.e., a larger threshold value is applied to the smaller scale (high frequency), and a smaller threshold value is applied to the larger scale (low frequency), which effectively suppresses the noise while preserving the key features of the signal. In the practical application of blood vessel monitoring, de-noising is particularly important because factors such as user activity, respiration, and environmental vibration can introduce noise, affecting the detection accuracy. For example, in elderly patients, the decrease in skin elasticity leads to an increase in sensor signal noise, and at this time, soft threshold de-noising can effectively improve the signal quality without losing key physiological features.

[0102] The reconstruction sub-unit 413 is configured to dynamically adjust the weights of the coefficients at different scales according to the signal energy distribution, and generate the reconstructed multi-scale feature coefficients. The signal reconstruction process can be represented as:

[0103] S'(t) =∑ j，k w j ·C j，k ·ψ j，k (t),

[0104] wherein S'(t) is the reconstructed signal, which is a function of time t; w j is the weight coefficient of the j-th scale, which is used to adjust the contribution of different scales to the reconstructed signal; C j，k is the wavelet coefficient; ψ j，k (t) is the wavelet basis function; and the summation symbol ∑ j，k indicates the accumulation of all scales j and shifts k. The weight coefficient w j is dynamically adjusted according to the energy proportion of the signal at this scale:

[0105]

[0106] wherein E j=∑ k |C j，k | 2 Ej is the energy of the jth scale, calculated as the sum of the squares of all wavelet coefficients at this scale; E is the sum of all scale energies; J is the total number of scales, J = 8 in this embodiment. This dynamic weight adjustment strategy can emphasize the frequency bands containing pathological features, improving the sensitivity of abnormality detection.

[0107] Preferably, the reconstruction subunit 413 updates the reconstruction weights every 5 seconds to adapt to different body states and activity levels. The processing window length is chosen to be 8 seconds, ensuring that at least 6-10 complete pulse cycles are included, providing sufficient information for feature extraction and analysis. The choice of window length takes into account the range of heart rate variation: the resting heart rate of a normal adult is about 60-80 beats per minute, and can reach 100-120 beats per minute in a motion state, while some elderly or athletes have a resting heart rate as low as 40-50 beats per minute. Therefore, an 8-second window can contain a sufficient number of complete cardiac cycles in various situations, ensuring the reliability of the analysis.

[0108] Referring to Figure 6 , the nonlinear feature extraction and pattern recognition unit 42 includes an autocorrelation calculation subunit 421, a nonlinear fitting subunit 422, a feature extraction subunit 423, and a feature optimization subunit 424.

[0109] The autocorrelation calculation subunit 421 is used to calculate the autocorrelation sequence of the pulse wave under different time windows. The autocorrelation function R(τ) is defined as:

[0110]

[0111] where R(τ) is the autocorrelation function, indicating the degree of correlation between the signal and its own delay τ time; S(t) is the pulse wave signal; T is the total length of the signal, in seconds; τ is the time delay, in seconds; the integral sign represents the integral from 0 to T-τ; is the normalization factor, ensuring that the results of different delays τ are comparable. The autocorrelation calculation subunit 421 selects four different time windows (0.5 seconds, 1 second, 2 seconds, and 4 seconds) to calculate the autocorrelation sequence, capturing the signal characteristics at different time scales. In vascular monitoring, autocorrelation analysis of different window lengths is aimed at different vascular pathological states: the 0.5-second window mainly captures changes within a single cardiac cycle, suitable for detecting heart valve diseases; the 1-second window captures the relationship between adjacent heartbeats, suitable for detecting arrhythmias; the 2-second window is suitable for detecting changes in vascular compliance; and the 4-second window is suitable for detecting longer-term hemodynamic changes, such as venous reflux disorders. Through multi-window analysis, the system can comprehensively evaluate different types of vascular abnormalities.

[0112] The nonlinear fitting subunit 422 is configured to perform nonlinear fitting of the autocorrelation sequence with a Gaussian function. The Gaussian fitting function is defined as:

[0113]

[0114] where R'(τ) is the fitting function, which is a function of time delay τ; A is the amplitude parameter, reflecting the signal intensity; a is the decay coefficient, with a unit of s -2 , reflecting the signal decay speed; τ0is the time delay, with a unit of second; f b is the pulse wave pulsation frequency, with a unit of Hz; e is the base of natural logarithm, approximately equal to 2.71828; and cos is the cosine function. The nonlinear fitting is solved by using the least square method to obtain the parameters:

[0115]

[0116] where represents summation over N sampling points; R(τ i ) is the actual autocorrelation value; R'(τ i ) is the fitting value; and N is the number of sampling points. In the actual application of blood vessel monitoring, the fitting parameters have clear physiological significance: the amplitude parameter A reflects the intensity of the pulse wave signal, which is related to the blood vessel perfusion level; the decay coefficient a reflects the blood vessel elasticity and blood flow resistance, and an increase in a indicates a decrease in elasticity or an increase in resistance; and the pulsation frequency f b directly reflects the heart rate. This nonlinear fitting method can capture the nonlinear characteristics in the pulse wave signal and is more sensitive to small changes in blood vessel elasticity and blood flow resistance. For example, in the early stage of deep vein thrombosis, a slight increase in blood flow resistance will cause the decay coefficient a to increase by 15-20%, and the traditional linear analysis method may not be able to detect such a small change.

[0117] The feature extraction subunit 423 is configured to extract feature parameters reflecting blood vessel elasticity, blood flow resistance, cycle stability, amplitude variability, and waveform complexity from the autocorrelation sequence and the fitting result. The main features extracted include:

[0118] 1. Decay coefficient a: reflecting blood vessel elasticity and blood flow resistance, which is usually in the range of 0.2-0.8 s -2 in a healthy state, and increases when the blood vessel elasticity decreases or the blood flow resistance increases. For example, the a value of a young and healthy adult is usually in the range of 0.2-0.4 s -2 , while the a value of an elderly person is usually in the range of 0.4-0.6 s -2 due to decreased blood vessel elasticity, and the a value may rise to 0.6-0.8 s -2 in the early stage of deep vein thrombosis.

[0119] 2. Autocorrelation amplitude A: reflects the stability of pulse wave amplitude, usually greater than 0.85 in healthy state, and decreases when pulse is unstable. Amplitude stability is related to cardiac ejection function and peripheral vascular resistance, and A value of patients with heart failure or peripheral vascular disease may decrease to below 0.7.

[0120] 3. Period stability index: calculated by the autocorrelation function at the period point, in the range of 0.75-0.95 in healthy state. Decreased period stability may indicate arrhythmia or autonomic nervous dysfunction. The specific calculation is Where T p is the pulse period, R(T p ) is the autocorrelation value with one period interval, and R(0) is the autocorrelation value with zero delay (equal to signal energy).

[0121] 4. Spectral entropy: reflects the complexity of the waveform, usually in the range of 1.2-2.0 in healthy state, and increases in vascular disease. Spectral entropy calculation is based on the energy distribution of wavelet decomposition coefficients: Where is the energy proportion of the jth scale. Turbulence caused by vascular disease increases signal complexity, resulting in increased spectral entropy.

[0122] 5. Transmission time difference: the time difference of pulse wave arrival at different parts, reflecting vascular compliance, usually in the range of 0.1-0.3s. For example, the transmission time from the heart to the wrist is about 0.2s, and to the ankle is about 0.3s. Prolonged transmission time may indicate reduced vascular compliance or the presence of obstruction.

[0123] The feature selection subunit 424 is used to select the most discriminant features from the initial feature set according to the sample discrimination degree and stability score. The discrimination degree score is defined as:

[0124]

[0125] Where Score disc is the discrimination degree score, the larger the value, the stronger the discrimination ability; μ1 and μ2 are the feature means of the healthy group and the abnormal group respectively; σ1 and σ2 are the standard deviations of the two groups respectively; |μ1-μ2| represents the absolute value of the mean difference of the two groups. The stability score is defined as:

[0126]

[0127] Where Scoree stab is the stability score, the larger the value, the more stable the feature; CV is the coefficient of variation, representing the ratio of standard deviation to mean; μ is the feature mean; σ is the feature standard deviation. In practical application, the comprehensive score is calculated as the weighted sum of discrimination and stability: Score total=0.7·Scoree disc +0.3 Score stab The weights are set based on clinical validation results, reflecting the fact that discrimination is relatively more important in vascular abnormality detection. Feature optimization subunit 424 comprehensively considers both discrimination and stability, selecting the 20 features with the highest overall scores to form the feature vector F:

[0128] F = [f1, f2, ..., f 20 ] T ,

[0129] Where F is the eigenvector, a 20×1 column vector; f i This represents the i-th eigenvalue; the superscript T indicates vector transpose. Finally, the feature optimization subunit 424 performs Z-score standardization on the features to make features of different dimensions comparable:

[0130]

[0131] Among them, f i ′ represents the standardized eigenvalue; f i The original eigenvalues; μ i and σ i These are the mean and standard deviation of the feature in the training dataset, respectively. This standardization process ensures that different features can be compared on the same scale, such as the decay coefficient α (in seconds). -2 ) and periodic stability index (dimensionless).

[0132] Reference Figure 7 The multi-parameter adaptive judgment standard unit 43 includes a static threshold subunit 431, a group threshold subunit 432, a personalized threshold subunit 433, a state recognition subunit 434, and an anomaly classification subunit 435.

[0133] The static threshold subunit 431 is used to set fixed thresholds based on physiological common sense. These fixed thresholds are suitable for identifying extreme abnormalities, such as a sudden drop in pulse wave amplitude of more than 50% or its complete disappearance, which usually indicates severe blood flow obstruction requiring immediate intervention. Static thresholds are typically determined based on clinical experience, for example:

[0134] Pulse wave frequency: The normal range for adults is 0.8-2Hz (corresponding to a heart rate of 48-120 beats / minute). A frequency outside this range is considered abnormal.

[0135] Pulse wave amplitude: A decrease of >50% relative to the baseline is considered abnormal and may indicate acute vascular occlusion;

[0136] Pulse wave propagation time: A sudden increase of >50% is considered abnormal and may indicate acute thrombosis.

[0137] These static thresholds are universal and suitable for all user groups, mainly for capturing acute and severe vascular abnormal events. In clinical applications, for example, for high-risk patients who may have lacunar cerebral infarction, when the carotid artery pulse amplitude suddenly drops by more than 40%, the system will immediately issue an emergency warning, indicating that a vascular obstruction may occur.

[0138] The group threshold subunit 432 is used to set the group threshold according to factors such as age, gender, etc. The group threshold is based on large-scale population data statistics, taking into account physiological differences in different age groups and genders. For example, the vascular elasticity of the elderly group is usually lower than that of young people, so the upper limit of the normal range of the decay coefficient a is adjusted to 1.0 s -2 (compared to 0.8 s -2 for young people); the pulse amplitude of women is usually slightly lower than that of men, so an adjustment factor of 0.85 is considered when setting the threshold. In clinical applications, for example, for people over 70 years old, the system will automatically adjust the judgment threshold of the vascular elasticity-related parameters to reduce false positives due to physiological reduction in vascular elasticity; similarly, for child users, the system will adjust to a threshold range more suitable for their physiological characteristics.

[0139] The personalized threshold subunit 433 is used to learn the user's normal state range and establish a personalized model. The personalized threshold is calculated based on the user's historical data:

[0140] Threshold personal = μ personal ± 1.96·σper sonal ,

[0141] where Threshold personal is the personalized threshold, which can be an upper threshold (plus) or a lower threshold (minus); μ personal is the average value of the feature in the user's normal state; σ personal is the standard deviation; 1.96 corresponds to a 95% confidence interval, meaning that under normal conditions, the feature value has a 95% probability of falling within this interval. Preferably, the personalized threshold subunit 433 needs to accumulate at least 7 days of data to form a preliminary personalized model, and more than 30 days of data to form a stable personalized threshold. The application of personalized thresholds greatly improves detection accuracy, especially for people with special physiological characteristics. For example, some professional athletes may have lower resting heart rates and higher vascular elasticity, which may lead to false judgments if using universal thresholds; while personalized thresholds can adapt to these individual differences and reduce false positive rates.

[0142] The state recognition subunit 434 is used to distinguish between resting, light activity, and strenuous exercise states. State recognition is based on accelerometer data and heart rate variability analysis, and different criteria are applied in different states. The state division criteria are as follows:

[0143] Resting state: acceleration root mean square <0.1g, heart rate variability SDNN <50ms;

[0144] Light activity: acceleration root mean square 0.1-0.5g, heart rate variability SDNN 50-100ms;

[0145] Strenuous exercise: acceleration root mean square >0.5g, heart rate variability SDNN >100ms;

[0146] Wherein, SDNN is the standard deviation of heart rate interval, unit is millisecond (ms), which is a commonly used index of heart rate variability; g is the acceleration of gravity, about 9.8 m / s 2 In the resting state, the judgment update frequency is once per minute; in the active state, the frequency is increased to once every 30 seconds to adapt to the rapid changes in physiological state. In actual application, state recognition greatly improves the detection accuracy of the system. For example, after strenuous exercise, the vascular elasticity and hemodynamic parameters of the user will temporarily change significantly, at this time the system will not misjudge as a pathological state, but adjust the judgment criteria according to the activity level; when the user is in a resting state, similar changes may indicate vascular abnormalities, and the system will issue a warning.

[0147] The abnormal classification subunit 435 is used to classify abnormalities into five categories: vascular elasticity abnormality, blood flow resistance abnormality, vascular compliance abnormality, blood flow velocity abnormality, and blood flow pattern abnormality. Abnormal classification uses a decision tree method based on the combination of different features to determine the type of abnormality. For example:

[0148] Attenuation coefficient α increases (>0.8s -2 + propagation time shortens (<0.08s) → vascular elasticity abnormality, indicating arteriosclerosis;

[0149] Attenuation coefficient α increases (>0.8s -2) + propagation time lengthens (>0.3s) → blood flow resistance abnormality, indicating thrombosis;

[0150] Pulse wave amplitude decreases (<50% baseline) + spectral entropy increases (>2.2) → vascular compliance abnormality, indicating vascular stenosis;

[0151] Pulse wave propagation velocity changes (>±30%) → blood flow velocity abnormality, indicating abnormal blood pressure;

[0152] Waveform complexity increases (spectral entropy >2.5) + cycle stability decreases (<0.7) → blood flow pattern abnormality, indicating turbulent flow;

[0153] Different types of abnormalities correspond to different clinical significance and intervention measures: abnormal vascular elasticity suggests that there may be atherosclerosis, which requires lipid-lowering treatment; abnormal blood flow resistance suggests that there may be thrombosis, which requires anticoagulant treatment; abnormal vascular compliance suggests that there may be vascular stenosis, which requires vasodilator treatment; abnormal blood flow velocity suggests that there may be abnormal blood pressure, which requires adjustment of antihypertensive drugs; abnormal blood flow pattern suggests that there may be heart valve disease or large vessel disease, which requires further examination.

[0154] Preferably, the abnormality confirmation adopts a "3-time confirmation" mechanism: single threshold triggering observation, and continuous three times of threshold triggering warning, which effectively reduces the false positive rate. In clinical application, for example, in the monitoring of deep vein thrombosis of postoperative patients in bed, when the system detects a possible abnormality of blood flow resistance, it will continuously confirm for 3 times, and if all are over the threshold, an alarm will be issued to avoid false positives caused by temporary factors such as body position changes.

[0155] Reference Figure 8 The deep learning feature automatic extraction unit 44 includes an input layer 441, a feature extraction layer 442, an attention mechanism module 443, a residual connection module 444, and an output layer 445.

[0156] The input layer 441 is used to receive a 32-channel pulse wave signal, each channel containing 128 data points (corresponding to 0.64 seconds, sampling frequency 200 Hz). The input data is organized as a two-dimensional matrix of 32x128, with each row representing the time series signal of a sensor channel. In the actual application of vascular monitoring, the 32-channel design can cover a large enough monitoring area and capture signal features at different vascular positions. For example, when monitoring deep vein thrombosis of the lower extremities, the sensor array can cover the posterior tibial vein, the peroneal vein, and the great saphenous vein, etc. multiple key sites, providing a more comprehensive assessment of the vascular status.

[0157] The feature extraction layer 442 includes multiple convolutional layers and pooling layers for feature extraction from different perspectives. The present application adopts an improved LeNet architecture, which is specifically designed as follows:

[0158] First convolutional layer: 8 3x1 convolutional kernels (short-time convolution), step size 1, output size 32x126x8;

[0159] First pooling layer: 2x1 max pooling, step size 2, output size 32x63x8;

[0160] Second convolutional layer: 16 5x1 convolutional kernels (medium-time convolution), step size 1, output size 32x59x16;

[0161] Second pooling layer: 2x1 max pooling, step size 2, output size 32x29x16;

[0162] Third convolutional layer: 32 9x1 convolutional kernels (long-time convolution), stride 1, output size 32x21x32;

[0163] Third pooling layer: 3x1 max pooling, stride 3, output size 32x7x32;

[0164] This multi-view convolution design can capture local features, medium-range features and long-period changes at the same time, forming a comprehensive understanding of the pulse wave signal. All convolutional layers use the ReLU activation function, defined as:

[0165] ReLU(x) = max(0, x),

[0166] where ReLU(x) is the output value of the activation function; x is the input value; max represents the larger of the two. The ReLU function has the advantages of simple calculation and stable gradient, and is suitable for deep network training. In the application of blood vessel monitoring, different sizes of convolutional kernels are used to capture different types of blood vessel abnormalities: short-time convolution (3x1) is suitable for capturing pulse wave peak features and is sensitive to blood vessel resistance changes; medium-time convolution (5x1) is suitable for capturing pulse wave shape changes and is sensitive to blood vessel elasticity changes; long-time convolution (9x1) is suitable for capturing changes in multiple cardiac cycles and is sensitive to heart rhythm and blood flow dynamics stability. This multi-scale design enables the network to comprehensively analyze different types of blood vessel abnormalities.

[0167] The attention mechanism module 443 is used to automatically identify and enhance the features of important sensor channels. The channel attention mechanism calculates the importance weight of each channel:

[0168] w c = σ(W2·ReLU(W1·GAP(F c ))),

[0169] where w c is the attention weight of the cth channel, with a value range of [0, 1]; F c is the feature map of the cth channel, with a dimension of LxD (L is the sequence length and D is the feature dimension); GAP represents the global average pooling operation, which compresses the LxD feature map into a 1xD vector; W1 and W2 are learnable weight matrices that map the feature dimension from D to D / r and back to D, respectively, where r is the dimension reduction ratio, usually set to 16; ReLU is the activation function; and σ is the Sigmoid activation function:

[0170]

[0171] where σ(x) is the output value of the Sigmoid function, with a value range of (0, 1); x is the input value; and e is the base of the natural logarithm, approximately 2.71828. The attention weight is applied to the original feature to obtain the weighted feature:

[0172] F c ′=w c ·F c ,

[0173] where F′ c is the weighted feature map; w c is the attention weight; F c is the original feature map; and · denotes the scalar multiplication of a matrix operation. This attention mechanism can automatically identify and enhance sensor channels containing useful information, suppress the influence of noise or irrelevant channels, and improve the effectiveness of feature extraction. In the practical application of vascular monitoring, attention mechanisms are particularly useful because not all sensor channels can capture valuable signals. For example, when monitoring deep vein thrombosis, only the sensor channels located above the main veins contain key information. Attention mechanisms can automatically identify these channels and enhance their weights while suppressing the interference of other channels, improving detection accuracy.

[0174] The residual connection module 444 is used to solve the gradient vanishing problem in deep networks. The residual connection is defined as:

[0175] H(x) = F(x) + x,

[0176] where H(x) is the output of this layer; F(x) is the output of the regular convolutional layer; x is the input; and + denotes element-wise addition. By adding a jump connection, the gradient can be directly propagated from the deep layer to the shallow layer, effectively solving the gradient vanishing problem in deep neural network training. In the application of vascular monitoring, residual connections ensure that deep networks can effectively learn, especially when analyzing complex vascular abnormality patterns, such as combined lesions of atherosclerosis and venous thrombosis, which require the powerful representation ability of deep networks.

[0177] The output layer 445 is used to generate the vascular state classification results. The output layer includes two fully connected layers:

[0178] First fully connected layer: 128 neurons, ReLU activation;

[0179] Second fully connected layer (output layer): 2 neurons, Softmax activation;

[0180] The Softmax function is defined as:

[0181]

[0182] where Softmax(z_i) is the output probability of the i-th class; z iis the original output for the i-th class; K is the total number of classes, K = 2 in this example, representing normal and abnormal states; e is the base of natural logarithm, approximately 2.71828; represents the summation over all K classes. The Softmax function converts the original output into a probability distribution, with the output values representing the probabilities of the sample belonging to each class. In the blood vessel monitoring application, the final output is [p_normal, p_abnormal], representing the probabilities of the blood vessel being in normal and abnormal states, respectively. The system usually sets p_abnormal > 0.7 as the abnormal alarm threshold, while 0.5 < p_abnormal < 0.7 is considered as the observation state.

[0183] Preferably, the deep learning feature automatic extraction unit 44 uses a cross-entropy loss function for model training:

[0184]

[0185] where Loss is the loss function value; N is the number of samples; K is the number of classes, K = 2 in this example; y ij is the true label (0 or 1) of sample i belonging to class j; p ij is the predicted probability; log is the natural logarithm function; represents the summation over all N samples; represents the summation over all K classes. Training uses small batch gradient descent with a batch size of 32, an initial learning rate of 0.001, and a learning rate decay of 10% after every 50 training rounds. To prevent overfitting, the model uses Dropout technology with a proportion of 0.5, as well as L2 regularization with a weight decay coefficient of 0.0001. In practical applications for blood vessel monitoring, the model training uses a labeled data set from clinical studies, including data from healthy controls and patients with different types of vascular abnormalities. Through transfer learning technology, the model can quickly adapt to the physiological characteristics of new users, improving the accuracy of personalized detection.

[0186] Referring to Figure 9 , the multi-site information fusion decision unit 45 includes a weight allocation subunit 451, a time series analysis subunit 452, a risk assessment subunit 453, and a report generation subunit 454.

[0187] The weight allocation subunit 451 is used to allocate weights to the monitoring results of different sites. The weight allocation uses a combination of static weights and dynamic weights:

[0188] The static weights are pre-set based on medical knowledge, for example:

[0189] Lower extremity site (ankle): weight 0.4, most prone to deep vein thrombosis;

[0190] Upper limb (wrist): weight 0.3, good vascular status and stable monitoring;

[0191] Neck: weight 0.3, reflecting the brain blood supply situation;

[0192] Dynamic weight is adjusted according to signal quality and historical accuracy:

[0193]

[0194] where w′ i is the adjusted weight; w i is the static weight of part i; q i is the signal quality score, with a value range of [0, 1], 1 indicating the best quality; a i is the historical accuracy, with a value range of [0, 1]; and are the average signal quality and average accuracy, respectively; λ and μ are adjustment coefficients, usually set to 0.2 and 0.3; is the signal quality adjustment factor; is the accuracy adjustment factor. Finally, the weight is normalized:

[0195]

[0196] where w″ i is the normalized weight, satisfying M is the total number of monitoring parts; is the sum of all adjusted weights. In the actual application of vascular monitoring, dynamic weight adjustment can adapt to individual differences of different users. For example, for obese patients, the ankle signal quality may be poor, and the system will automatically reduce its weight and increase the weight of parts with better signal quality to ensure the accuracy of the overall evaluation.

[0197] The time series analysis subunit 452 is used to analyze short-term trend changes, medium-term periodic changes, and long-term evolution trends. The short-term trend analysis detects the change trend of consecutive 3-5 measurements, and defines the trend index as:

[0198]

[0199] where Trend is the trend index, with a value range of [-1, 1]; N is the number of consecutive measurements, usually 5; x i is the i-th measurement result; sign is the sign function, returning 1 when the input is positive, -1 when the input is negative, and 0 when the input is zero; represents the sum of all i from 1 to N-1; |Trend|>0.6 indicates a significant trend, positive value indicates an upward trend, and negative value indicates a downward trend. In vascular monitoring applications, short-term trend analysis can capture acute changes, such as rapid increases in postoperative patient deep vein thrombosis risk, allowing medical staff to intervene in time.

[0200] Mid-term pattern recognition analyzes periodic changes within 24 hours, using autocorrelation analysis to detect periodicity:

[0201]

[0202] where R day (τ) is the autocorrelation function of 24-hour data; X(t) is the measurement sequence within 24 hours; T is the total duration, which is 24 hours; τ is the time delay; represents the integral from 0 to T-τ; is the normalization factor. The peak position indicates the period length, and the peak height indicates the periodicity strength. In clinical applications, mid-term pattern recognition can find patterns of vascular state changes related to daily activities, such as slowed venous return in the lower extremities after long periods of sitting, or postprandial vasodilation, providing personalized lifestyle recommendations for users.

[0203] Long-term evolution tracking monitors slow change trends over 7-30 days, using linear regression analysis:

[0204] X(t) = β0 + β1·t + ∈,

[0205] where X(t) is the measurement at time t; β0 is the intercept, indicating the initial value; β1 is the slope, indicating the rate of change; t is the time variable; ∈ is the error term, following a normal distribution with mean 0. A significant P<0.05 of β1 indicates a significant long-term trend. In vascular monitoring applications, long-term trend analysis is particularly valuable for evaluating treatment effectiveness or disease progression. For example, for hypertensive patients, monitoring the long-term trend of vascular elasticity can evaluate the effectiveness of antihypertensive drugs; for patients with atherosclerosis, it can evaluate the effectiveness of lipid-lowering therapy.

[0206] Risk assessment sub-unit 453 is used to determine the risk level based on comprehensive information. Risk assessment uses a five-level risk scale:

[0207] 1. Attention level: single abnormality, risk index 0.2-0.4, no immediate intervention required;

[0208] 2. Warning level: consecutive abnormalities but mild, risk index 0.4-0.6, recommended to reduce activity;

[0209] 3. Alarm level: consecutive significant abnormalities, risk index 0.6-0.7, recommended to rest and monitor;

[0210] 4. Emergency level: severe abnormality, risk index 0.7-0.8, suggesting immediate medical treatment;

[0211] 5. Critical level: extreme abnormality, risk index > 0.8, automatically notifying emergency contacts;

[0212] The risk index R is calculated as a weighted average:

[0213]

[0214] where R is the comprehensive risk index, with a value range of [0, 1]; r i is the risk score of part i, with a value range of [0, 1]; w i is the normalized weight; which represents the summation of all M monitoring parts. In practical applications of vascular monitoring, risk levels are associated with specific intervention measures. For example, for warning level risk, the system will suggest that the user reduce sedentary time, be more active, or elevate the lower limbs; for alarm level risk, the system will suggest that the user rest in bed and continue monitoring; and for emergency level and critical level risks, the system will suggest immediate medical treatment, and can automatically notify the preset emergency contacts.

[0215] The report generation subunit 454 is used to integrate information and generate a comprehensive assessment report containing vascular elasticity, blood flow resistance, and thrombus risk. The report content includes: current risk level and trend; abnormal type and its characteristic description; detailed data of each monitoring part; historical data comparison and analysis and personalized health recommendations;

[0216] Preferably, the report uses visual charts to display the results, making it easy for users to understand. The regular report is generated once a day, and the abnormal situation is pushed in real time, ensuring that the user can understand their own health status in a timely manner. In clinical applications, for inpatient users, the report can be directly integrated into the electronic medical record system for medical staff to consult; for home users, the report is displayed through a mobile application and can be selectively shared with family doctors, forming a complete health management closed loop.

[0217] Referring to Figure 10 , the application also provides a pulse analysis method applied to the above-mentioned vascular hemodynamic abnormality monitoring system, comprising the following steps:

[0218] Step S1: Collecting pulse wave signals and sending the collected pulse wave signals to a data processing host.

[0219] In this step, the flexible piezoelectric sensor array collects the human vascular pulse wave signal with a sampling frequency of 200 Hz. After pre-processing the signal, the wearable monitoring device transmits the data to the data processing host through Bluetooth 5.0. The transmitted data includes the original pulse wave signal and sensor position information. To ensure data quality, the system will evaluate the signal quality in real time, and trigger a re-collection or position adjustment prompt when the quality score is below 0.6 (full score 1.0). In actual application scenarios, for example, for critically ill patients who are long-term bedridden, the system will automatically perform a complete collection cycle every hour to monitor the risk of deep vein thrombosis; while for office workers who sit for a long time, the system can be set to automatically increase the collection frequency after sitting for 2 hours.

[0220] Step S2: Multi-scale decomposition of pulse wave signal to obtain feature coefficient set at different scales.

[0221] In this step, the multi-scale signal decomposition and reconstruction unit receives the pulse wave signal and decomposes it into multiple scale coefficient sets through wavelet transform. Daubechies wavelet (db4) is used for 8-layer decomposition to capture signal characteristics in the 0.5-40 Hz range. The coefficients are subjected to soft threshold denoising, and the weights are dynamically adjusted according to the energy distribution to generate the reconstructed multi-scale feature coefficients. This step can effectively separate the characteristics of different frequency components, for example, in the early stage of monitoring deep vein thrombosis, the energy distribution change in the 4-8 Hz frequency band is a key indicator, and multi-scale decomposition can effectively extract the characteristics of this frequency band without interference from other frequency band signals.

[0222] Step S3: Extract nonlinear features from the feature coefficient set, including vessel elasticity, blood flow resistance, cycle stability, amplitude variability, and waveform complexity features.

[0223] In this step, the nonlinear feature extraction and pattern recognition unit calculates the autocorrelation sequence of the pulse wave in different time windows, and uses a Gaussian function for nonlinear fitting. From the autocorrelation sequence and fitting results, feature parameters reflecting the state of the blood vessels are extracted, including attenuation coefficient, autocorrelation amplitude, cycle stability index, spectral entropy, and propagation time difference. The feature selection sub-unit selects the 20 most discriminative features according to the discriminability and stability score, and performs Z-score standardization. In clinical applications, different types of vascular abnormalities exhibit different feature combinations, for example, venous thrombosis usually shows increased blood flow resistance and prolonged propagation time, while atherosclerosis mainly shows reduced vessel elasticity and pulse wave deformation. By extracting a comprehensive feature set, the system can distinguish different types of vascular abnormalities.

[0224] Step S4: Establish multi-level adaptive judgment criteria based on individual differences and physiological state to evaluate nonlinear features.

[0225] In this step, the multi-parameter adaptive judgment criterion unit establishes a three-level judgment criterion system: static threshold, group threshold, and individualized threshold. The state recognition subunit distinguishes between rest, light activity, and intense exercise states, and applies different criteria in different states. The abnormality classification subunit classifies abnormalities into five categories, and uses a "3-time confirmation" mechanism to reduce the false positive rate. In actual application, for example, for elderly diabetic patients over 65 years old, the system will automatically adjust the group threshold of the vascular elasticity parameter, and adjust the normal upper limit of the attenuation coefficient a from 0.8 s -2 to 1.0 s -2 to adapt to the physiological characteristics of the natural reduction of vascular elasticity in elderly diabetic patients; at the same time, for individual differences, the system will establish an individualized baseline based on at least 7 days of historical data to make the threshold more accurate.

[0226] Step S5: Automatically extract deep features from the original signal using a deep learning network, and perform vascular state classification.

[0227] In this step, the deep learning feature automatic extraction unit receives the 32-channel pulse wave signal, and performs feature learning and classification through an improved LeNet network. The network includes multi-view convolution design, attention mechanism, and residual connection, which can capture complex time and frequency domain features. The output layer uses the Softmax function to generate the vascular state classification result and its confidence. In the application of vascular monitoring, the deep learning network can identify complex patterns that traditional feature extraction methods are difficult to capture, especially in the case of coexistence of multiple pathological states. For example, when a patient has both hypertension and venous insufficiency, the signal features will become complex, and traditional methods may have difficulty distinguishing them, while the deep learning network can effectively identify these complex patterns by learning a large amount of labeled data.

[0228] Step S6: Fuse the evaluation results of multiple monitoring sites, perform time series pattern analysis and comprehensive risk assessment.

[0229] In this step, the multi-site information fusion decision unit assigns weights to the monitoring results of different sites, analyzes short-term trend changes, medium-term periodic changes, and long-term evolution trends. The risk assessment subunit determines the risk level according to the comprehensive information, and uses a five-level risk level system. In clinical applications, multi-site monitoring and information fusion significantly improve detection accuracy. For example, when monitoring the risk of deep vein thrombosis in the lower extremities, the ankle and knee positions are monitored simultaneously. If an abnormality occurs in a single site, it may be caused by local factors, while when multiple sites simultaneously show increased blood flow resistance signals, it is more likely to be a true thrombosis risk increase, which requires timely intervention.

[0230] Step S7: Generate a vascular health report, including abnormal type, risk level, and health recommendations.

[0231] In this step, the report generation subunit integrates various types of information to generate a comprehensive evaluation report. The report content includes the current risk level and trend, the type of abnormality and its characteristic description, the detailed data of each monitoring site, the comparative analysis of historical data, and personalized health recommendations. The routine report is generated once a day, and the abnormal situation is pushed in real time. In clinical application, the timeliness and pertinence of the report are crucial. For example, for patients at high risk of postoperative venous thrombosis, when the system detects a warning level risk, it will immediately notify the nursing station and provide specific risk sites and characteristic descriptions in the report to guide nursing staff to conduct targeted intervention; for home users, the report will provide easy-to-understand health recommendations, such as "detected slow venous return in lower extremities, recommended to avoid long-term maintaining the same posture, stand for 5 minutes every hour, and if necessary, elevate the lower extremities for rest".

[0232] Through the above steps, the present application can realize accurate assessment and early warning of vascular health status, especially early identification of deep venous thrombosis. Compared with the prior art, the present application has the advantages of non-invasiveness, continuous monitoring, personalized assessment, multi-site collaborative judgment, etc., and provides an effective tool for the prevention and early intervention of vascular diseases.

[0233] The present application is not limited to the above embodiments, and those skilled in the art can adjust the above embodiments without departing from the spirit of the present application. The scope of protection of the present application is subject to the claims.

Claims

1. A system for monitoring abnormal vascular hemodynamics, characterized in that, include: Flexible piezoelectric sensor arrays are used to collect pulse wave signals from human blood vessels; A wearable monitoring device is electrically connected to the flexible piezoelectric sensor array to receive the pulse wave signal and perform preliminary preprocessing. The data processing host is wirelessly connected to the wearable monitoring device and is used to receive and process the pulse wave signal; The pulse wave analysis module is located in the data processing host and is used to perform multi-level analysis on the pulse wave signal. The pulse wave analysis module includes a multi-scale signal decomposition and reconstruction unit, a nonlinear feature extraction and pattern recognition unit, a multi-parameter adaptive judgment standard unit, a deep learning feature automatic extraction unit, and a multi-part information fusion decision unit connected in sequence. as well as The application platform, connected to the data processing host via a network, is used to receive processing results and generate vascular health reports; The nonlinear feature extraction and pattern recognition unit includes: The autocorrelation calculation subunit is used to calculate the autocorrelation sequence of pulse waves under different time windows; A nonlinear fitting subunit is used to perform nonlinear fitting on the autocorrelation sequence using a Gaussian function; The feature extraction subunit is used to extract feature parameters reflecting vascular elasticity, blood flow resistance, periodic stability, amplitude variability, and waveform complexity from the autocorrelation sequence and fitting results; and The feature selection subunit is used to select the most discriminative features from the initial feature set based on sample discriminability and stability scores. The multi-parameter adaptive judgment criterion unit includes: The static threshold subunit is used to set a fixed threshold based on common physiological knowledge; The group threshold subunit is used to set the grouping threshold based on factors such as age and gender. The personalized threshold subunit is used to learn the range of normal user states and build a personalized model. A state recognition subunit is used to distinguish between resting, mild activity, and vigorous exercise states; and The abnormality classification subunit is used to classify abnormalities into five categories: abnormal vascular elasticity, abnormal blood flow resistance, abnormal vascular compliance, abnormal blood flow velocity, and abnormal blood flow pattern.

2. The vascular hemodynamic abnormality monitoring system according to claim 1, characterized in that, The multi-scale signal decomposition and reconstruction unit includes: The signal decomposition subunit is used to receive the pulse wave signal and decompose the pulse wave signal into a set of coefficients at multiple scales using various wavelet basis functions. A denoising processing subunit is used to perform soft-threshold shrinking denoising on the coefficient set; and The reconstruction subunit is used to dynamically adjust the weights of coefficients at different scales based on the signal energy distribution and generate reconstructed multi-scale characteristic coefficients.

3. The vascular hemodynamic abnormality monitoring system according to claim 1, characterized in that, The deep learning feature extraction unit includes: The input layer is used to receive 32-channel pulse wave signals. The feature extraction layer includes multiple convolutional and pooling layers for feature extraction from different perspectives; Attention mechanism module, used to automatically identify and enhance the features of important sensor channels; Residual connection modules are used to solve the vanishing gradient problem in deep networks; and The output layer is used to generate the blood vessel status classification results.

4. The vascular hemodynamic abnormality monitoring system according to claim 1, characterized in that, The multi-part information fusion decision-making unit includes: The weight allocation subunit is used to assign weights to the monitoring results of different parts; The time series analysis sub-unit is used to analyze short-term trend changes, medium-term cyclical changes, and long-term evolution trends. The risk assessment subunit is used to determine the risk level based on comprehensive information; and The report generation subunit is used to integrate information and generate a comprehensive assessment report that includes vascular elasticity, blood flow resistance, and thrombosis risk.

5. The vascular hemodynamic abnormality monitoring system according to claim 1, characterized in that, The flexible piezoelectric sensor array includes 32 piezoelectric sensors, which are made of medical-grade polyvinylidene fluoride piezoelectric film and arranged in a 4×8 matrix structure. Each sensor has a size of 3mm×3mm and a thickness of 0.2mm.

6. The vascular hemodynamic abnormality monitoring system according to claim 1, characterized in that, The wearable monitoring device includes: The data acquisition unit is used for multi-channel synchronous acquisition and control. The signal preprocessing unit is used to perform bandpass filtering, notch filtering, and outlier detection on the acquired signal; A Bluetooth communication unit is used to transmit the processed data to the data processing host; and The power management unit is used to provide stable power and achieve low power management.

7. The vascular hemodynamic abnormality monitoring system according to claim 1, characterized in that, The application platform includes: A data receiving module is used to receive the processing results sent by the data processing host; Data storage module, used to store historical monitoring data; The analysis and visualization module is used to generate visual charts and trend analyses; The early warning and alert module is used to issue an alarm when an anomaly is detected; and The user interaction module is used to receive user feedback and configure system parameters.

Images