Deep vein thrombosis continuous monitoring system and method based on multi-wavelength optoelectronic sensor

Abstract

The present application belongs to the technical field of intelligent wearable medical health monitoring, and particularly relates to a deep vein thrombosis continuous monitoring system and method based on a multi-wavelength photoelectric sensor. The monitoring system comprises sequentially connected lower limb blood flow signal acquisition module, intelligent signal processing module and early risk assessment module. The lower limb blood flow signal acquisition module comprises a master control module, a multi-wavelength photoelectric signal acquisition module and a cuff module. The intelligent signal processing module comprises a signal preprocessing module and an alternating component time domain feature extraction module, an alternating component frequency domain feature extraction module, a direct current component feature extraction module and a venous vessel resistance feature extraction module connected with the signal preprocessing module respectively. The early risk assessment module comprises a feature processing dimension reduction module, a lower limb blood flow state evaluation module, a risk area identification module and a risk level assessment module. The present application can judge the lower limb state, possible diseased area and risk level of a user, and realize early warning of deep vein thrombosis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent wearable medical and health monitoring, and particularly relates to a deep vein thrombosis continuous monitoring system and method based on a multi-wavelength photoelectric sensor. BACKGROUND

[0002] Deep Venous Thrombosis (DVT) is a backflow disorder caused by abnormal blood clotting in veins, commonly found in the upper femur, popliteal region, posterior tibial region and peroneal veins of the lower limbs, and the main causes are damage to the vein wall, blood clotting and slow blood flow. The main manifestations of acute lower limb DVT are fever and swelling of the affected limb, and the induced Pulmonary Embolism (PE) is one of the major risks of sudden death. According to statistics, about 50% of users who undergo knee replacement surgery will develop DVT, with an annual overall incidence of about 0.1%, and 1 / 3 of the users will develop a worsening condition. However, more than 60% of clinical users belong to asymptomatic DVT, which shows that the disease has considerable concealment.

[0003] At present, compression ultrasound, venography and magnetic resonance imaging can be used for clinical lower limb venous thrombosis screening. Although the above methods have relatively accurate results, they are subject to tight medical resources, long diagnosis time, high diagnosis cost and heavy burden on medical staff, and are more commonly used for highly detailed measurements, such as before surgery planning and during intervention. The current clinical diagnosis strategy cannot achieve early continuous diagnosis of DVT, which brings great risk to acute clinical users and postoperative users. Therefore, improvement of the DVT diagnosis method has a positive effect on reducing the incidence.

[0004] Currently, there are mainly three types of related technologies for the diagnosis of deep vein thrombosis in the lower extremities: (1) Post-processing method based on magnetic resonance imaging (MRI). There are examples of using a YOLOv3 detection network to build a deep learning model, using K-means clustering analysis to optimize the length-width ratio of the thrombus detection box, and inputting a five-channel image matrix to train the thrombus detection model, which can realize the detection of venous thrombosis and the framing of the thrombus area in new MRI images (Patent No.: ZL202010387879.5). Another example uses a semantic segmentation network method for model training, extracts high-dimensional semantic features, and selects features through a clustering algorithm to build a prediction model for the binary classification diagnosis of thrombolysis efficacy (Patent No.: ZL201911229603.8). Users of this type of diagnostic method still cannot avoid the constraints of the shortage of medical imaging resources when collecting images, and cannot effectively monitor and diagnose DVT in the early stage. (2) Post-processing method based on medical information database. There are examples of using clinical feature data collected from hospital medical data management systems to build a decision model through XGBoost algorithm to estimate the incidence of DVT in users after lower extremity fracture surgery (Patent No.: ZL202210521778.1). Another example uses collected user information for statistical modeling analysis. By determining the risk factors, a Logistic regression model and a scoring system are established to realize high-risk prediction (Patent No.: ZL202110307118.9). This type of diagnostic method ignores the individual differences of users and the timeliness of data, and is more likely to have inaccurate predictions, making it difficult to achieve effective early diagnosis. (3) Blood oxygen signal processing method based on near-infrared light. There are examples of using a light transmission detection system to continuously monitor the changes in microcirculation blood volume using infrared sensors to compare the changes in blood volume between the two limbs, and to evaluate unilateral deep vein thrombosis and vascular lesions (Patent No.: ZL200580035023.5). However, this diagnostic method has the shortcomings of missing DVT in both limbs and diagnostic errors caused by congenital constitution, and cannot be used as a sufficient condition for DVT diagnosis.

[0005] The functions proposed in the above patents are not effective in early diagnosis and continuous monitoring of DVT, mainly being improvements on post-processing of existing signal collection methods. There are few patents on early diagnosis of deep vein thrombosis, and the principles and accuracy of the models and algorithms need to be improved, and the functions are relatively single. Most patents do not have early warning functions for lower extremity blood flow, which may lead to missed detection and misdiagnosis of deep vein thrombosis. SUMMARY

[0006] To solve the above technical problems, the present application provides a deep vein thrombosis continuous monitoring system and method based on a multi-wavelength photoelectric sensor, which realizes early warning and evaluation of lower extremity venous thrombosis.

[0007] The application is implemented by providing a deep vein thrombosis continuous monitoring system based on a multi-wavelength photoelectric sensor, which comprises sequentially connected lower limb blood flow signal acquisition module, intelligent signal processing module and early risk assessment module, the lower limb blood flow signal acquisition module comprises a main control module, a multi-wavelength photoelectric signal acquisition module and a cuff module, the multi-wavelength photoelectric signal acquisition module and the cuff module are connected with the main control module, the photoelectric signal collected by the multi-wavelength photoelectric signal acquisition module comprises infrared light, red light, yellow light and green light signals, the intelligent signal processing module comprises a signal preprocessing module and an alternating component time domain feature extraction module, an alternating component frequency domain feature extraction module, a direct current component feature extraction module and a venous vessel resistance feature extraction module connected with the signal preprocessing module respectively, the main control module is connected with the signal preprocessing module, the early risk assessment module comprises a feature processing dimension reduction module, a lower limb blood flow state evaluation module, a risk area identification module and a risk level assessment module, the alternating component time domain feature extraction module, the alternating component frequency domain feature extraction module, the direct current component feature extraction module and the venous vessel resistance feature extraction module are connected with the feature processing dimension reduction module, the feature processing dimension reduction module is connected with the lower limb blood flow state evaluation module, and the lower limb blood flow state evaluation module is connected with the risk area identification module and the risk level assessment module respectively.

[0008] Preferably, the main control module comprises a shell, a main control MCU, a power supply and a storage unit are arranged in the shell, an inflatable cuff interface, a multi-channel signal transmission interface and a charging transmission interface are arranged on the surface of the shell, the main control MCU is connected with the power supply, the storage unit, the inflatable cuff interface and the multi-channel signal transmission interface respectively, the charging transmission interface is connected with the power supply, and the main control MCU is connected with the signal preprocessing module;

[0009] The multi-wavelength photoelectric signal acquisition module comprises a flexible PCB, a multi-wavelength photoelectric sensor is arranged on the flexible PCB, magic tapes are arranged on the two sides of the flexible PCB, and the multi-channel signal transmission interface is connected with the multi-wavelength photoelectric sensor through a transmission line.

[0010] The cuff module comprises an inflatable cuff and a cuff pressure control module, the inflatable cuff interface is connected with the cuff pressure control module through a skin catheter, and the cuff pressure control module is connected with the inflatable cuff.

[0011] Further preferably, the multi-wavelength photoelectric sensor comprises an infrared LED, a yellow LED, a red LED, a green LED and a PD light receiver, all of which are connected with the main control MCU.

[0012] The application further provides a deep vein thrombosis continuous monitoring and early warning method based on a multi-wavelength photoelectric sensor, which uses the deep vein thrombosis continuous monitoring system based on a multi-wavelength photoelectric sensor, and comprises the following steps:

[0013] 1) the user wears the multi-wavelength photoelectric signal acquisition module on the monitoring area, acquires blood flow state data, wears the cuff module on the area to be pressurized, realizes the pressurization of the area to be pressurized and the dynamic measurement of the lower limb pulse wave signal under different pressure conditions, and the data of the monitoring area, the data of the area to be pressurized, the blood flow state data and the lower limb pulse wave data form an acquisition information feature data set;

[0014] 2) the signal preprocessing module carries out noise reduction and smoothing processing on the lower limb pulse wave data, separates the direct current component DC(t) and the alternating current component AC(t) of the lower limb pulse wave signal, demultiplexes the four source signals of the alternating current component to obtain AC IR (t), AC R (t), AC Y (t) and AC G (t);

[0015] 3) the alternating current component time domain feature extraction module carries out feature extraction on the alternating current component AC(t) of the preprocessed lower limb pulse wave data in the time domain, and obtains the feature points of the original pulse wave, the feature points of the first-order pulse wave, the feature points of the second-order pulse wave and the feature points of the third-order pulse wave;

[0016] 4) the alternating current component frequency domain feature extraction module extracts the features of the alternating current component AC(t) based on frequency domain analysis, including power spectrum relationship features and frequency spectrum passband relationship features;

[0017] 5) the direct current component feature extraction module carries out feature extraction on the direct current component DC(t) of the preprocessed lower limb pulse wave data in the dynamic measurement;

[0018] 6) the venous vascular resistance feature extraction module carries out venous vascular resistance feature extraction on AC IR (t), AC R (t), AC Y (t) and AC G (t) obtained in step 2);

[0019] 7) the early risk assessment module analyzes the features obtained in steps 3), 4), 5) and 6), and evaluates the lower limb blood flow state, the risk level and the risk area:

[0020] 7.1) the feature processing dimension reduction module processes and reduces the dimensions of the features obtained in steps 3), 4), 5) and 6);

[0021] 7.2) the lower limb blood flow state evaluation module constructs a lower limb blood flow state classifier according to the blood flow state data, predicts the lower limb blood flow state data of the user through the classifier, obtains the lower limb blood flow state label of the user, and is used for evaluating the blood flow state;

[0022] 7.3) The risk area identification module compares the results of step 7.1) with the data of the area to be monitored and the data of the area to be pressurized, constructs a risk area decision tree, and predicts and identifies the area that is likely to be diseased;

[0023] 7.4) The risk level evaluation module constructs a risk evaluation score S, compares the blood flow state label of the user's lower limbs with the feature result S obtained through training, and according to the matching degree, obtains a preliminary risk evaluation score in three evaluation levels of low, medium and high;

[0024] 8) According to the results of the early risk evaluation module, an alarm is sent to the user.

[0025] Preferably, step 3) specifically comprises:

[0026] Step 3.1) The feature points of the original pulse wave, using the local maximum and minimum method to obtain the starting point value onset, the systolic peak value sys, the dicrotic notch dic, the diastolic peak value dia and the ending point value end;

[0027] Step 3.2) The feature points of the first-order pulse wave, using the local maximum to obtain the maximum slope point ms;

[0028] Step 3.3) The feature points of the second-order pulse wave, using a zero-phase second-order Butterworth filter with a passband of 0.5-15Hz to eliminate the components of the frequency band that do not contribute; the negative part of the filtered second-order pulse wave Z[n] is set to 0, and y(n)=Z 2 [n] is obtained, that is, the squared second-order pulse wave; using two moving average value formulas to generate the target area:

[0029]

[0030]

[0031] Wherein, W1 represents the systolic peak duration, W2 represents the average size of each cycle of the pulse wave, the size of MA peak [n] and MA beak [n] is compared to remove invalid areas; finally, the threshold method is used to obtain the effective area, and the region maximum absolute value method is used to detect the points a, c, d and e; the point b is the first minimum value after the point a, and the local minimum value method is used to detect the point b, wherein the point a is the first positive maximum value point of the systolic phase of the second-order pulse wave; the point b is the first negative maximum value point of the systolic phase of the second-order pulse wave; the point c is the second maximum value point of the systolic phase of the second-order pulse wave; the point d is the second minimum value point of the systolic phase of the second-order pulse wave; the point e is the first maximum value point of the diastolic phase of the second-order pulse wave;

[0032] Step 3.4) Feature points of the third order pulse wave, using the local maximum method to search the first local maximum p1 of the third derivative after b point, using the local minimum method to search the last local minimum p2 of the third derivative before d point.

[0033] Further preferably, step 4) specifically comprises:

[0034] Step 4.1) Extraction of power spectrum peak feature points, Welch algorithm is used to calculate the power spectrum density of the alternating current component AC(t) of the lower limb pulse wave signal;

[0035] Step 4.1.1) Zero-mean AC(t) to make the center at zero point; segment AC(t) with 50% overlap between segments; take the first 8 segments, smooth them with Hamming window, and then calculate the corresponding periodogram of each segment; average all periodograms to obtain the power spectrum density of AC(t);

[0036] Step 4.1.2) Use the local maximum method to extract the first 6 relatively obvious power spectrum peak points PSD1-PSD6;

[0037] Step 4.2) Extraction of spectral passband feature values, using the characteristic frequency band framework method to calculate the spectral energy;

[0038] Step 4.2.1) Set the fingertip lower limb pulse wave signal Tip(t) as the reference signal, use fast Fourier transform and filter through S-G filter to obtain the smoothed reference signal spectrum FTip(ω);

[0039] Step 4.2.2) Use the reference signal to extract three groups of central characteristic frequencies to generate the reference signal spectrum framework; set FP1 as the frequency of the peak point of FTip(ω) in the interval [0.8Hz, 1.5Hz], FP2 as the frequency of the peak point of FTip(ω) in the interval [2×fp1-0.5Hz, 2×fp1+0.5Hz], and FP3 as the frequency of the peak point of FTip(ω) in the interval [3×fp1-0.5Hz, 3×fp1+0.5Hz]; according to the -3db bandwidth principle, determine the upper and lower limit cutoff frequencies of the adjacent frequency domain of FP1, FP2 and FP3 to obtain three reference signal spectral passband characteristic intervals

[0040] Step 4.2.3) Use the reference signal spectral passband characteristic interval to obtain three spectral passband characteristic values of the alternating current component; extract the characteristic values in the spectral passband characteristic interval respectively Obtain P1, P2, P3;

[0041] Step 4.3) Use the power spectrum peak feature points and the spectral passband feature values to construct the relationship feature;

[0042] Step 4.3.1) Constructing power spectrum relationship features using power spectrum peak feature points PSD1-PSD6;

[0043]

[0044]

[0045] ∑θ n = 1

[0046] wherein θ n is the regression coefficient, which is fitted from the self-built data set;

[0047] Step 4.3.2) Constructing frequency spectrum passband relationship features FF n using frequency spectrum passband feature values P1, P2, P3;

[0048] FF n = θ1×P1+ θ2×P2+ θ3×P3

[0049] θ1+ θ2+ θ3= 1

[0050] wherein θ n is the regression coefficient, which is fitted from the self-built data set.

[0051] Further preferably, step 5) specifically comprises:

[0052] Step 5.1) Feature extraction on DC(t) in the venous emptying phase, including start point S, end point E, mid-point M feature points, and calculating dynamic baseline DBL and static baseline SBL, etc. feature values; the average value of the direct current component of the lower limb pulse wave signal 10 seconds before the start of dynamic measurement is defined as the dynamic baseline DBL, and the average value 30 seconds after the end of dynamic measurement represents the static baseline SBL;

[0053] Step 5.2) Defining 6 feature parameters of the direct current component of the lower limb pulse wave signal; V1 is the change value of the venous pump volume between DBL and SBL, V2 is the change value of the venous pump volume between S and SBL, V3 is the change value of the venous pump volume between E and DBL, K1 is the slope of the blood return curve from S point to E point, K2 is the slope of the blood return curve from S point to M point, and K3 is the slope of the blood return curve from M to E point.

[0054] Further preferably, step 6) specifically comprises:

[0055] Step 6.1) According to the separated AC IR (t), AC R (t), AC Y (t), AC GThe tissue transmission ability of (t) is marked in turn from shallow to deep as 4 different types of tissue pulsatile components; the infrared light contains the pulsatile components of arteries, arterioles and capillaries, the red light and yellow light contain different degrees of arteriole and capillary pulsatile components, and the green light contains capillary components;

[0056] Step 6.2) Arterial pulse wave reconstruction; using an improved difference method based on Beer-Lambert law to remove AC IR (t) from (t) to obtain the pulsatile component of (t); R (t) from (t) to obtain the pulsatile component of (t); Y (t) from (t) to obtain the pulsatile component of (t);

[0057] Step 6.3) Extraction of relevant features of venous vascular resistance VVR; TD RES is the time difference between the peak values of the reconstructed arterial pulse wave and the capillary pulse wave; TD IR is the time difference between the peak values of (t) and AC IR (t); the time constant t comes from AC G (t); the time constant t comes from AC IR (t); the heart rate HR is calculated from the peak-to-peak interval PPI of the AC G (t) with the shortest wavelength.

[0058] Further preferably, step 7) specifically comprises:

[0059] Step 7.1.1) Feature processing of the collected information feature dataset, AC component time domain features, AC component frequency domain features, DC component features and venous vascular resistance features using a bidirectional long short-term memory network BiLSTM method;

[0060] Step 7.1.1.1) Construct the collected information feature dataset, AC component time domain features, AC component frequency domain features, DC component features and venous vascular resistance features into an n-dimensional feature vector containing x groups, and input them into BiLSTM to obtain feature vectors A and A' in the front and back directions;

[0061] Step 7.1.1.2) Calculate the feature weights and sort the contribution degrees of different features; input A and A' into the feature attention model and normalize them into feature weights by the softmax function;

[0062] Step 7.1.1.3) Weighted sum of feature weights to obtain feature vector B; input the feature vector B into BiLSTM to obtain the long-term dependence relationship in the sequence data and output the hidden state S; integrate the hidden state S through the full connection layer to complete the feature extraction of multiple physiological signals;

[0063] Step 7.1.2) Dimensionality reduction of physiological signal features using principal component analysis PCA method;

[0064] Step 7.1.2.1) Centralize the feature data set to C and solve the covariance matrix C cov ;

[0065] Step 7.1.2.2) Calculate the eigen vector of C cov , and arrange them in order of size, and finally extract the effective features of the physiological signal feature data set;

[0066] Step 7.2) Classification and prediction of lower limb blood flow state; construct a lower limb blood flow state classifier; predict the lower limb blood flow state data of the user through the classifier, obtain the lower limb blood flow state label of the user, and use it to evaluate the blood flow state;

[0067] Step 7.2.1) Directly connect the features obtained in step 7.1) to realize feature fusion, use the support vector machine SVM method with linear kernel function to classify the fused features, construct a threshold classifier to obtain the classification label; construct a self-built lower limb blood flow state data label set;

[0068] Step 7.2.2) Then use the threshold classifier to predict the lower limb blood flow state data of the user, evaluate the blood flow state, and obtain the lower limb blood flow state label of the user;

[0069] Step 7.3) Risk area identification; compare the classification results obtained in step 7.2) with the acquisition area and acquisition pressure, construct a risk area decision tree, and predict the possible diseased area;

[0070] Step 7.3.1) Use the ankle acquisition area, knee acquisition area, proximal end pressurization area, distal end pressurization area, and self-built lower limb blood flow state data label set obtained in step 7.2) to jointly construct a lower limb deep vein thrombosis risk area decision tree;

[0071] Step 7.3.2) Compare the lower limb blood flow state label of the user with the risk area decision tree to obtain the risk area identification result;

[0072] Step 7.4) Risk level evaluation; construct a risk evaluation score S, compare the lower limb blood flow state label of the user with the trained feature result S, and according to the matching degree, obtain the preliminary risk evaluation score according to the low, medium and high evaluation levels;

[0073] Step 7.4.1) Use the acquisition area, pressurization area and blood flow state label to construct a lower limb deep vein thrombosis risk evaluation score S:

[0074] S = λ * MA + β * PA + θ * SL

[0075] Wherein, S is the lower limb deep vein thrombosis risk assessment score; MA is the collection area and the lower limb deep vein thrombosis risk correlation score; PA is the compression area and the lower limb deep vein thrombosis risk correlation score; SL is the blood flow state score; Lambda, beta, theta are weights obtained according to the relative importance of MA, PA and SL, respectively;

[0076] Step 7.4.2) compares the lower limb deep vein thrombosis risk assessment score obtained in step 7.5.2) with the set reference risk score, and if the score exceeds the set reference disease prediction standard, the user is at risk of having the disease, and outputs three assessment levels of low, medium and high according to the matching degree.

[0077] Compared with the prior art, the advantages of the present application are that:

[0078] The deep vein thrombosis continuous monitoring system based on the multi-wavelength photoelectric sensor provided by the present application has small device volume, low power consumption, can realize continuous monitoring, and has high user comfort degree; the intelligent signal processing module provided by the present application can estimate the venous blood flow state without image analysis, reduces the burden of patients and hospitals, and ensures the miniaturization and intelligentization of the peripheral blood pressure monitoring device; the early risk assessment module provided by the present application can judge the lower limb state, possible diseased area and risk level of the user, and realizes early warning of deep vein thrombosis. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 The structural block diagram of the deep vein thrombosis continuous monitoring system based on the multi-wavelength photoelectric sensor provided by the present application is shown in the figure;

[0080] Figure 2 The structural schematic diagram of the lower limb blood flow signal collection module provided by the present application is shown in the figure;

[0081] Figure 3 The carrying position map of the multi-wavelength photoelectric signal collection module provided by the present application is shown in the figure;

[0082] Figure 4 The compression position schematic diagram of the cuff module provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0083] The specific embodiments of the present application will be further described in detail below in combination with the drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application.

[0084] In the present application, the deep vein thrombosis continuous monitoring system based on the multi-wavelength photoelectric sensor is used for continuous monitoring and disease risk assessment of deep vein thrombosis, such as Figure 1As shown, it comprises a lower limb blood flow signal acquisition module, an intelligent signal processing module and an early risk assessment module; the lower limb blood flow signal acquisition module is used for continuously collecting blood flow state signals of a predetermined area of the lower limbs of a user and transmitting them to the intelligent signal processing module; the intelligent signal processing module can preprocess the lower limb blood flow signals and extract the change characteristics of the blood volume and vascular resistance of the user; the early risk assessment module can analyze the state of the lower limb venous vessels based on the change characteristics of the blood volume and vascular resistance of the user, realize the identification of the state of the lower limb veins, the early warning of abnormal areas and the assessment of the risk level of deep vein thrombosis.

[0085] In the present application, the specific structure of the lower limb blood flow signal acquisition module is as shown in Figure 2 As shown, it comprises a multi-wavelength photoelectric signal acquisition module 2, a cuff module 3 and a main control module 1. The multi-wavelength photoelectric signal acquisition module 2 comprises a flexible PCB board 26 placed on a magic tape 27. The multi-wavelength photoelectric sensor comprises four LED light emitters, i.e. an infrared LED 21, a red LED 23, a yellow LED 22 and a green LED 24, and a 3*3cm PD light receiver 25 for receiving blood flow state signals; the cuff module 3 comprises an inflatable cuff and a cuff pressure control module 31; the cuff pressure control module 31 can control the inflation and deflation of the inflatable cuff to a preset pressure through a motor-controlled valve, so as to actively hinder venous return and realize the pressurization of the lower limbs to dynamically measure the pulse wave signals of the lower limbs under different pressures. The main control module 1 comprises a multi-channel signal transmission interface 13, an inflatable cuff interface 12, a main control MCU 11, a TF storage unit 17, a power supply 16 and a Type-C charging transmission interface 18; the multi-channel signal transmission interface 13 is used for receiving the data of the multi-wavelength photoelectric signal acquisition module; the inflatable cuff interface 12 is used for connecting the cuff pressure control module 31; the main control MCU 11 can set the working mode of the multi-wavelength photoelectric sensor, the preset value of the cuff pressure and the acquisition time according to the instructions of the upper computer; the TF storage unit 17 is used for locally storing the collected signals; the power supply 16 comprises one rechargeable lithium ion battery; the Type-C charging data interface 18 is used for charging the power supply battery, connecting the upper computer and transmitting data. The lower limb blood flow signal acquisition module can work in two modes, i.e. static measurement and dynamic measurement; the static measurement is to continuously monitor the selected area for a long time when the inflatable cuff is not worn or not pressurized; the dynamic measurement is to pressurize the lower limbs by implementing a preset inflation and deflation method of the inflatable cuff, artificially intervene the venous return and realize the dynamic measurement of the pulse wave signals of the lower limbs under different pressures.

[0086] In the present application, the intelligent signal processing module comprises a signal preprocessing module, an alternating component time domain feature extraction module, an alternating component frequency domain feature extraction module, a direct current component feature extraction module and a venous vessel resistance feature extraction module. The signal preprocessing module can perform noise reduction and smoothing processing on the lower limb pulse wave signal collected by the multi-wavelength photoelectric signal acquisition module 2, including removing baseline drift, power frequency interference and high frequency interference, separating the direct current component (DC) and alternating current component (AC) of the lower limb pulse wave signal, and demultiplexing the four source signals of the alternating current component; the alternating component time domain feature extraction module can extract the time domain features of the alternating current component AC; the alternating component frequency domain feature extraction module can extract the frequency domain features of the alternating current component AC; the direct current component feature extraction module can extract the features of the direct current component DC; and the venous vessel resistance feature extraction module can calculate the venous vessel resistance using the four source signals.

[0087] In the present embodiment, the early risk assessment module comprises a feature processing and dimension reduction module, a lower limb blood flow state evaluation module, a risk area identification module and a risk level assessment module. The feature processing and dimension reduction module processes and reduces the dimension of the features, the lower limb blood flow state evaluation module can evaluate and warn the degree of lower limb venous vessel obstruction through feature extraction during continuous monitoring; the risk area identification module can predict the possible parts of DVT based on the diagnosis position and classification; and the risk level assessment module can determine the DVT risk level based on the blood flow state warning module and the risk area assessment module.

[0088] The specific method for monitoring and warning deep vein thrombosis by using the deep vein thrombosis continuous monitoring system based on a multi-wavelength photoelectric sensor provided by the present application is as follows:

[0089] 1) The user wears the multi-wavelength photoelectric signal acquisition module 2 on the monitoring area, collects blood flow state data, and wears the cuff module 3 on the pressurized area to realize the pressurization of the pressurized area and the dynamic measurement of the lower limb pulse wave signal under different pressure conditions, and the monitoring area data, the pressurized area data, the blood flow state data and the lower limb pulse wave data form the acquisition information feature data set; specifically:

[0090] Step 1.1) Use magic tape 27 to fix the multi-wavelength photoelectric signal acquisition module 2 on the medial malleolus, the lateral malleolus and the medial knee, and the predetermined collection area is as shown in Figure 3 1101, 1102 and 1103; and the inflatable cuff is worn on the proximal end of the lower limb and the distal end of the lower limb, and the predetermined pressurized area is as shown in Figure 4 1104 and 1105;

[0091] Step 1.2) Adjust the air bag pressure to the preset pressure P cuff, the preset pressure is set by the master control module 1; the LEDs in the multi-wavelength photoelectric signal acquisition module 2 flash in turn according to the preset time length, and appropriate full-bright intervals are left; then the lower limb pulse wave signals are continuously collected, and the collected signals are transmitted to the intelligent signal processing system;

[0092] Step 1.3) forming the acquisition information feature data set by using the user's to-be-monitored region data, to-be-pressurized region data, blood flow state data and lower limb pulse wave data;

[0093] 2) the signal preprocessing module performs noise reduction and smoothing processing on the lower limb pulse wave data, separates the direct current component DC(t) and the alternating current component AC(t) of the lower limb pulse wave signal, demultiplexes the four source signals of the alternating current component to obtain AC IR (t), AC R (t), AC Y (t) and AC G (t); specifically,

[0094] Step 2.1) the noise reduction and smoothing processing includes baseline drift removal, power frequency interference and high frequency interference, and signal reconstruction; first, a difference filtering method is used to identify and mark the peak and valley values of the lower limb pulse wave signal, and Savitzky-Golay filtering is used to calculate the polynomial fitting parameters of each peak-valley interval sequence for signal smoothing processing; then, a synchrosqueezing wavelet transform (SWT) is used to process the signal, decompose the signal into wavelet coefficients of different scales corresponding to different frequency components and time characteristics, sort the wavelet coefficients, identify and retain the most important wavelet coefficients, and discard other smaller coefficients to reduce noise; finally, the selected wavelet coefficients are used for signal reconstruction;

[0095] Step 2.2) the AC / DC separation of the lower limb pulse wave signal is to separate the vibration component of the pulse wave, i.e. the alternating current component, from the venous blood flow, i.e. the direct current component; the direct current component is extracted by using a "peak-middle-valley" three-point interpolation method to smooth the lower limb pulse wave signal, using the average of the peak and valley values as the middle value point, and using the three consecutive middle value points to construct a PPG interpolation signal through a second-order polynomial function, and the interpolation signal is the direct current component DC(t) of the pulse wave; the alternating current component is extracted by subtracting the direct current component from the original signal, and the alternating current component AC(t) is obtained;

[0096] Step 2.3) the alternating current component AC(t) is separated into AC IR (t), AC R (t), AC Y (t), ACG (t);

[0097] 3) The AC component time domain feature extraction module extracts features from the AC(t) of the preprocessed lower limb pulse wave data in the time domain, obtaining feature points of the original pulse wave, feature points of the first-order pulse wave, feature points of the second-order pulse wave, and feature points of the third-order pulse wave; specifically including:

[0098] Step 3.1) The feature points of the original pulse wave, using the local maximum and minimum method to obtain the onset point value onset, the systolic peak value sys, the dicrotic notch dic, the diastolic peak value dia, and the end point value end;

[0099] Step 3.2) The feature points of the first-order pulse wave, using the local maximum to obtain the maximum slope point ms;

[0100] Step 3.3) The feature points of the second-order pulse wave, using a zero-phase second-order Butterworth filter with a passband of 0.5-15Hz to eliminate the components of the frequency band that do not contribute; setting the negative part of the filtered second-order pulse wave Z[n] to 0, and obtaining y(n) = Z 2 [n], which is the squared second-order pulse wave; using two moving average formulas to generate the target region:

[0101]

[0102]

[0103] Wherein, W1 represents the systolic peak duration, W2 represents the average size of each cycle of the pulse wave, and the size of MA peak [n] is compared with MA beak [n] to remove the invalid region; finally, the threshold method is used to obtain the effective region, and the region maximum absolute value method is used to detect the points a, c, d, and e; the point b is the first minimum value after the point a, and the local minimum value method is used to detect the point b, wherein the point a is the first positive maximum value point of the systolic phase of the second-order pulse wave; the point b is the first negative maximum value point of the systolic phase of the second-order pulse wave; the point c is the second maximum value point of the systolic phase of the second-order pulse wave; the point d is the second minimum value point of the systolic phase of the second-order pulse wave; the point e is the first maximum value point of the diastolic phase of the second-order pulse wave;

[0104] Step 3.4) The feature points of the third-order pulse wave, using the local maximum method to search for the first local maximum p1 of the third derivative after the point b, and using the local minimum method to search for the last local minimum p2 of the third derivative before the point d.

[0105] 4) the AC component frequency domain feature extraction module extracts the features of the AC component AC(t) based on frequency domain analysis, including power spectrum relationship features and frequency spectrum passband relationship features; specifically including:

[0106] Step 4.1) extraction of power spectrum peak feature points, Welch algorithm is used to calculate the power spectrum density of the AC component AC(t) of the lower limb pulse wave signal;

[0107] Step 4.1.1) zero-mean AC(t) to make the center at zero; segment AC(t) with 50% overlap between segments; take the first 8 segments, smooth them with Hamming window, and calculate the corresponding periodogram of each segment; average all periodograms to obtain the power spectrum density of AC(t);

[0108] Step 4.1.2) use the local maximum value method to extract the first 6 more obvious power spectrum peak points PSD1-PSD6;

[0109] Step 4.2) extraction of frequency spectrum passband feature values, using the characteristic frequency band framework method to calculate the frequency spectrum energy;

[0110] Step 4.2.1) set the fingertip lower limb pulse wave signal Tip(t) as the reference signal, use fast Fourier transform and filter through S-G filter to obtain the smoothed reference signal spectrum FTip(ω);

[0111] Step 4.2.2) use the reference signal to extract three groups of central characteristic frequencies to generate the reference signal spectrum framework; set FP1 as the frequency of the peak point of FTip(ω) in the interval [0.8Hz, 1.5Hz], FP2 as the frequency of the peak point of FTip(ω) in the interval [2×fp1-0.5Hz, 2×fp1+0.5Hz], and FP3 as the frequency of the peak point of FTip(ω) in the interval [3×fp1-0.5Hz, 3×fp1+0.5Hz]; according to the -3db bandwidth principle, determine the upper and lower limit cutoff frequencies of the adjacent frequency domain of FP1, FP2 and FP3, and obtain three reference signal spectrum passband feature intervals

[0112] Step 4.2.3) use the reference signal spectrum passband feature interval to obtain three frequency spectrum passband feature values of the AC component; respectively extract the feature values in the frequency spectrum passband feature interval obtain P1, P2, P3;

[0113] Step 4.3) use the power spectrum peak feature points and the frequency spectrum passband feature values to construct the relationship features;

[0114] Step 4.3.1) use the power spectrum peak feature points PSD1-PSD6 to construct the power spectrum relationship features;

[0115]

[0116]

[0117] ∑θ n =1

[0118] Where, θ n These are the regression coefficients, obtained by fitting from a self-built dataset;

[0119] Step 4.3.2) Construct the spectral passband relationship feature FF using the spectral passband feature values ​​P1, P2, and P3. n ;

[0120] FF n =θ1×P1+θ2×P2+θ3×P3

[0121] θ1+θ2+θ3=1

[0122] Where, θ n The regression coefficients are obtained by fitting from a self-built dataset.

[0123] 5) The DC component feature extraction module extracts features from the DC component DC(t) of the preprocessed lower limb pulse wave data under dynamic measurement; specifically including:

[0124] Step 5.1) Extract features from DC(t) during the venous emptying phase, including the start point S, end point E, and measurement midpoint M, and calculate feature values ​​such as dynamic baseline DBL and static baseline SBL; the average value of the DC component of the lower limb pulse wave signal 10 seconds before the start of dynamic measurement is defined as dynamic baseline DBL, and the average value 30 seconds after the end of dynamic measurement is defined as static baseline SBL.

[0125] Step 5.2) Define six characteristic parameters for the DC component of the lower limb pulse wave signal: V1 is the change in venous pump volume between DBL and SBL, V2 is the change in venous pump volume between S and SBL, V3 is the change in venous pump volume between E and DBL, K1 is the slope of the blood return curve from point S to point E, K2 is the slope of the blood return curve from point S to point M, and K3 is the slope of the blood return curve from point M to point E.

[0126] 6) The venous vascular resistance feature extraction module extracts the AC obtained in step 2). IR (t), AC R (t), AC Y (t), AC G (t) Extraction of venous vascular resistance features; specifically including:

[0127] Step 6.1) Based on the separated ACIR (t), AC R (t), AC Y (t), AC G (t) are marked in turn from shallow to deep as 4 different types of tissue pulsatile components; the infrared light contains the pulsatile components of arteries, arterioles and capillaries, the red light and yellow light contain different degrees of arteriolar and capillary pulsatile components, and the green light contains capillary components;

[0128] Step 6.2) Arterial pulse wave reconstruction; using the improved difference method based on Beer-Lambert law to remove AC IR (t) from AC R (t), AC Y (t) of the pulsatile component;

[0129] Step 6.3) Extraction of relevant features of venous vascular resistance VVR; TD RES is the time difference between the peak values of the reconstructed arterial pulse wave and capillary pulse wave; TD IR is the time difference between the peak values of AC IR (t) and AC G (t) peak; the time constant t comes from AC IR (t); the heart rate HR is calculated from the peak-to-peak interval PPI of the AC G (t) with the shortest wavelength.

[0130] 7) The early risk assessment module analyzes the features obtained in steps 3), 4), 5), 6), and evaluates the lower limb blood flow state, risk level and risk area:

[0131] 7.1) The feature processing and dimension reduction module processes and reduces the dimension of the features obtained in steps 3), 4), 5), 6); specifically including:

[0132] Step 7.1.1) Use the bidirectional long short-term memory network BiLSTM method to process the feature data set of the collected information, the time domain features of the alternating component, the frequency domain features of the alternating component, the DC component features and the venous vascular resistance features;

[0133] Step 7.1.1.1) Construct the feature data set of the collected information, the time domain features of the alternating component, the frequency domain features of the alternating component, the DC component features and the venous vascular resistance features into an n-dimensional feature vector containing x groups, and input BiLSTM to obtain the feature vectors A and A' in the front and back directions;

[0134] Step 7.1.1.2) Calculate the feature weight and sort the contribution degree of different features; input A and A' into the feature attention model, and normalize them into feature weights by the softmax function;

[0135] Step 7.1.1.3) Weighted sum of feature weights to obtain feature vector B; input the feature vector B into BiLSTM to obtain long-term dependencies in sequence data, output hidden state S; integrate the hidden state S through the full connection layer to complete the feature extraction of multiple physiological signals;

[0136] Step 7.1.2) Use principal component analysis (PCA) method to reduce the dimension of physiological signal features;

[0137] Step 7.1.2.1) Center the feature data set to C, and solve the covariance matrix C cov ;

[0138] Step 7.1.2.2) Calculate the eigenvectors of C cov , and arrange them in order of size, and finally extract the effective features of the physiological signal feature data set;

[0139] 7.2) The lower limb blood flow state evaluation module constructs a lower limb blood flow state classifier according to the blood flow state data, predicts the user's lower limb blood flow state data through the classifier, obtains the user's lower limb blood flow state label, and is used for evaluating the blood flow state; specifically including

[0140] Step 7.2.1) Directly connect the features obtained in step 7.1) to realize feature fusion, use the support vector machine (SVM) method with a linear kernel function to classify the fused features, construct a threshold classifier, and obtain a classification label; construct a self-built lower limb blood flow state data label set;

[0141] Step 7.2.2) Then use the threshold classifier to machine learning prediction of the user's lower limb blood flow state data, evaluate the blood flow state, and obtain the user's lower limb blood flow state label;

[0142] 7.3) Compare the classification results obtained in step 7.2) with the collection area and collection pressure to construct a risk area decision tree and predict the area that may be diseased;

[0143] Step 7.3.1) Use the ankle collection area, knee collection area, proximal end compression area, distal end compression area, and self-built lower limb blood flow state data label set obtained in step 7.2) to jointly construct a lower limb deep vein thrombosis risk area decision tree;

[0144] Step 7.3.2) Compare the user's lower limb blood flow state label with the risk area decision tree to obtain a risk area identification result;

[0145] 7.4) The risk level assessment module builds a risk assessment score S, and compares the user's lower limb blood flow state label with the trained feature result S, and according to the matching degree, the preliminary risk assessment score is obtained in three assessment levels of low, medium and high;

[0146] Step 7.4.1) Use the acquisition area, compression area and blood flow state label to build the lower limb deep vein thrombosis risk assessment score S:

[0147] S = λ * MA + β * PA + θ * SL

[0148] Wherein, S is the lower limb deep vein thrombosis risk assessment score; MA is the acquisition area and the lower limb deep vein thrombosis risk correlation score; PA is the compression area and the lower limb deep vein thrombosis risk correlation score; SL is the blood flow state score; λ, β, θ are the weights obtained according to the relative importance of MA, PA and SL, respectively;

[0149] Step 7.4.2) Compare the lower limb deep vein thrombosis risk assessment score obtained in step 7.5.2) with the set reference risk score, if the score exceeds the set reference disease prediction standard, the user is at risk of having the disease, and according to the matching degree, three assessment levels of low, medium and high are output.

[0150] 8) According to the results of the early risk assessment module, an alarm is sent to the user.

[0151] Finally, it should be pointed out that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present application.

Claims

1. A continuous monitoring system for deep vein thrombosis based on multi-wavelength photoelectric sensors, characterized in that, The system includes a lower limb blood flow signal acquisition module, an intelligent signal processing module, and an early risk assessment module connected sequentially. The lower limb blood flow signal acquisition module includes a main control module (1), a multi-wavelength photoelectric signal acquisition module (2), and a cuff module (3). Both the multi-wavelength photoelectric signal acquisition module (2) and the cuff module (3) are connected to the main control module (1). The photoelectric signals acquired by the multi-wavelength photoelectric signal acquisition module (2) include infrared light, red light, yellow light, and green light signals. The intelligent signal processing module includes a signal preprocessing module and an AC component time-domain feature extraction module and an AC component frequency-domain feature extraction module, which are respectively connected to the signal preprocessing module. The module includes a block, a DC component feature extraction module, and a venous blood flow resistance feature extraction module. The main control module (1) is connected to the signal preprocessing module. The early risk assessment module includes a feature processing dimensionality reduction module, a lower limb blood flow status evaluation module, a risk area identification module, and a risk level assessment module. The AC component time domain feature extraction module, the AC component frequency domain feature extraction module, the DC component feature extraction module, and the venous blood flow resistance feature extraction module are all connected to the feature processing dimensionality reduction module. The feature processing dimensionality reduction module is connected to the lower limb blood flow status evaluation module. The lower limb blood flow status evaluation module is connected to the risk area identification module and the risk level assessment module, respectively.

2. The deep vein thrombosis continuous monitoring system based on a multi-wavelength photoelectric sensor according to claim 1, characterized in that, The main control module (1) includes a housing, inside which are a main control MCU (11), a power supply (16) and a storage unit (17). The surface of the housing is provided with an inflatable cuff interface (12), a multi-channel signal transmission interface (13) and a charging transmission interface (18). The main control MCU (11) is connected to the power supply (16), the storage unit (17), the inflatable cuff interface (12) and the multi-channel signal transmission interface (13), respectively. The charging transmission interface (18) is connected to the power supply (16). The main control MCU (11) is connected to the signal preprocessing module. The multi-wavelength photoelectric signal acquisition module (2) includes a flexible PCB board (26), a multi-wavelength photoelectric sensor is provided on the flexible PCB board (26), and a magic strip (27) is provided on both sides of the flexible PCB board (26). The multi-channel signal transmission interface (13) is connected to the multi-wavelength photoelectric sensor through a transmission line (15). The cuff module (3) includes an inflatable cuff and a cuff pressure control module (31). The inflatable cuff interface (12) is connected to the cuff pressure control module (31) through a tubular conduit (14), and the cuff pressure control module (31) is connected to the inflatable cuff.

3. The deep vein thrombosis continuous monitoring system based on a multi-wavelength photoelectric sensor according to claim 2, characterized in that, The multi-wavelength photoelectric sensor includes an infrared LED (21), a yellow LED (22), a red LED (23), a green LED (24), and a PD photodetector (25), all of which are connected to the main control MCU (11).

4. A method for continuous monitoring and early warning of deep vein thrombosis based on multi-wavelength photoelectric sensors, characterized in that, The deep vein thrombosis continuous monitoring system based on a multi-wavelength photoelectric sensor as described in claim 1 includes the following steps: 1) The user wears the multi-wavelength photoelectric signal acquisition module (2) on the area to be monitored to collect blood flow status data, and wears the cuff module (3) on the area to be pressurized to realize the pressurization of the area to be pressurized and the dynamic measurement of the lower limb pulse wave signal under different pressure conditions. The data of the area to be monitored, the data of the area to be pressurized, the blood flow status data, and the lower limb pulse wave data form the data acquisition information feature dataset. 2) The signal preprocessing module performs noise reduction and smoothing on the lower limb pulse wave data, separates the DC component DC(t) and AC component AC(t) of the lower limb pulse wave signal, and demultiplexes the four source signals of the AC component to obtain AC... IR (t), AC R (t), AC Y (t), AC G (t); 3) The AC component time-domain feature extraction module extracts features from the AC component AC(t) of the preprocessed lower limb pulse wave data in the time domain to obtain the feature points of the original pulse wave, the feature points of the first-order pulse wave, the feature points of the second-order pulse wave, and the feature points of the third-order pulse wave. 4) The AC component frequency domain feature extraction module extracts the AC component AC(t) based on frequency domain analysis features, including power spectrum relationship features and spectral passband relationship features; 5) The DC component feature extraction module extracts features from the DC component DC(t) of the preprocessed lower limb pulse wave data under dynamic measurement. 6) The venous vascular resistance feature extraction module extracts the AC obtained in step 2). IR (t), AC R (t), AC Y (t), AC G (t) Extract venous resistance features; 7) The early risk assessment module analyzes the features obtained in steps 3), 4), 5), and 6), and assesses the lower limb blood flow status, risk level, and risk area: 7.1) The feature processing and dimensionality reduction module processes and reduces the dimensionality of the features obtained in steps 3), 4), 5), and 6); 7.2) The lower limb blood flow status evaluation module constructs a lower limb blood flow status classifier based on blood flow status data. The lower limb blood flow status data of the user is predicted by the classifier to obtain the lower limb blood flow status label of the user, which is used to evaluate the blood flow status. 7.3) The risk area identification module compares the results of step 7.1) with the data of the area to be monitored and the data of the area to be pressurized to construct a risk area decision tree and predict and identify areas that may be affected by the disease. 7.4) The risk level assessment module constructs a risk assessment score S, compares the user's lower limb blood flow status label with the feature result S obtained from training, and obtains a preliminary risk assessment score according to the degree of matching at three assessment levels: low, medium, and high. 8) Issue alerts to users based on the results of the early risk assessment module.

5. The method for continuous monitoring and early warning of deep vein thrombosis based on a multi-wavelength photoelectric sensor according to claim 4, characterized in that, Step 3) specifically includes: Step 3.1) Obtain the following characteristic points from the original pulse wave: onset, systolic peak, dicrotic notch, diastolic peak, and end point using the local maximum / minimum method; Step 3.2) For the characteristic points of the first-order pulse wave, use local maxima to obtain the maximum slope point ms; Step 3.3) For the characteristic points of the second-order pulse wave, use a zero-phase second-order Butterworth filter with a bandpass of 0.5-15Hz to eliminate non-contributing frequency band components; set the negative part of the filtered second-order pulse wave Z[n] to 0, and obtain y(n) = Z 2 [n] represents the squared second-order pulse wave; the target region is generated using two moving average formulas: Where W1 represents the duration of the contraction peak, W2 represents the average size of each pulse wave cycle, and MA is compared. peak [n]with MA beak The size of [n] is used to remove invalid regions; finally, a threshold method is used to obtain the valid regions, and the maximum absolute value method of the regions is used to detect points a, c, d, and e; point b is the first minimum value after point a, and the local minimum method is used to detect point b. Here, point a is the first positive maximum point of the second-order pulse wave systole; point b is the first negative maximum point of the second-order pulse wave systole; point c is the second maximum point of the second-order pulse wave systole; point d is the second minimum point of the second-order pulse wave systole; and point e is the first maximum point of the second-order pulse wave diastole. Step 3.4) For the characteristic points of the third-order pulse wave, use the local maximum method to search for the first local maximum p1 of the third derivative after point b, and use the local minimum method to search for the last local minimum p2 of the third derivative before point d.

6. The method for continuous monitoring and early warning of deep vein thrombosis based on a multi-wavelength photoelectric sensor according to claim 4, characterized in that, Step 4) specifically includes: Step 4.1) Extraction of power spectrum peak feature points, and use Welch algorithm to calculate the power spectral density of AC(t) of lower limb pulse wave signal; Step 4.1.1) Zero-mean AC(t) is applied so that the center is at zero; AC(t) is segmented with 50% overlap between segments; the first 8 segments are taken, smoothed by Hamming window, and the corresponding periodogram of each segment is calculated; the power spectral density of AC(t) is obtained by averaging all periodograms. Step 4.1.2) Use the local maximum method to extract the first 6 more obvious power spectrum peaks PSD1-PSD6; Step 4.2) Extraction of spectral passband features, and calculation of spectral energy using the feature band framework method; Step 4.2.1) Set the fingertip lower limb pulse wave signal Tip(t) as the reference signal, use fast Fourier transform and filter it through SG filter to obtain the smooth reference signal spectrum FTip(ω); Step 4.2.2) Extract three sets of center characteristic frequencies using the reference signal to generate the reference signal spectrum framework; set FP1 as the frequency of the peak point of FTip(ω) in the interval [0.8Hz, 1.5Hz], FP2 as the frequency of the peak point of FTip(ω) in the interval [2×fp1-0.5Hz, 2×fp1+0.5Hz], and FP3 as the frequency of the peak point of FTip(ω) in the interval [3×fp1-0.5Hz, 3×fp1+0.5Hz]; based on the -3dB bandwidth principle, determine the upper and lower cutoff frequencies of the adjacent frequency domains of FP1, FP2, and FP3 to obtain the three reference signal spectrum passband characteristic intervals. Step 4.2.3) Obtain three spectral passband feature values ​​of the AC component using the reference signal spectral passband feature interval; extract the feature values ​​within the spectral passband feature interval respectively. Obtain P1, P2, and P3; Step 4.3) Construct relational features using power spectrum peak feature points and spectral passband feature values; Step 4.3.1) Construct power spectrum relationship features using power spectrum peak feature points PSD1-PSD6; ∑θ n =1 Where, θ n These are the regression coefficients, obtained by fitting from a self-built dataset; Step 4.3.2) Construct the spectral passband relationship feature FF using the spectral passband feature values ​​P1, P2, and P3. n ; FF n =θ1×P1+θ2×P2+θ3×P3 θ1+θ2+θ3=1 Where, θ n The regression coefficients are obtained by fitting from a self-built dataset.

7. The method for continuous monitoring and early warning of deep vein thrombosis based on a multi-wavelength photoelectric sensor according to claim 4, characterized in that, Step 5) specifically includes: Step 5.1) Extract features from DC(t) during the venous emptying phase, including the start point S, end point E, and measurement midpoint M, and calculate feature values ​​such as dynamic baseline DBL and static baseline SBL; the average value of the DC component of the lower limb pulse wave signal 10 seconds before the start of dynamic measurement is defined as dynamic baseline DBL, and the average value 30 seconds after the end of dynamic measurement is defined as static baseline SBL. Step 5.2) Define six characteristic parameters for the DC component of the lower limb pulse wave signal: V1 is the change in venous pump volume between DBL and SBL, V2 is the change in venous pump volume between S and SBL, V3 is the change in venous pump volume between E and DBL, K1 is the slope of the blood return curve from point S to point E, K2 is the slope of the blood return curve from point S to point M, and K3 is the slope of the blood return curve from point M to point E.

8. The method for continuous monitoring and early warning of deep vein thrombosis based on a multi-wavelength photoelectric sensor according to claim 4, characterized in that, Step 6) specifically includes: Step 6.1) Based on the separated AC IR (t), AC R (t), AC Y (t), AC G The tissue transmittance of (t) is labeled as four different types of tissue pulsation components from light to dark; infrared light contains the pulsation components of arteries, arterioles and capillaries, red light and yellow light contain arterioles and capillaries pulsation components of varying degrees, and green light contains capillary components. Step 6.2) Arterial pulse wave reconstruction; using an improved differential method based on Beer-Lambert's law to reconstruct the AC pulse wave. IR Remove AC from (t) R (t), AC Y The pulsating component of (t); Step 6.3) Extraction of relevant features of venous vascular resistance (VVR); TD RES It is the time difference between the peak values ​​of the reconstructed arterial pulse wave and the capillary pulse wave; TD IR It is AC IR The peak value of (t) and AC G (t) The time difference between peak values; the time constant t comes from AC. IR (t); Heart rate (HR) is determined by the shortest wavelength AC. G Calculation of interpeak PPI for (t).

9. The method for continuous monitoring and early warning of deep vein thrombosis based on a multi-wavelength photoelectric sensor according to claim 4, characterized in that, Step 7) specifically includes: Step 7.1.1) Use the Bi-directional Long Short-Term Memory (BiLSTM) network method to perform feature processing on the collected information feature dataset, AC component time-domain features, AC component frequency-domain features, DC component features, and venous vascular resistance features; Step 7.1.1.1) Construct an n-dimensional feature vector containing x groups from the collected information feature dataset, AC component time-domain features, AC component frequency-domain features, DC component features, and venous resistance features, and input it into BiLSTM to obtain feature vectors A and A' in the front and back directions; Step 7.1.1.2) Calculate the feature weights and rank the contributions of different features; input A and A' into the feature attention model and normalize them into feature weights using the softmax function; Step 7.1.1.3) The feature weights are weighted and summed to obtain the feature vector B; the feature vector B is input into BiLSTM to obtain the long-term dependencies in the sequence data and output the hidden state S; the hidden state S is integrated through a fully connected layer to complete the feature extraction of multiple physiological signals. Step 7.1.2) Use Principal Component Analysis (PCA) to reduce the dimensionality of physiological signal features; Step 7.1.2.1) Center the feature dataset to C and solve for the covariance matrix C. cov ; Step 7.1.2.2) Calculate C cov The feature vectors are sorted in order of size, and finally the effective features of the physiological signal feature dataset are extracted. Step 7.2) Classification and prediction of lower limb blood flow status; construct a lower limb blood flow status classifier; predict the lower limb blood flow status data of the user through the classifier to obtain the lower limb blood flow status label of the user, which is used to evaluate the blood flow status; Step 7.2.1) Directly connect the features obtained in Step 7.1) to achieve feature fusion. Use the Support Vector Machine (SVM) method with linear kernel function to classify the fused features and construct a threshold classifier to obtain classification labels; construct a self-built lower limb blood flow status data label set. Step 7.2.2) Then, a threshold classifier is used to perform machine learning prediction on the user's lower limb blood flow status data to evaluate the blood flow status and obtain the user's lower limb blood flow status label. Step 7.3) Risk area identification; compare the classification results obtained in step 7.2) with the collection area and collection pressure, construct a risk area decision tree, and predict areas that may be affected by the disease; Step 7.3.1) Use the ankle acquisition area, knee acquisition area, proximal compression area, distal compression area and the self-built lower limb blood flow status data label set obtained in step 7.2) to jointly construct a decision tree for the risk area of ​​lower limb deep vein thrombosis; Step 7.3.2) Compare the user's lower limb blood flow status labels with the risk area decision tree to obtain the risk area identification results; Step 7.4) Risk level assessment; Construct a risk assessment score S, compare the user's lower limb blood flow status label with the feature results S obtained from training, and obtain a preliminary risk assessment score according to the degree of matching, classified into three assessment levels: low, medium, and high. Step 7.4.1) Construct a lower extremity deep vein thrombosis risk assessment score S using the collection area, pressure area, and blood flow status labels: S=λ*MA+β*PA+θ*SL Wherein, S is the risk assessment score for deep vein thrombosis in the lower extremities; MA is the correlation score between the collection area and the risk of deep vein thrombosis in the lower extremities; PA is the correlation score between the compression area and the risk of deep vein thrombosis in the lower extremities; SL is the blood flow status score; λ, β, and θ are the weights obtained according to the relative importance of MA, PA, and SL, respectively. Step 7.4.2) Compare the lower extremity deep vein thrombosis risk assessment score obtained in step 7.5.2) with the set reference risk score. If the score exceeds the set reference disease prediction standard, the user is at risk of having the disease, and output three assessment levels: low, medium, and high according to the degree of matching.

Images