Noninvasive femoral artery blood flow estimation method and device based on characteristic difference

Abstract

The invention discloses a non-invasive femoral artery blood flow estimation method based on characteristic difference, which comprises the following steps of: acquiring PPG signal data of a finger tip and an instep of a subject, and performing baseline drift correction, noise reduction and zero value expansion preprocessing on the PPG signal data of the finger tip and the instep of the subject to obtain the PPG signal data of the finger tip and the instep of the subject; acquiring the preprocessed PPG signal data of the finger tip and the instep; according to the preprocessed PPG signal data of the finger tip and the instep, based on a pre-trained femoral artery blood flow estimation model, femoral artery blood flow estimation processing is conducted, and an estimation result of the femoral artery blood flow is obtained. The invention discloses non-invasive femoral artery blood flow estimation based on characteristic differences. The femur artery blood flow is evaluated by analyzing the characteristic difference of the photoelectric volume pulse wave signals obtained by the finger tip and the instep. By utilizing a deep learning algorithm, physiological signals which are easy to collect can be effectively utilized, and innovation is brought to treatment of lower limb artery diseases.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and device for non-invasive femoral artery blood flow estimation based on feature differences. Background Art

[0002] Peripheral Artery Disease (PAD) is the main cause of disability and loss of mobility in the elderly, especially atherosclerotic disease of the lower limbs. The main causes of lower limb PAD are ischemia and vascular damage, which may also be accompanied by intermittent claudication, and eventually tissue necrosis. Severe cases require amputation, which seriously affects the patient's quality of life. In recent years, supervised exercise rehabilitation training has been applied to the treatment of PAD as a non-invasive, non-drug, and non-interventional method, improving patients' walking ability.

[0003] However, since the intensity of exercise is difficult to control, there is much controversy about the improvement of lower limb vascular function in patients, and high exercise intensity may aggravate the patient's claudication. Therefore, it is of great clinical significance to seek new non-invasive alternative treatment methods to improve the lower limb vascular function of PAD patients. Passive exercise rehabilitation enhanced external counterpulsation (EECP) originated from China. It is a non-invasive auxiliary circulation device. Under the synchronous triggering of the ECG R wave, the air bag wrapped around the calf, thigh and buttocks is inflated and pressurized from bottom to top during the diastole of the heart, driving the blood to form a double pulse and perfuse the whole body, accelerating the blood flow velocity and increasing the blood flow shear stress, which has an effect similar to exercise. At present, in clinical practice, ultrasound technology is used to monitor the changes in hemodynamic indicators of lower limb arteries to evaluate the immediate efficacy of EECP. However, this method has the following defects: (1) Ultrasound is relatively expensive, which increases clinical costs. (2) After obtaining the blood flow spectrum curve through ultrasound, data processing and calculation are required to obtain the blood flow, which increases the complexity and time cost of the operation. Summary of the invention

[0004] The present invention overcomes the shortcomings of the prior art and provides a non-invasive femoral artery blood flow estimation based on feature differences; by analyzing the feature differences of the photoplethysmographic pulse wave signals obtained from the fingertips and the dorsum of the foot, the femoral artery blood flow can be accurately assessed. The deep learning algorithm can effectively utilize the physiological signals that are easy to collect, bringing innovation to the treatment of lower extremity arterial diseases.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a non-invasive femoral artery blood flow estimation method based on feature difference, comprising the following steps: Acquire PPG signal data of the fingertips and dorsum of the subject, and perform baseline drift correction, noise reduction, and zero value extension preprocessing on the PPG signal data of the fingertips and dorsum of the foot to obtain the preprocessed PPG signal data of the fingertips and dorsum of the foot; According to the pre-processed PPG signal data of the fingertips and the dorsum of the foot, based on the pre-trained femoral artery blood flow estimation model, a femoral artery blood flow estimation process is performed to obtain an estimation result of the femoral artery blood flow; The construction of the femoral artery blood flow estimation model includes: constructing a femoral artery blood flow estimation model based on a convolutional neural network-long short-term memory neural network, and incorporating an attention mechanism into the femoral artery blood flow estimation model; Wherein, the training of the femoral artery blood flow estimation model includes: The femoral artery blood flow estimation model is trained using the training set in the data set to obtain the trained femoral artery blood flow estimation model; the signal sequence of the test set in the data set is input into the trained femoral artery blood flow estimation model, different evaluation indicators are calculated, and the estimation result of the femoral artery blood flow is obtained.

[0006] In a preferred embodiment of the present invention, baseline drift correction, noise reduction, and zero extension preprocessing include the following steps: Step a1, performing baseline drift correction and noise reduction preprocessing on the PPG signals collected from the fingertips and dorsum of the subject; Step a2, intercepting the preprocessed PPG signals of the fingertips and dorsum of the foot; Step a3, calculating the difference between the intercepted fingertip PPG and dorsum PPG signals, and downsampling the calculated signal sequence to reduce the amount of data; Step a4: performing a zero-value extension operation on each down-sampled signal segment.

[0007] In a preferred embodiment of the present invention, the lower limb artery blood flow spectrum is obtained by ultrasonic measurement, and the average blood flow is calculated by the lower limb artery blood flow spectrum as the target value for femoral artery blood flow estimation.

[0008] In a preferred embodiment of the present invention, in step a1, baseline drift correction and noise reduction preprocessing are performed on the PPG signals of the fingertips and the dorsum of the foot, and discrete wavelet transform is performed on the input signal using the wavedec function and wrcoef function in MATLAB. The basic formula of discrete wavelet transform is: ; in, are discrete wavelet coefficients, is the wavelet basis function; parameter j is the scale factor, which determines The shape of the scale factor includes width and amplitude, and k is the time shift factor, which determines Position along the horizontal axis, x represents the discrete input signal, and t represents the time along the horizontal axis; In step a2, each subject collected PPG signals of fingertips and dorsum of foot under 7 states, which are static state, counterpulsation pressure set to 0.030MPa without calf cuff, counterpulsation pressure set to 0.030MPa with calf cuff, counterpulsation pressure set to 0.035MPa, counterpulsation pressure set to 0.025MPa, counterpulsation pressure set to 0.020MPa, and recovery after completion. From the PPG signals of the subjects in different states, signal segments are intercepted according to the start and end time of the ultrasonic blood flow spectrum tracing points, and the length of the signal segment is determined by the number of cardiac cycles between the ultrasonic blood flow spectrum tracing points. The interception method is to first determine the conditions for finding the peak, including the minimum peak height and the minimum peak spacing, then invert the original PPG signal, use the find peaks function in MATLAB to find the main peak position of the pulse wave, as the position of the trough found, and finally intercept the corresponding segment according to the position of the trough to obtain the required data; In step a3, the difference between the intercepted fingertip PPG and dorsum PPG signals is calculated, and the calculated signal sequence is downsampled to reduce the amount of data. The specific method is to first subtract the two signal sequences to obtain a new sequence, and then calculate the size of each downsampling window according to the target sampling frequency. Each downsampling window is traversed on the new sequence, and the following operations are performed for each window: if there is a peak in the window, the peak is used as the representative value of the window; if there is no peak but there is a valley value, the valley value is used as the representative value of the window; if there is neither a peak nor a valley value in the window, the data in the window is averaged to obtain the representative value.

[0009] In step a4, in order to compensate for the inconsistency of the data length of the down-sampled signal segments caused by the dynamic change of the cardiac cycle over time, a zero-value extension operation is performed on each down-sampled signal segment to ensure that they are uniformly expanded to a preset fixed length, thereby achieving standardization and alignment in the dimension of time series data.

[0010] In a preferred embodiment of the present invention, obtaining the target value of femoral artery blood flow estimation includes the following steps: The average blood flow was calculated using the lower limb arterial blood flow spectrum obtained by ultrasound measurement of the subjects. First, the blood flow spectrum curve was extracted by image processing. The main steps included removing the yellow mark generated by the scanner integrated software, selecting ROI, converting the image to grayscale, closing operation for filling, Sobel filtering for edge extraction and envelope extraction. Then, the average blood flow in the time period between the tracing points was calculated based on the product of the integral of the points on the blood flow spectrum curve and the cross-sectional area of ​​the blood vessel, which was used as the target value for femoral artery blood flow estimation.

[0011] In a preferred embodiment of the present invention, a femoral artery blood flow estimation model based on a convolutional neural network-long short-term memory neural network is constructed, comprising: a convolutional neural network and a long short-term memory network; The convolutional neural network includes: two consecutive convolutional modules, each of which contains three one-dimensional convolutional layers, a jump connection is used in the third layer to achieve feature fusion, and a normalization layer and a Swish activation function are set after each convolutional layer. , down-sampling is performed through maximum pooling; where x' is the feature of the input signal sequence after convolution and standardization; The long short-term memory network comprises: Initialize the cell state and hidden state. The cell state is set to the zero vector to accumulate information in the sequence; the hidden state is initialized to the zero vector, representing the output state at the initial moment.

[0012] Specifically, the true value is the true value of the femoral artery blood flow, which is calculated by the hospital using ultrasound to collect the blood flow spectrum and vascular inner diameter during the EECP process.

[0013] In a preferred embodiment of the present invention, in the long short-term memory network, for each time step in the sequence, the long short-term memory network performs the following steps: Step b1, according to the input signal of the current time step and the hidden state of the previous time step, the input gate is calculated by the sigmoid function to obtain the input gate vector; Step b2, using the current input and the previous hidden state to perform forget gate calculation to obtain the forget gate vector; Step b3, generating a new vector by inputting candidate calculation; the input candidate calculation includes: candidate information that is desired to be added to the cell state; and is completed by a tanh activation function; Step b4, combining the results of the forget gate and the input gate to obtain the updated cell state by using the cell state update; Step b5, calculating the output gate through the sigmoid function in the output gate calculation, which is used to determine which information is important to the outside world; Step b6, calculate the new hidden state through hidden state update based on the current cell state and the value of the output gate.

[0014] In a preferred embodiment of the present invention, the attention mechanism includes two steps: Step c1, linearly transform and unbias the hidden state obtained by the long short-term memory network, and use the tanh function to help the femoral artery blood flow estimation model learn how to assign different weights according to the importance of the hidden state to obtain the attention weight; In step c2, the calculated attention weights are applied to the output of the LSTM network, and weighted summation is achieved through batch matrix multiplication. The attention weights are used to emphasize the time step information that is most useful for the task.

[0015] In a preferred embodiment of the present invention, 15% of the data set used to train the constructed femoral artery blood flow estimation model is allocated as a test set, and the remaining 85% of the data set is used as a training set; And / or, use AdamW optimizer to optimize data. The Epoch of AdamW optimizer is set to 100, the batch_size is set to 8, and the loss function of AdamW optimizer is set to mean square error MSE. The formula of mean square error MSE is: ;in, is the true value, is the predicted value.

[0016] In a preferred embodiment of the present invention, a non-invasive femoral artery blood flow estimation device based on characteristic difference comprises: Memory; processor; as well as Computer programs; The computer program is stored in the memory and is configured to be executed by the processor to implement a non-invasive femoral artery blood flow estimation method based on feature differences.

[0017] The present invention solves the defects existing in the background technology: The present invention discloses a non-invasive femoral artery blood flow estimation method based on feature differences, which accurately evaluates the femoral artery blood flow by analyzing the feature differences of the photoplethysmography (PPG) signals obtained from the fingertips and the dorsum of the foot. By using a deep learning algorithm, physiological signals that are easy to collect can be effectively utilized to bring innovation to the treatment of lower limb arterial diseases. In particular, during the implementation of enhanced external counterpulsation therapy, convenient and timely feedback is provided to doctors, assisting doctors in selecting the best treatment plan for patients with lower limb arterial diseases based on changes in femoral artery blood flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0019] FIG1 is a flow chart of non-invasive femoral artery blood flow estimation based on feature difference in a preferred embodiment of the present invention; FIG2 is a general structural diagram of the CNN-LSTM-Attention model constructed in a preferred embodiment of the present invention; FIG. 3 is a diagram showing the effect of a down-sampling method combining local average and feature points in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention will now be further described in detail with reference to the accompanying drawings and embodiments. These drawings are simplified schematic diagrams that only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention. Embodiment 1

[0021] like Figure 1 , Figure 2 As shown, a non-invasive femoral artery blood flow estimation method based on feature difference includes the following steps: Acquire PPG signal data of the fingertips and dorsum of the subject, and perform baseline drift correction, noise reduction, and zero value extension preprocessing on the PPG signal data of the fingertips and dorsum of the foot to obtain the preprocessed PPG signal data of the fingertips and dorsum of the foot; According to the pre-processed PPG signal data of the fingertips and the dorsum of the foot, based on the pre-trained femoral artery blood flow estimation model, a femoral artery blood flow estimation process is performed to obtain an estimation result of the femoral artery blood flow; The construction of the femoral artery blood flow estimation model includes: constructing a femoral artery blood flow estimation model based on a convolutional neural network-long short-term memory neural network, and incorporating an attention mechanism into the femoral artery blood flow estimation model; Wherein, the training of the femoral artery blood flow estimation model includes: The femoral artery blood flow estimation model is trained using the training set in the data set to obtain the trained femoral artery blood flow estimation model; the signal sequence of the test set in the data set is input into the trained femoral artery blood flow estimation model, different evaluation indicators are calculated, and the estimation result of the femoral artery blood flow is obtained.

[0022] Specifically, baseline drift correction, noise reduction, and zero extension preprocessing include the following steps: Step a1: perform baseline drift correction and noise reduction preprocessing on the PPG signals collected from the fingertips and dorsum of the subject.

[0023] Specifically, in step a1, baseline drift correction and noise reduction preprocessing are performed on the PPG signals of the fingertips and the dorsum of the foot, and discrete wavelet transform is performed on the input signal using the wavedec function and wrcoef function in MATLAB. The basic formula of discrete wavelet transform is: ; in, are discrete wavelet coefficients, is the wavelet basis function; parameter j is the scale factor, which determines The shape of the scale factor includes width and amplitude, and k is the time shift factor, which determines Position along the horizontal axis, x represents the discrete input signal, and t represents the time along the horizontal axis; The specific method is: in step a1, the input signal is first subjected to 10-layer wavelet decomposition, and Daubechies7 wavelet is used as the basic wavelet. According to the result of the wavelet decomposition, the approximate coefficients of the 10-layer wavelet decomposition are reconstructed, and then the approximate coefficients are subtracted from the original signal to obtain the signal after baseline drift correction. The signal after baseline drift correction is subjected to 4-layer wavelet decomposition, and according to the result of the second wavelet decomposition, the approximate coefficients of the 4-layer wavelet decomposition are reconstructed to obtain the denoised signal. Finally, the mean value and standard deviation of the signal are calculated and normalized.

[0024] Step a2, intercepting the preprocessed PPG signals of the fingertips and dorsum of the foot; In step a2, each subject collected PPG signals of fingertips and dorsum of foot in 7 states, which are static state, counterpulsation pressure set to 0.030MPa without calf cuff, counterpulsation pressure set to 0.030MPa with calf cuff, counterpulsation pressure set to 0.035MPa, counterpulsation pressure set to 0.025MPa, counterpulsation pressure set to 0.020MPa, and recovery after completion. From the PPG signals of the subjects in different states, signal segments are intercepted according to the start and end time of the ultrasonic blood flow spectrum tracing points, and the length of the signal segment is determined by the number of cardiac cycles between the ultrasonic blood flow spectrum tracing points. The interception method is to first determine the conditions for finding the peak, including the minimum peak height and the minimum peak spacing, then invert the original PPG signal, use the findpeaks function in MATLAB to find the main peak position of the pulse wave, as the position of the trough found, and finally intercept the corresponding segment according to the position of the trough to obtain the required data.

[0025] Step a3, calculating the difference between the intercepted fingertip PPG and dorsum PPG signals, and downsampling the calculated signal sequence to reduce the amount of data; In step a3, the difference between the intercepted fingertip PPG and dorsum PPG signals is calculated, and the calculated signal sequence is downsampled to reduce the amount of data. The specific method is to first subtract the two signal sequences to obtain a new sequence, and then calculate the size of each downsampling window according to the target sampling frequency. Each downsampling window is traversed on the new sequence, and the following operations are performed for each window: if there is a peak in the window, the peak is used as the representative value of the window; if there is no peak but there is a valley value, the valley value is used as the representative value of the window; if there is neither a peak nor a valley value in the window, the data in the window is averaged to obtain the representative value.

[0026] Step a4: performing a zero-value extension operation on each down-sampled signal segment.

[0027] In step a4, in order to compensate for the inconsistency of the data length of the down-sampled signal segments caused by the dynamic change of the cardiac cycle over time, a zero-value extension operation is performed on each down-sampled signal segment to ensure that they are uniformly expanded to a preset fixed length, thereby achieving standardization and alignment in the dimension of time series data.

[0028] Specifically, the lower limb artery blood flow spectrum obtained by ultrasound measurement is used, and the average blood flow is calculated by the lower limb artery blood flow spectrum as the target value of femoral artery blood flow estimation. Specifically, the target value serves as a label for supervised training of the model.

[0029] Furthermore, obtaining the target value of the femoral artery blood flow estimation includes the following steps: The average blood flow was calculated using the lower limb arterial blood flow spectrum obtained by ultrasound measurement of the subjects. First, the blood flow spectrum curve was extracted by image processing. The main steps included removing the yellow mark generated by the scanner integrated software, selecting ROI, converting the image to grayscale, closing operation for filling, Sobel filtering for edge extraction and envelope extraction. Then, the average blood flow in the time period between the tracing points was calculated based on the product of the integral of the points on the blood flow spectrum curve and the cross-sectional area of ​​the blood vessel, which was used as the target value for femoral artery blood flow estimation.

[0030] Specifically, a femoral artery blood flow estimation model based on a convolutional neural network-long short-term memory neural network is constructed, including: a convolutional neural network and a long short-term memory network; Specifically, the convolutional neural network includes: two consecutive convolutional modules, each of which contains three one-dimensional convolutional layers. A skip connection is used in the third layer to achieve feature fusion, and a normalization layer and a Swish activation function are set after each convolutional layer. , down-sampling is performed through maximum pooling; where x' is the feature of the input signal sequence after convolution and standardization; Specifically, the long short-term memory network includes: Initialize the cell state and hidden state. The cell state is set to the zero vector to accumulate information in the sequence; the hidden state is initialized to the zero vector, representing the output state at the initial moment.

[0031] Specifically, in the long short-term memory network, for each time step in the sequence, the long short-term memory network performs the following steps: Step b1, according to the input signal of the current time step and the hidden state of the previous time step, the input gate vector is obtained by performing input gate calculation through the sigmoid function; specifically, this vector determines how much new information will be added to the cell state.

[0032] Step b2, using the current input and the previous hidden state to perform forget gate calculation to obtain a forget gate vector; specifically, the forget gate calculation determines which information in the cell state needs to be forgotten; specifically, a forget gate value close to 1 indicates that the information is retained, and a forget gate value close to 0 indicates that the information is forgotten.

[0033] Step b3, generating a new vector by inputting candidate calculation; the input candidate calculation includes: candidate information that is desired to be added to the cell state; and is completed by a tanh activation function; specifically, the tanh activation function can ensure that the value range is between (-1,1).

[0034] Step b4, combining the results of the forget gate and the input gate, and using the cell state update to obtain the updated cell state. Specifically, the cell state update includes: forgetting part of the old information, and then adding new information. The specific operation is to multiply the forget gate by the previous cell state element by element (forgetting the old information), and at the same time multiply the input gate by the input candidate vector and add them (adding new information).

[0035] Step b5, calculate the output gate through the sigmoid function in the output gate calculation, which is used to decide which information is important to the outside; specifically, the output gate calculation: decides which parts of the cell state will be used as the output of the current time step.

[0036] Step b6, based on the current cell state and the value of the output gate, a new hidden state is calculated by updating the hidden state. Specifically, the cell state is compressed to the interval (-1, 1) by the tanh function, and then multiplied element by element by the output gate to obtain the final hidden state, which is used as the input of the next layer or as the output of the sequence model.

[0037] Specifically, the attention mechanism consists of two steps: Step c1, linearly transform and unbias the hidden state obtained by the long short-term memory network, and use the tanh function to help the femoral artery blood flow estimation model learn how to assign different weights according to the importance of the hidden state to obtain the attention weight; In step c2, the calculated attention weights are applied to the output of the LSTM network, and weighted summation is achieved through batch matrix multiplication. The attention weights are used to emphasize the time step information that is most useful for the task.

[0038] Specifically, 15% of the dataset used to train the constructed femoral artery blood flow estimation model was allocated as a test set, and the remaining 85% of the dataset was used as a training set; And / or, use AdamW optimizer to optimize data. The Epoch of AdamW optimizer is set to 100, the batch_size is set to 8, and the loss function of AdamW optimizer is set to mean square error MSE. The formula of mean square error MSE is: ;in, is the true value, is the predicted value. Embodiment 2

[0039] A non-invasive femoral artery blood flow estimation method based on feature difference comprises the following steps: Step 1: Collect PPG signals from the fingertips and dorsum of the subject.

[0040] Step 2: Perform baseline drift correction and noise reduction preprocessing on the PPG signals of the fingertips and dorsum of the foot.

[0041] Specifically, baseline drift correction and noise reduction preprocessing are performed on the PPG signals of the fingertips and dorsum of the foot, and discrete wavelet transform is performed on the input signal using the wavedec function and wrcoef function in MATLAB. The basic formula of discrete wavelet transform is: ;in are discrete wavelet coefficients, It is a wavelet basis function. The specific method is to first perform 10-layer wavelet decomposition on the input signal, use Daubechies7 wavelet as the basic wavelet, reconstruct the approximate coefficients of the 10-layer wavelet decomposition according to the result of the wavelet decomposition, and then subtract this approximate coefficient from the original signal to obtain the signal after baseline drift correction. Perform 4-layer wavelet decomposition on the signal after baseline drift correction, and reconstruct the approximate coefficients of the 4-layer wavelet decomposition according to the result of the second wavelet decomposition to obtain the denoised signal. Finally, calculate the mean and standard deviation of the signal and perform normalization.

[0042] Step 3: intercept the preprocessed PPG signals of the fingertips and dorsum of the foot.

[0043] Specifically, each subject collected PPG signals of fingertips and dorsum of foot in 7 states, namely static state, counterpulsation pressure set to 0.030MPa without calf cuff, counterpulsation pressure set to 0.030MPa with calf cuff, counterpulsation pressure set to 0.035MPa, counterpulsation pressure set to 0.025MPa, counterpulsation pressure set to 0.020MPa, and recovery after completion. From the PPG signals of the subjects in different states, signal segments were intercepted according to the start and end time of the ultrasonic blood flow spectrum tracing points, and the length of the signal segment was determined by the number of cardiac cycles between the ultrasonic blood flow spectrum tracing points. The interception method is to first determine the conditions for finding the peak, including the minimum peak height and the minimum peak spacing, then invert the original PPG signal, use the findpeaks function in MATLAB to find the main peak position of the pulse wave, as the position of the trough found, and finally intercept the corresponding segment according to the position of the trough to obtain the required data.

[0044] Step 4: Calculate the difference between the intercepted fingertip PPG and dorsum PPG signals, and downsample the calculated signal sequence to reduce the amount of data.

[0045] Specifically, the difference between the intercepted fingertip PPG and dorsum PPG signals is calculated, and the calculated signal sequence is downsampled to reduce the amount of data. The specific method is to first subtract the two signal sequences to obtain a new sequence, and then calculate the size of each downsampling window according to the target sampling frequency. Each downsampling window is traversed on the new sequence, and the following operations are performed for each window: if there is a peak in the window, the peak is used as the representative value of the window; if there is no peak but there is a valley value, the valley value is used as the representative value of the window; if there is neither a peak nor a valley value in the window, the data in the window is averaged to obtain the representative value.

[0046] Step 5: Perform a zero-value extension operation on each down-sampled signal segment.

[0047] Specifically, in order to compensate for the inconsistency of the data length of the downsampled signal segments caused by the dynamic change of the cardiac cycle over time, a zero-value extension operation is performed on each downsampled signal segment to ensure that they are uniformly expanded to a preset fixed length, thereby achieving standardization and alignment in the dimension of time series data.

[0048] Step 6, using the lower limb arterial blood flow spectrum obtained by ultrasound measurement of the subject, calculate the average blood flow as the target value for femoral artery blood flow estimation.

[0049] Specifically, the average blood flow was calculated using the lower limb arterial blood flow spectrum obtained by ultrasound measurement of the subjects. First, the blood flow spectrum curve was extracted through image processing. The main steps included removing the yellow mark generated by the scanner integrated software, selecting ROI, converting the image to grayscale, closing operations for filling, Sobel filtering for edge extraction and envelope extraction, and then calculating the average blood flow in the time period between the tracing points based on the integral of the points on the blood flow spectrum curve and the product of the cross-sectional area of ​​the blood vessel as the target value for femoral artery blood flow estimation.

[0050] Step 7: Build a femoral artery blood flow estimation model based on convolutional neural network-long short-term memory neural network (CNN-LSTM) and integrate the attention mechanism (Attention).

[0051] Specifically, a femoral artery blood flow estimation model based on convolutional neural network-long short-term memory neural network (CNN-LSTM) is constructed and the attention mechanism is integrated. The overall structure of the model is shown in the figure below: Figure 2 As shown in the figure, the convolutional neural network (CNN) includes two consecutive convolution modules, each of which contains three one-dimensional convolution layers. A jump connection is used in the third layer to achieve feature fusion, and a normalization layer (Batch Normalization, BN) and a Swish activation function are set after each convolution layer. Swish has the characteristics of lower bound, smoothness, and non-monotonicity. The formula is: , which works better than ReLU on deep models, and then downsampled by maximum pooling; where x' is the feature of the input signal sequence after the convolution layer and normalization.

[0052] The long short-term memory network (LSTM) first initializes the cell state (CellState) and hidden state (HiddenState). The cell state is set to a zero vector to accumulate information in the sequence; the hidden state is also initialized to a zero vector, representing the output state at the initial moment.

[0053] For each time step in the sequence, the LSTM neural network performs the following steps: Input gate calculation: Based on the input signal of the current time step and the hidden state of the previous time step, the input gate vector is calculated through the sigmoid function. This vector determines how much new information will be added to the cell state.

[0054] Forget gate calculation: The forget gate vector is calculated using the current input and the previous hidden state to determine which information in the cell state needs to be forgotten. A forget gate value close to 1 indicates that the information is retained, and a value close to 0 indicates that the information is forgotten.

[0055] Input candidate (CellInput) calculation: Generate a new vector representing the candidate information that you want to add to the cell state. This is done using the tanh activation function to ensure that the value range is between (-1,1).

[0056] Cell state update: Combine the results of the forget gate and the input gate to update the cell state. Forget some old information and then add new information. The specific operation is to multiply the forget gate with the previous cell state element by element (forget the old information), and multiply the input gate with the input candidate vector and add them together (add new information).

[0057] Output gate calculation: determines which parts of the cell state will be used as the output of the current time step. The output gate is calculated through the sigmoid function to determine which information is important to the outside world.

[0058] Hidden state update: Calculate the new hidden state based on the current cell state and the value of the output gate. Compress the cell state into the (-1,1) interval through the tanh function, and then multiply it element by element with the output gate to get the final hidden state, which is used as the input of the next layer or as the output of the sequence model.

[0059] The attention mechanism described consists of two steps: Calculate attention weights: Perform a linear transformation on the hidden state obtained by the LSTM layer without bias, and use the tanh function to help the model learn how to assign different weights based on the importance of the hidden state.

[0060] Attention weighting: The calculated attention weight is applied to the output of the LSTM layer, and weighted summation is achieved through batch matrix multiplication (bmm) to emphasize the time step information that is most useful for the task. Specifically, in step 8, the constructed CNN-LSTM-Attention model is trained, 15% of the data set is allocated as a test set, and 85% is used as a training set. Epoch is set to 100, batch_size is set to 8, and the AdamW optimizer is used. The loss function is set to mean square error MSE, and the formula is as follows: ,in is the true value, is the predicted value.

[0061] Step 8: Use the training set to train the femoral artery blood flow estimation model to obtain a trained femoral artery blood flow estimation model.

[0062] Step 9: Input the test set signal sequence into the trained femoral artery blood flow estimation model, calculate different evaluation indicators, and obtain the estimation result of the femoral artery blood flow. Embodiment 3

[0063] In a preferred embodiment of the present invention, a non-invasive femoral artery blood flow estimation device based on characteristic difference comprises: Memory; processor; as well as Computer programs; The computer program is stored in the memory and is configured to be executed by the processor to implement a non-invasive femoral artery blood flow estimation method based on feature differences.

[0064] Working principle: like Figure 1-Figure 2 As shown, the present invention provides a non-invasive femoral artery blood flow estimation method and device based on feature difference, providing doctors with convenient and timely feedback. Compared with the existing technology and experimental methods, the present invention brings a series of significant technical effects and innovative advantages: (1) Innovative non-invasive assessment method. The present invention creatively utilizes the special correlation between pulse wave and blood flow. That is, pulse wave is formed by the pulsation of blood pressure, blood flow velocity and blood flow, as well as the deformation and vibration of blood vessel wall propagating in the vascular system. By using deep learning technology, only the easy-to-collect pulse wave signal is used to assess femoral artery blood flow, providing an alternative to ultrasound and a convenient, fast and low-cost means for assessing the immediate efficacy of EECP in patients with lower limb diseases. (2) Innovative signal analysis method. The present invention creatively utilizes the time difference between the fingertip PPG and the dorsalis pedis PPG signals, that is, the time difference between the blood traveling from the heart to the fingertips and the dorsalis pedis, which is approximately equal to the time it takes for the pulse to pass through the femoral artery. The difference between the fingertip PPG and the dorsalis pedis PPG signals is used as the input feature, which solves the problem that the time interval between the fingertip PPG systolic peak and the dorsalis pedis PPG systolic peak that exists in the traditional sequential connection of the two signals as input may disappear. This method better extracts the difference in physiological characteristics between the two signals, and improves the model's sensitivity to blood flow changes and recognition accuracy.

[0065] (3) Innovative signal downsampling strategy: The present invention creatively selects representative values ​​or average values ​​dynamically according to the peak and valley values ​​in the signal, which not only maintains the key information of the signal, but also effectively reduces the amount of data and improves processing efficiency.

[0066] (4) Accelerate the application of deep learning algorithms in the medical field and promote the popularization and application of EECP in the field of cardiac rehabilitation.

[0067] It can be understood that the various embodiments described in this specification can be implemented individually or in combination, and the embodiments of this specification are not limited to this.

[0068] Unless otherwise specified, all technical and scientific terms used in the embodiments of this specification are the same as those commonly understood by those skilled in the art of the technical field of this specification. The terms used in this specification are only for the purpose of describing specific embodiments and are not intended to limit the scope of this specification.

[0069] The term "and / or" used in this specification includes any and all combinations of one or more of the related listed items. The singular forms "a", "above", and "the" used in the embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0070] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this specification.

[0071] In the several embodiments provided in this specification, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0072] Based on the ideal embodiment of the present invention, through the above description, relevant personnel can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the content in the specification, and the technical scope must be determined according to the scope of the claims.

Claims

1. A non-invasive femoral artery blood flow estimation method based on feature difference, characterized in that: The following steps are involved: Acquire PPG signal data of the fingertips and dorsum of the subject, and perform baseline drift correction, noise reduction, and zero value extension preprocessing on the PPG signal data of the fingertips and dorsum of the foot to obtain the preprocessed PPG signal data of the fingertips and dorsum of the foot; According to the pre-processed PPG signal data of the fingertips and the dorsum of the foot, based on the pre-trained femoral artery blood flow estimation model, a femoral artery blood flow estimation process is performed to obtain an estimation result of the femoral artery blood flow; The construction of the femoral artery blood flow estimation model includes: constructing a femoral artery blood flow estimation model based on a convolutional neural network-long short-term memory neural network, and incorporating an attention mechanism into the femoral artery blood flow estimation model; Wherein, the training of the femoral artery blood flow estimation model includes: The femoral artery blood flow estimation model is trained using the training set in the data set to obtain the trained femoral artery blood flow estimation model; the signal sequence of the test set in the data set is input into the trained femoral artery blood flow estimation model, different evaluation indicators are calculated, and the estimation result of the femoral artery blood flow is obtained.

2. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 1, characterized in that: The baseline drift correction, noise reduction, and zero extension preprocessing include the following steps: Step a1, performing baseline drift correction and noise reduction preprocessing on the PPG signals collected from the fingertips and dorsum of the subject; Step a2, intercepting the preprocessed PPG signals of the fingertips and dorsum of the foot; Step a3, calculating the difference between the intercepted fingertip PPG and dorsum PPG signals, and downsampling the calculated signal sequence to reduce the amount of data; Step a4: performing a zero-value extension operation on each down-sampled signal segment.

3. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 2, characterized in that: The lower limb arterial blood flow spectrum was obtained by ultrasound measurement, and the average blood flow was calculated by the lower limb arterial blood flow spectrum as the target value for femoral artery blood flow estimation.

4. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 3, characterized in that: In step a1, baseline drift correction and noise reduction preprocessing are performed on the PPG signals of the fingertips and the dorsum of the foot, and discrete wavelet transform is performed on the input signal using the wavedec function and wrcoef function in MATLAB. The formula for discrete wavelet transform is: ； in, are discrete wavelet coefficients, is the wavelet basis function, where the parameter j is the scale factor, which determines The shape of the scale factor includes width and amplitude, and k is the time shift factor, which determines Position along the horizontal axis, x represents the discrete input signal, and t represents the time along the horizontal axis; In the step a2, each subject collected the PPG signals of the fingertips and dorsum of the foot under 7 states, and the 7 states were static state, counterpulsation pressure was set to 0.030MPa without adding a calf cuff, counterpulsation pressure was set to 0.030MPa with a calf cuff, counterpulsation pressure was set to 0.035MPa, counterpulsation pressure was set to 0.025MPa, counterpulsation pressure was set to 0.020MPa, and recovery after completion. From the PPG signals of the subjects in different states, signal segments were intercepted according to the start and end time of the ultrasonic blood flow spectrum tracing points, and the length of the signal segment was determined by the number of cardiac cycles between the ultrasonic blood flow spectrum tracing points. The interception method is to first determine the conditions for finding the peak, including the minimum peak height and the minimum peak spacing, then invert the original PPG signal, use the find peaks function in MATLAB to find the main peak position of the pulse wave, as the position of the trough found, and finally intercept the corresponding segment according to the position of the trough to obtain the required data; In step a3, the difference between the intercepted fingertip PPG and dorsum PPG signals is calculated, and the calculated signal sequence is downsampled to reduce the amount of data. The specific method is to first subtract the two signal sequences to obtain a new sequence, and then calculate the size of each downsampling window according to the target sampling frequency. Traverse each downsampling window on the new sequence, and perform the following operations for each window: if there is a peak in the window, use the peak as the representative value of the window; if there is no peak but a valley value, use the valley value as the representative value of the window; if there is neither a peak nor a valley value in the window, use the data in the window as the average value to obtain the representative value; In step a4, in order to compensate for the inconsistency of the data length of the down-sampled signal segments caused by the dynamic change of the cardiac cycle over time, a zero-value extension operation is performed on each down-sampled signal segment to ensure that they are uniformly expanded to a preset fixed length, thereby achieving standardization and alignment in the dimension of time series data.

5. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 4, characterized in that: The acquisition of the target value of the femoral artery blood flow estimation includes the following steps: The average blood flow was calculated using the lower limb arterial blood flow spectrum obtained by ultrasound measurement of the subjects. First, the blood flow spectrum curve was extracted by image processing. The main steps included removing the yellow mark generated by the scanner integrated software, selecting ROI, converting the image to grayscale, closing operation for filling, Sobel filtering for edge extraction and envelope extraction. Then, the average blood flow in the time period between the tracing points was calculated based on the product of the integral of the points on the blood flow spectrum curve and the cross-sectional area of ​​the blood vessel, which was used as the target value for femoral artery blood flow estimation.

6. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 5, characterized in that: Construct a femoral artery blood flow estimation model based on convolutional neural network-long short-term memory neural network, including: convolutional neural network and long short-term memory network; The convolutional neural network includes: two consecutive convolutional modules, each convolutional module includes three one-dimensional convolutional layers, a jump connection is used in the third layer to achieve feature fusion, and a normalization layer and a Swish activation function are set after each convolutional layer, and the Swish activation function is: , down-sampling is performed through maximum pooling, where x' is the feature of the input signal sequence after convolution and standardization; The long short-term memory network comprises: Initialize the cell state and hidden state. The cell state is set to the zero vector to accumulate information in the sequence; the hidden state is initialized to the zero vector, representing the output state at the initial moment.

7. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 6, characterized in that: In the LSTM network, for each time step in the sequence, the LSTM network performs the following steps: Step b1, according to the input signal of the current time step and the hidden state of the previous time step, the input gate is calculated by the sigmoid function to obtain the input gate vector; Step b2, using the current input and the previous hidden state to perform forget gate calculation to obtain the forget gate vector; Step b3, generating a new vector by inputting candidate calculation; the input candidate calculation includes: candidate information that is desired to be added to the cell state; and is completed by a tanh activation function; Step b4, combining the results of the forget gate and the input gate to obtain the updated cell state by using the cell state update; Step b5, calculating the output gate through the sigmoid function in the output gate calculation, which is used to determine which information is important to the outside world; Step b6, calculate the new hidden state through hidden state update based on the current cell state and the value of the output gate.

8. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 7, characterized in that: The attention mechanism consists of two steps: Step c1, linearly transform and unbias the hidden state obtained by the long short-term memory network, and use the tanh function to help the femoral artery blood flow estimation model learn how to assign different weights according to the importance of the hidden state to obtain the attention weight; In step c2, the calculated attention weights are applied to the output of the LSTM network, and weighted summation is achieved through batch matrix multiplication. The attention weights are used to emphasize the time step information that is most useful for the task.

9. The method for non-invasive femoral artery blood flow estimation based on feature difference according to claim 8, characterized in that: 15% of the dataset used to train the constructed femoral artery blood flow estimation model was allocated as a test set, and the remaining 85% of the dataset was used as a training set; And / or, use AdamW optimizer to optimize data. The Epoch of AdamW optimizer is set to 100, the batch_size is set to 8, and the loss function of AdamW optimizer is set to mean square error MSE. The formula of mean square error MSE is: ;in, is the true value, is the predicted value.

10. A non-invasive femoral artery blood flow estimation device based on characteristic difference, characterized in that: include: Memory; processor; as well as Computer programs; Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the non-invasive femoral artery blood flow estimation method based on feature difference as described in any one of claims 1 to 9.

Images