Automatic detection method, device and equipment for internal arteriovenous fistula stenosis based on wrist strap equipment and storage medium

Abstract

The invention provides an internal arteriovenous fistula stenosis automatic detection method, device and equipment based on wrist strap equipment, and a storage medium. According to the technical scheme of the invention, PPG signals of a single-side wrist of a user are automatically acquired through intelligent wrist strap equipment, so that the tedious operation of using a plurality of sensors is avoided; the detection convenience and the wearing comfort are greatly improved, and the device is particularly suitable for long-time monitoring in a daily home environment. A time-frequency deep neural network model based on a multi-head spectrum attention mechanism is constructed, and the time domain and frequency domain characteristics of PPG signals are combined, so that the degree of internal fistula stenosis can be accurately predicted. According to the technical scheme, the complexity of traditional detection is simplified, internal fistula stenosis detection is easier to operate and popularize, serious consequences caused by delayed treatment are avoided, the complexity of traditional detection is greatly simplified, the method is suitable for daily use and long-term monitoring of patients, potential problems can be found in time, and the detection efficiency is improved. And the complication risk caused by internal fistula stenosis is reduced.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more specifically, to a wristband-based automatic detection method, device, equipment and storage medium for arteriovenous fistula stenosis. Background Art

[0002] Arteriovenous fistula (AVF) is an important vascular access for hemodialysis patients to maintain their lives, and its health status is directly related to the efficiency of dialysis treatment and the long-term health of patients. Arteriovenous fistula stenosis is one of the main causes of reduced dialysis effectiveness and fistula failure, which may cause a variety of serious complications, including thrombosis, infection, and complete loss of vascular access. Therefore, regular detection of arteriovenous fistula stenosis is essential to prevent these complications and prolong the service life of the fistula. Early detection of stenosis can not only reduce the patient's physical pain, but also significantly reduce the medical expenses caused by complications, improve the safety and effectiveness of dialysis treatment, and thus improve the patient's overall quality of life and survival rate.

[0003] Angiography and Doppler ultrasound are the gold standard for evaluating arteriovenous fistula stenosis, and they can provide precise anatomical and hemodynamic information. However, these methods are invasive and require expensive equipment and professional operators, which limits their application in daily monitoring and home care. In addition, frequent examinations may bring additional burden and risks to patients.

[0004] In the existing technology, capacitive sensors and sound sensors are used to evaluate the stenosis status of arteriovenous fistulas. However, capacitive sensors can cause fluctuations in measurement data due to interference from skin conditions, sweat, and humidity. In addition, the complex calibration and setup process limits their popularity and application in home care. Sound sensors assess blood vessel status by capturing the sound of blood flow, but are easily disturbed by environmental noise and patient activities. In addition, sound sensors need to be close to the skin, which may cause skin discomfort to patients. These problems make the two sensor technologies face challenges in terms of breadth and effectiveness in clinical applications.

[0005] Compared with the above methods, the existing PPG (photoplethysmography) technology has become a research hotspot for arteriovenous fistula stenosis detection due to its advantages of non-invasiveness, low cost and easy operation. PPG technology evaluates blood flow status by detecting light signals generated by changes in blood volume, and has the advantages of non-invasive and real-time monitoring. However, the existing PPG detection method also has some shortcomings: 1. The current research on fistula stenosis based on PPG signals is conducted in a clinical environment. The device uses a finger-clip pulse wave transmission sensor, which requires simultaneous measurement of the PPG signals of the patient's left and right fingers, and the difference in the PPG signals of the left and right hands is used to judge the fistula blockage, which increases the complexity and difficulty of operation of the device; 2. Bilateral monitoring may increase the patient's wearing discomfort and compliance problems, resulting in patients being unwilling to use the device for a long time or at a prescribed frequency.

[0006] Traditional machine learning methods require manual feature extraction and input into model training, resulting in the separation of feature extraction and model training processes and a lack of automatic optimization capabilities. Previous deep learning methods usually focus on the time domain and pay less attention to the frequency domain, failing to fully utilize the fusion information of the time and frequency domains of the PPG signal, resulting in certain limitations in the model's ability to capture complex signal patterns. Summary of the invention

[0007] The present invention provides a wristband-based automatic detection method, device, equipment and storage medium for arteriovenous fistula stenosis. By specifically targeting the fistula vascular stenosis data set collected by the wristband device, a time-frequency deep neural network model based on a multi-head spectral attention mechanism is proposed, which solves the technical problems of existing arteriovenous fistula stenosis detection methods that are complex to operate, time-consuming, costly and inconvenient to wear.

[0008] The present invention provides an automatic detection method for arteriovenous fistula stenosis based on a wristband device, comprising:

[0009] Obtaining a user PPG signal collected by the wristband device;

[0010] Arteriovenous fistula stenosis analysis is performed based on a pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism to generate arteriovenous fistula stenosis risk classification results.

[0011] Furthermore, the present invention also provides an automatic detection device for arteriovenous fistula stenosis based on a wristband device, comprising:

[0012] A PPG signal acquisition and processing module, used to acquire the user's PPG signal in real time and perform PPG signal processing;

[0013] A detection module is used to perform arteriovenous fistula stenosis analysis based on a pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism, and generate arteriovenous fistula stenosis risk classification results;

[0014] The early warning module is used to make early warning selection according to the risk classification result of the arteriovenous fistula stenosis.

[0015] Furthermore, the present invention also provides an automatic detection device for arteriovenous fistula stenosis based on a wristband device, comprising a processor and a memory;

[0016] The memory stores a computer program.

[0017] The processor is used to execute one or more programs stored in the memory to implement the steps of the above-mentioned wristband device-based method for automatic detection of arteriovenous fistula stenosis.

[0018] Furthermore, the present invention also provides a computer-readable storage medium, which stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the above-mentioned wristband device-based automatic detection method for arteriovenous fistula stenosis.

[0019] Beneficial Effects

[0020] The present invention provides an automatic detection method, device, equipment and storage medium for arteriovenous fistula stenosis based on a wristband device. The technical solution of the present invention automatically collects the PPG signal of a single wrist through an intelligent wristband device, avoiding the cumbersome operation of using multiple sensors, greatly improving the convenience of detection and the comfort of wearing, and is particularly suitable for long-term monitoring in daily home environments. By constructing a time-frequency deep neural network model based on a multi-head spectral attention mechanism, the time domain and frequency domain characteristics of the PPG signal are combined to accurately predict the degree of fistula stenosis. Through the technical solution of the present invention, the complexity of traditional detection is simplified, making fistula stenosis detection easier to operate and popularize, avoiding the serious consequences of delayed treatment, greatly simplifying the complexity of traditional detection, suitable for daily use and long-term monitoring of patients, and helping to timely discover potential problems and reduce the risk of complications caused by fistula stenosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained like these drawings without paying creative work.

[0022] Figure 1 A flowchart of a method for automatically detecting arteriovenous fistula stenosis based on a wristband device provided in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of the structure of a time-frequency deep neural network model based on a multi-head spectral attention mechanism provided in an embodiment of the present invention;

[0024] Figure 3 It is a schematic structural diagram of the automatic detection device for arteriovenous fistula stenosis based on a wristband device of the present invention. DETAILED DESCRIPTION

[0025] The following is a description of an automatic detection method for arteriovenous fistula stenosis based on a wristband device in an embodiment of the present application in conjunction with the accompanying drawings. Figure 1 , a schematic diagram of an automatic detection method for arteriovenous fistula stenosis based on a wristband device is provided, which can be applied to wristband devices such as smart watches and smart bracelets with PPG (photoplethysmography) signal acquisition function in the prior art.

[0026] This embodiment is described by taking the method applied to a smart watch with a PPG signal acquisition function as an example. The method for automatically detecting arteriovenous fistula stenosis based on a wristband device in this embodiment includes the following steps:

[0027] S101, obtaining a user PPG signal collected by the wristband device;

[0028] In this step, the method for obtaining the user's PPG signal may be through a smart watch, which obtains the PPG signal and pre-processes the PPG signal.

[0029] In actual applications, users can customize the frequency of periodic acquisition of the user's PPG signal through smart watches, such as collecting data every 2 hours or 3 hours, with each collection time of 12 seconds. This flexible period setting allows users to adjust the detection frequency according to their own needs and medical advice to ensure early detection and timely response to fistula stenosis. Since fistula stenosis is not a health problem that requires real-time monitoring, periodic detection can effectively balance the timeliness of warning and the convenience of users' lives. After the user collects the data, the smart watch pre-processes the collected PPG signals through a preset algorithm, including denoising and standardization, to ensure the stability and comparability of the signal.

[0030] S102. Perform arteriovenous fistula stenosis analysis based on a pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism to generate arteriovenous fistula stenosis risk classification results.

[0031] In this step, the time-frequency deep neural network model based on the multi-head spectral attention mechanism can be obtained by training the artificial neural network using the prior art, such as lightweight network models such as FCN, Unet, Unet++, SegNet, ENet and ESPNet, and the time-frequency deep neural network model based on the multi-head spectral attention mechanism is pre-trained; the PPG signal is used to perform arteriovenous fistula stenosis analysis on the user using the pre-trained time-frequency deep neural network model based on the multi-head spectral attention mechanism to obtain the arteriovenous fistula stenosis risk classification result of the user.

[0032] In practical application, the risk classification results of arteriovenous fistula stenosis can be divided into three groups:

[0033] Grade 1: stenosis rate between 0-30%;

[0034] Grade II: stenosis rate between 30% and 50%;

[0035] Grade III: Stenosis rate exceeds 50%.

[0036] In other embodiments, Figure 2 As shown, the time-frequency deep neural network model based on the multi-head spectrum attention mechanism includes a time domain branch and a frequency domain branch;

[0037] In practical applications, the time domain branch can be the LSTM network of the prior art to extract the time domain features of the PPG signal, and the frequency domain branch can be the time series convolution network of the prior art to extract the frequency domain features of the PPG signal. The time domain branch and frequency domain branch settings of the time-frequency deep neural network model based on the multi-head spectrum attention mechanism of the present invention make full use of the fusion information of the time domain and frequency domain of the PPG signal, and solve the limitations of the prior art model in capturing the complex pattern of the PPG signal. The original signal input by the time domain branch directly reflects the characteristics of the PPG signal changing over time. For patients with fistula stenosis, the dynamic characteristics of the amplitude, waveform, rise and fall time in the PPG signal can reveal the immediate changes in blood flow, and these changes in time are directly related to pathological conditions such as blood flow obstruction and stenosis. The frequency domain branch converts the signal into different frequency components by means of fast Fourier transform (FFT) and the like. For patients with fistula stenosis, abnormalities in blood flow can cause changes in specific frequency components, such as changes in high-frequency components caused by eddy currents or irregular flows. Such frequency features are often difficult to observe directly through the time domain, so the frequency domain branch helps to identify hidden non-periodic features.

[0038] In another embodiment, the pre-trained time-frequency deep neural network model based on the multi-head spectral attention mechanism includes:

[0039] S201, constructing a PPG signal dataset for internal fistula stenosis;

[0040] In this step, the construction of the fistula stenosis dataset adopted a cross-sectional research method, and the research subjects were 70 patients on maintenance hemodialysis. All patients were first screened by professional medical staff to ensure that they met the experimental research conditions. For the screened patients, medical staff performed detailed measurements of their arteriovenous fistulas (AVFs) using color Doppler ultrasound, evaluated the ratio of the stenotic inner diameter to the normal blood vessel diameter, obtained the stenosis rate, and grouped and labeled the patients according to different stenosis rates. Specifically divided into the following three groups:

[0041] Level 1: patients with stenosis rate of 0-30%;

[0042] Grade II: patients with stenosis between 30% and 50%;

[0043] Level 3: Patients with stenosis rate exceeding 50%.

[0044] PPG signals are collected on the dorsal and inner wrists of each eligible patient. The selection of collection points is based on the common wearing positions of wristband devices. During the data collection process, each patient wears the device to continuously collect PPG signals for 2-3 minutes. The patient remains quiet and the wrist is naturally relaxed to reduce the impact of environmental noise, muscle activity and external interference on the signal. The PPG signal acquisition equipment is calibrated to ensure the consistency and signal quality of each measurement.

[0045] After the acquisition is completed, the original PPG signal is preprocessed, which mainly includes the following steps:

[0046] 1. Filtering: Use a low-pass filter to remove high-frequency noise in the signal, and use polynomial fitting and band-pass filtering to remove baseline drift caused by breathing and other reasons to ensure the smoothness and usability of the data.

[0047] 2. Standardization: The PPG signal is standardized and the amplitude of the data is adjusted to a standard normal distribution with a mean of 0 and a standard deviation of 1, so as to ensure that the data amplitudes of different patients are within the same range, which is convenient for subsequent reasonable comparison and analysis.

[0048] 3. Sliding window segmentation: Cut each long PPG signal into shorter segments, set a fixed window size (12 seconds) and sliding step size (10 seconds), that is, the distance the window moves each time. For each signal segment generated by the sliding window, its label is consistent with the original data label.

[0049] Through the above steps, a reliable wristband-based PPG signal dataset of fistula stenosis was successfully constructed, covering patient data with different degrees of stenosis, providing data support for subsequent model training and evaluation.

[0050] S202, extracting the time domain features of the PPG signal in the internal fistula stenosis PPG signal dataset through the time domain branch to obtain a time domain feature vector;

[0051] In this step, each PPG signal in the internal fistula stenosis PPG signal data set is processed by a lightweight neural network model in the prior art, such as an LSTM network, to obtain a time domain feature vector of the PPG signal.

[0052] In other embodiments, the time domain branch includes an LSTM layer, a temporal attention mechanism layer, and a global average pooling layer;

[0053] Step S202 includes:

[0054] S212, the PPG signal time series is processed by an LSTM layer to obtain a hidden state sequence of each time step of the PPG signal time series;

[0055] In this step, the input PPG signal time series data is first processed by the LSTM layer. The LSTM layer is a prior art, and the relevant formula for processing the PPG signal time series data is as follows:

[0056] f t =σ(W f ·h t-1 ,x t ]+b f ) (Forget Gate)

[0057] i t =σ(W i ·[h t-1 , x t ]+b i ) (Input Gate)

[0058]

[0059] o t =σ(W o ·h t-1 ,x t ]+b o ) (Output Gate)

[0060] h t =o t tanh(C t ) (Hidden state)

[0061] LSTM solves the problem that traditional RNN is difficult to capture long-term dependencies through the gating mechanism, using input gate, forget gate and output gate to control the transmission of information. t is the input of the current time step, h t-1 is the hidden state of the previous time step, C t is the current memory cell state, σ is the sigmoid activation function, and tanh is the hyperbolic tangent activation function. The output of the LSTM layer is the hidden state sequence H = {h1,h2,...,h T}, where T is the total length of the input PPG signal time series. The hidden state is continuously updated and transmitted with the input of each time step. LSTM calculates and updates the hidden state through a series of gate mechanisms. The generation process of the hidden state mainly depends on the interaction of three gates: input gate, forget gate, output gate, and memory unit (Cell State, Ct); the hidden state is combined with the input x of the current time step t , the hidden state h of the previous time step t-1 and the memory cell state C t-1 The forget gate determines which memories of the previous time step are discarded, the input gate determines how much new information is saved in the current time step, and the output gate determines how much of the current memory is passed to the hidden state h of the next time step. t .h t-1 is the hidden state calculated by LSTM at time step t-1, representing the summary of the sequence information up to time step t-1, h t-1 It also contains the LSTM’s memory of the input at previous time steps, which is passed to time step t as part of the input.

[0062] S222, the hidden state sequence calculates the attention weight of each time step of the hidden state sequence through the temporal attention mechanism layer to obtain a weighted hidden state sequence;

[0063] In this step, the hidden state sequence of each time step of the PPG signal time series output by the LSTM layer is further processed by the temporal attention mechanism. The temporal attention mechanism enables the model to focus on the key time steps in the PPG signal by calculating the attention weight of each time step of the hidden state sequence.

[0064] The temporal attention mechanism calculates the attention weights through the query matrix Q, the key matrix K and the value matrix V:

[0065] Q=XW Q , K = XW K , V = XW V

[0066] Among them, WQ , W K and W V are the weight matrices for query, key, and value, respectively, and X is the feature of the input hidden state sequence.

[0067] Then calculate the attention score (scaled dot-productattention):

[0068]

[0069] Among them, d k is a scaling factor used to prevent the gradient from disappearing due to excessive values. Softmax is used to convert the score into a probability distribution to obtain the attention weight matrix Attention(Q,K,V).

[0070] The value matrix is ​​weighted and summed by the attention weights to obtain the weighted time domain feature representation:

[0071]

[0072] Among them, i is the attention weight at the i-th time step, V i is the value corresponding to the attention weight at the i-th time step.

[0073] In another embodiment, position encoding is performed before the hidden state sequence passes through the temporal attention mechanism layer.

[0074] In this step, the position encoding of the hidden state sequence is calculated by the following formula:

[0075]

[0076] Among them, pos is the position of the hidden state sequence time step, i is the index of the feature dimension, and d is the size of the feature dimension.

[0077] In practical applications, position encoding is introduced into the temporal attention mechanism to enable the model to identify the relative positions of different time steps. The weight of each time step is dynamically calculated through the temporal attention mechanism, so that the model can focus on the key time steps in the signal. At the same time, the introduction of position encoding enhances the model's ability to perceive the relative position of the time step, and can identify the position relationship of each time step in the sequence, improving the model's effect in processing non-stationary signals or periodic signals.

[0078] S232. The weighted hidden state sequence passes through the global average pooling layer to obtain the time domain feature vector.

[0079] In this step, the hidden state sequence weighted by the temporal attention mechanism passes through the global average pooling layer to calculate the average value of each feature dimension in the time dimension, which is calculated by the following formula:

[0080]

[0081] Among them, X is the weighted hidden state sequence, T is the total length of the hidden state sequence, and x t is the characteristic value of the tth time step in the hidden state sequence, representing the characteristics of the hidden state sequence at the tth moment.

[0082] This operation converts dynamic time series features into feature vectors of fixed length to facilitate subsequent classification operations.

[0083] S203, extracting the frequency domain features of the PPG signal in the internal fistula stenosis PPG signal data set through the frequency domain branch to obtain a frequency domain feature vector;

[0084] In this step, each PPG signal in the internal fistula stenosis PPG signal data set is processed by a prior art temporal convolutional network, such as TCN, to obtain a frequency domain feature vector of the PPG signal.

[0085] In one embodiment, the frequency domain branch includes a multi-head spectrum attention mechanism layer and a temporal convolutional network; the multi-head spectrum attention mechanism layer includes multiple attention heads, a fully connected layer and a sparsification mechanism layer;

[0086] Step S203 includes:

[0087] S213, converting the PPG signal time series into a frequency domain signal by fast Fourier transform;

[0088] In this step, in the multi-head spectrum attention mechanism layer, the input PPG signal time series is first converted into a frequency domain signal through fast Fourier transform (FFT). In this way, the frequency characteristics of the PPG time domain signal can be decomposed into different frequency components, represented as the real and imaginary parts of the spectrum, respectively. The real part represents the amplitude of the signal, while the imaginary part represents the phase information of the signal. The formula is as follows:

[0089]

[0090] Where x(t) is the input signal of the PPG signal time series at time step t, f is the frequency, T is the total length of the sequence, and X(f) is the frequency domain signal in complex form corresponding to frequency f.

[0091] The real part features of different frequency features of the frequency domain signal are transformed through different attention heads of the multi-head spectrum attention mechanism. The formula is as follows:

[0092]

[0093] Where W i is the weight matrix of the i-th attention head of the multi-head spectral attention mechanism, b i is the bias term;

[0094] In this step, each attention head in the multi-head spectral attention mechanism layer independently processes the frequency domain features of the input and models different frequency ranges. Each attention head learns the feature transformation of a specific frequency range. First, the real part feature Re(X) of the frequency domain signal after Fourier transformation is transformed to obtain multiple real part features after transformation.

[0095] S233, passing the plurality of real features after feature transformation through the fully connected layer to obtain an attention score;

[0096] In this step, the multiple real features after feature transformation are transformed through the fully connected layer to generate an intermediate feature representation. On this basis, the importance of each feature is calculated to obtain the attention scores of the multiple real features. Through this step, more critical features of the frequency domain signal are extracted.

[0097] S243. Normalize the attention score through the Softmax function to generate attention weights of multiple real features.

[0098] In this step, the Softmax function ensures that the sum of the attention weights is 1, reflecting the importance of each feature.

[0099] S243. The normalized attention weight is processed through the sparse mechanism layer to obtain the sparse attention weight, and the formula is as follows:

[0100]

[0101] Among them i is the attention weight of the i-th time step, ɑ i ' is the attention weight of the i-th time step after sparse processing, τ is the weight threshold;

[0102] In this step, the sparsification mechanism is set to filter out unimportant or noisy components in the frequency domain. The normalized attention weights are processed through the sparsification mechanism layer, and unimportant weights are removed to obtain the sparsified attention weights. This process only retains the features most relevant to the target task by setting a threshold. By setting a weight threshold τ, the attention weights below the threshold are reset to 0. The sparsification mechanism can filter out unimportant or noisy components in the frequency domain. By setting a threshold, the attention weights below the threshold are reset to zero, and only the frequency components most relevant to the target task are retained. This sparsification process not only enhances the model's ability to extract key information, but also greatly reduces the overhead of calculating unnecessary spectral components. Compared with traditional processing methods, this sparsification mechanism effectively reduces the complexity of the model and reduces the calculation time.

[0103] The real part of the spectrum is weighted using the sparse attention weights. Each attention head weights the most relevant frequency components in the frequency domain. After weighting, the real part features of the spectrum are adjusted according to their importance. Each head can automatically select the most valuable frequency information when processing these spectral features.

[0104] S253. Process the attention weight after the sparse processing through the temporal convolutional network and the global average pooling layer to obtain the frequency domain feature vector.

[0105] In practical applications, since the multi-head spectral attention mechanism focuses on extracting features in different frequency ranges, these frequency domain features may still have dynamic changes in the time dimension. By inputting the attention weights after the sparse processing into the temporal convolutional network, the dependency of frequencies in the time dimension and the interaction between local frequencies can be captured through the characteristics of its dilated convolution. The temporal convolutional network introduces intervals between input data to expand the receptive field and capture long-term dependencies. Compared with traditional convolution, it can capture the dependencies between long-distance time steps with less computation, thereby capturing features at different frequency scales. The dilated convolution formula of the temporal convolutional network is:

[0106]

[0107] The frequency domain feature vector obtained after processing by the temporal convolutional network is a multi-scale representation based on frequency and time interaction modeling. It contains complete information about the dynamics of frequency changes over time, further enhancing the expressiveness of spectral features, so that the model can not only focus on local frequency characteristics, but also capture global frequency dependencies and change patterns.

[0108] Subsequently, the variable-length spectral feature sequence is converted into a frequency domain feature vector of fixed-length feature representation by calculating the global average value of each spectral feature through the global average pooling layer. This process provides a unified feature dimension for the subsequent adaptive fusion with the time domain feature vector, ensuring that the model can more effectively integrate and utilize the information of both.

[0109] S204, performing adaptive feature fusion on the time domain feature vector and the frequency domain feature vector to obtain the PPG signal feature vector;

[0110] In this step, the time domain feature vector and the frequency domain feature vector are adaptively fused. The adaptive feature fusion method may be an adaptive feature fusion algorithm of the prior art, such as ASFF (adaptivelyspatial feature fusion), to obtain the PPG signal feature vector.

[0111] In other implementations, step S204 includes:

[0112] S214, respectively calculating the weights of the time domain feature vector and the frequency domain feature vector through a fully connected layer to obtain a time domain feature weight and a frequency domain feature weight;

[0113] In this step, the weights of the time domain feature vector and the frequency domain feature vector are calculated respectively through the fully connected layer. These weights are dynamically adjusted through model learning, and the degree of contribution of the frequency domain and time domain features can be adaptively adjusted.

[0114] S224, normalizing the time domain feature weight and the frequency domain feature weight by a Softmax function, the formula is as follows:

[0115]

[0116] where α t is the normalized time domain feature weight, α f is the normalized frequency domain feature weight, w t and w f Respectively represent the time domain feature weight and the frequency domain feature weight;

[0117] In this step, the time domain feature weight and the frequency domain feature weight are normalized by the Softmax function to ensure that the total weight of the two is 1.

[0118] S234. Perform weighted fusion on the normalized time domain feature weight and the normalized frequency domain feature weight to obtain the PPG signal feature vector.

[0119] In this step, the normalized time domain feature weights and the normalized frequency domain feature weights are weighted and fused by weighted summation. The model can adaptively adjust its dependence on time domain and frequency domain information according to different input conditions.

[0120] S205, extracting features from the PPG signal feature vector through a fully connected layer, and outputting risk classification results through a batch normalization layer and a Dropout layer;

[0121] In this step, the adaptively fused PPG signal feature vector is passed through a fully connected layer for deep feature extraction, and a batch normalization layer (BatchNormalization) and a Dropout layer are used to prevent overfitting.

[0122] S206. Repeat the above process to complete the training of the time-frequency deep neural network model based on the multi-head spectral attention mechanism.

[0123] In this step, the above process is repeated, i.e., steps S201-S205 are repeated to complete the training of the time-frequency deep neural network model based on the multi-head spectral attention mechanism for each PPG signal in the fistula stenosis PPG signal dataset, thereby forming the pre-trained time-frequency deep neural network model based on the multi-head spectral attention mechanism.

[0124] In practical applications, the time-frequency deep neural network model based on the multi-head spectrum attention mechanism of the present invention extracts the time domain features of the PPG signal through the LSTM layer and the sequential attention mechanism, and captures the frequency domain features through the multi-head spectrum attention mechanism and the sequential convolutional network. The design of the model is lightweight-oriented, using fewer network layers and parameters while ensuring accuracy, and introducing a sparse mechanism in the multi-head attention mechanism, so that the model only focuses on the most important frequency information, reducing unnecessary computational overhead. In addition, the PPG signal input to the model is only a signal with a sampling frequency of 25Hz for 12 seconds, and the amount of data is small, which further reduces the computational complexity and is suitable for running on resource-constrained devices such as smart watches. After the time domain and frequency domain features of the PPG signal are extracted, the model uses a global average pooling layer instead of a fully connected layer, which effectively reduces the number of parameters and the amount of calculation, while retaining key information, further reducing memory usage and computational costs. Finally, the model completes the classification task of the degree of fistula stenosis through adaptive feature fusion, and outputs the classification result of fistula stenosis risk.

[0125] In one embodiment, Figure 3 As shown, an automatic detection device for arteriovenous fistula stenosis based on a wristband device comprises:

[0126] A PPG signal acquisition and processing module 301 is used to acquire the user's PPG signal in real time and perform PPG signal processing;

[0127] A detection module 302 is used to perform arteriovenous fistula stenosis analysis based on a pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism to generate an arteriovenous fistula stenosis risk classification result;

[0128] The early warning module 303 is used to perform early warning selection according to the arteriovenous fistula stenosis risk classification result.

[0129] In practical applications, the wristband-based automatic detection device for arteriovenous fistula stenosis can be integrated into wristband devices of the prior art, such as smart watches, smart bracelets, etc., or can be set in other forms of wristband devices, and the device has a PPG signal acquisition function. The PPG signal acquisition and processing module 301 is used to acquire the user's PPG signal in real time and process the PPG signal. The acquisition time interval can be set by the user in a customized manner, such as evaluating the fistula stenosis once every 2 hours or 3 hours, and each acquisition time is 12 seconds. This flexible cycle setting allows the user to adjust the detection frequency according to their own needs and medical advice to ensure early detection and timely response to fistula stenosis; specifically, the user's PPG signal is acquired through the existing technology, and the PPG signal is sent to the PPG signal acquisition and processing module 301 in real time by wire or wireless means. The PPG signal acquisition and processing module 301 processes the acquired real-time PPG signal, and the processing method includes denoising and standardization to ensure the stability and comparability of the signal.

[0130] The detection module 302 is used to perform arteriovenous fistula stenosis analysis based on a pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism, and generate a conclusion on the risk level of arteriovenous fistula stenosis; the pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism is obtained by training a fistula stenosis PPG signal data set; the detection module 302 generates a conclusion on the risk level of arteriovenous fistula stenosis.

[0131] The early warning module 303 performs early warning selection according to the conclusion of the risk degree of arteriovenous fistula stenosis generated by the detection module 302.

[0132] In practical applications, the detection module 302 detects the PPG signal of the user once, and the risk level conclusions obtained are divided into:

[0133] Grade 1: stenosis rate is 0-30%;

[0134] Grade II: stenosis rate is 30% to 50%;

[0135] Grade III: Stenosis rate exceeds 50%.

[0136] Each test result is automatically recorded to form continuous health data. By setting a minimum test frequency within a day, such as testing once every 3 hours, we ensure that a sufficient number of samples are used for comprehensive judgment, reducing the possibility of triggering a false alarm based on only one high-risk result.

[0137] The warning module 303 will only issue an alarm if the results show medium or high risk for at least two or more consecutive times in a day. The warning options are as follows:

[0138] Low risk reminder: For situations where the risk level conclusion is the second or third level or less than once, the early warning module 303 will not trigger an emergency alarm, but will provide the user with a gentle reminder in the form of text or voice to remind them to continue to maintain good daily management habits.

[0139] Moderate risk warning: When the risk level is concluded as the second level twice or more, indicating that there is a certain risk of blockage, but it is not serious. The early warning module 303 will issue a reminder and advise the user to take preventive measures, such as strengthening simple exercises of the arms and hands, promoting blood circulation, and increasing the monitoring frequency.

[0140] High-risk warning: If the detection result is the third level twice or more, it shows that the user's congestion is more serious. The early warning module 303 will immediately issue an emergency alarm, which can be in the form of text or voice, or vibrate to remind the user, and suggest that he go to the hospital for examination and treatment as soon as possible.

[0141] The present invention also provides an automatic detection device for arteriovenous fistula stenosis based on a wristband device, comprising a processor and a memory;

[0142] The memory stores a computer program.

[0143] The processor is used to execute one or more programs stored in the memory to implement the steps of the method for automatically detecting arteriovenous fistula stenosis based on a wristband device in the above embodiment.

[0144] The present invention also provides a computer-readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the steps of the automatic detection method of arteriovenous fistula stenosis based on a wristband device in the above embodiment.

[0145] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0146] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0147] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods of various embodiments of the present invention.

[0148] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A wristband device-based automatic detection method for arteriovenous fistula stenosis, characterized in that: include: Obtaining a user PPG signal collected by the wristband device; Arteriovenous fistula stenosis analysis is performed based on a pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism to generate arteriovenous fistula stenosis risk classification results.

2. The method for automatic detection of arteriovenous fistula stenosis based on a wristband device according to claim 1, characterized in that: The time-frequency deep neural network model based on the multi-head spectral attention mechanism includes a time domain branch and a frequency domain branch.

3. The method for automatic detection of arteriovenous fistula stenosis based on a wristband device according to claim 1, characterized in that: The pre-trained time-frequency deep neural network model based on the multi-head spectral attention mechanism includes: Construct a PPG signal dataset for fistula stenosis; Extracting the time domain features of the PPG signal in the internal fistula stenosis PPG signal data set through the time domain branch to obtain a time domain feature vector; Extracting the frequency domain features of the PPG signal in the internal fistula stenosis PPG signal data set through the frequency domain branch to obtain a frequency domain feature vector; Adaptively fusing the time domain feature vector and the frequency domain feature vector to obtain the PPG signal feature vector; Extract features from the PPG signal feature vector through a fully connected layer, and output a risk classification result through a batch normalization layer and a Dropout layer; Repeat the above process to complete the training of the time-frequency deep neural network model based on the multi-head spectral attention mechanism.

4. The method for automatic detection of arteriovenous fistula stenosis based on a wristband device according to claim 3 is characterized in that: The time domain branch includes an LSTM layer, a temporal attention mechanism layer, and a global average pooling layer; The step of extracting the time domain features of the PPG signal in the internal fistula stenosis PPG signal data set through the time domain branch to obtain the time domain feature vector comprises: The PPG signal time series is processed by an LSTM layer to obtain a hidden state sequence of each time step of the PPG signal time series; The hidden state sequence calculates the attention weight of each time step of the hidden state sequence through the temporal attention mechanism layer to obtain a weighted hidden state sequence; The weighted hidden state sequence passes through the global average pooling layer to obtain the time domain feature vector.

5. The method for automatic detection of arteriovenous fistula stenosis based on a wristband device according to claim 4 is characterized in that: Position encoding is performed before the hidden state sequence passes through the temporal attention mechanism layer.

6. The method for automatic detection of arteriovenous fistula stenosis based on a wristband device according to claim 3, characterized in that: The frequency domain branch includes a multi-head spectrum attention mechanism layer, a temporal convolutional network layer and a global average pooling layer; the multi-head spectrum attention mechanism layer includes multiple attention heads, a fully connected layer and a sparsification mechanism layer; The extracting the frequency domain features of the PPG signal in the internal fistula stenosis PPG signal data set by the frequency domain branch to obtain the frequency domain feature vector comprises: Convert the PPG signal time series into a frequency domain signal by fast Fourier transform; The real part features of different frequency features of the frequency domain signal are passed through the multi-head spectrum attention mechanism The feature transformation is performed on the different attention heads, and the formula is as follows: Where W i is the weight matrix of the i-th attention head of the multi-head spectral attention mechanism, b i is the bias term; Passing the plurality of real features after feature transformation through the fully connected layer to obtain an attention score; Normalizing the attention score by a Softmax function to obtain a normalized attention weight; The normalized attention weight is weighted by the sparse mechanism layer to obtain the attention weight after sparse processing, and the formula is as follows: Among them i is the attention weight of the i-th time step, ɑ i ' is the attention weight of the i-th time step after sparse processing, τ is the weight threshold; The attention weight after the sparse processing is processed through the temporal convolutional network and the global average pooling layer to obtain the frequency domain feature vector.

7. The method for automatic detection of arteriovenous fistula stenosis based on a wristband device according to claim 3, characterized in that: The time domain feature vector and the frequency domain feature vector are adaptively fused to obtain the PPG signal feature vector, including: Through the fully connected layer, the weights of the time domain feature vector and the frequency domain feature vector are calculated respectively to obtain the time domain feature weight and the frequency domain feature weight; The time domain feature weight and the frequency domain feature weight are normalized by the Softmax function, and the formula is as follows: where α t is the normalized time domain feature weight, α f is the normalized frequency domain feature weight, w t and w f Respectively represent the time domain feature weight and the frequency domain feature weight; The normalized time domain feature weight and the normalized frequency domain feature weight are weightedly fused to obtain the PPG signal feature vector.

8. An automatic detection device for arteriovenous fistula stenosis based on a wristband device, characterized in that: include: A PPG signal acquisition and processing module, used to acquire the user's PPG signal in real time and perform PPG signal processing; A detection module is used to analyze arteriovenous fistula stenosis based on a pre-trained time-frequency deep neural network model based on a multi-head spectral attention mechanism, and generate a conclusion on the risk level of arteriovenous fistula stenosis; The early warning module is used to make early warning selection according to the risk classification result of the arteriovenous fistula stenosis.

9. An automatic detection device for arteriovenous fistula stenosis based on a wristband device, characterized in that: Including processor and memory; The memory stores a computer program. The processor is used to execute one or more programs stored in the memory to implement the steps of the method for automatic detection of arteriovenous fistula stenosis based on a wristband device as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the automatic detection method of arteriovenous fistula stenosis based on a wristband device as described in any one of claims 1 to 7.

Images