A sleep apnea detection method based on wavelet transform

Abstract

The present application is suitable for sleep apnea detection technical field, and provides a sleep apnea detection method based on wavelet transform.The present application adopts discrete wavelet decomposition technology, decomposes electrocardiogram signals into multi-scale information as model input, and fuses a supervised contrast learning concept, constructs an end-to-end deep learning network model, and realizes efficient and accurate classification of sleep apnea.Compared with the traditional detection method which depends on expert manual analysis method or needs complex preprocessing steps, the present application improves the detection efficiency and simplifies the complexity of signal processing.More importantly, the present application directly uses time series signals as input, avoiding the complex secondary feature extraction process, which makes the present application more easily converted into actual hardware equipment for implementation and application.In addition, using the public ApneaECG dataset for testing, the prediction accuracy of the present application exceeds the current electrocardiogram-based automatic sleep apnea detection model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sleep apnea detection, and particularly relates to a sleep apnea detection method based on wavelet transform. BACKGROUND

[0002] Sleep apnea (SA) is a common sleep disorder characterized by short, periodic breathing stop events during sleep. This disorder often increases the risk of cardiovascular diseases, diabetes and other metabolic diseases. SA is divided into obstructive sleep apnea (OSA) and central sleep apnea, and most adult and child patients with sleep apnea belong to the former.

[0003] Polysomnography (PSG) is considered a diagnostic standard for sleep-related breathing disorders (SRBD), snoring or OSA and other sleep diseases. PSG can collect various physiological signals of patients during overnight sleep, including electroencephalogram, electrooculogram, electromyogram, electrocardiogram, thermal and oral-nasal pressure airflow, etc. However, due to its high cost and complex operation process, its wide application is limited. Therefore, finding a more convenient and accurate detection method has become a research hotspot. Since the correlation between electrocardiogram and respiratory status has been confirmed by clinical experiments, electrocardiogram has gradually become a new way for SA detection. With the rapid development of artificial intelligence technology, especially the introduction of machine learning and deep learning technology, automatic SA detection methods based on electrocardiogram have emerged.

[0004] Automatic SA detection methods based on electrocardiogram are mainly divided into traditional machine learning methods and deep learning methods. Among them, the method based on deep learning contains various front-end data processing and feature extraction methods to adapt to the neural network in the back-end. At present, there are two main feature engineering methods in this field, one is to manually extract statistical features of electrocardiogram such as heart rate, and the other is to extract R peak value and R peak interval from electrocardiogram as features, the latter is widely used in various deep learning methods as the input of the model. However, these methods all need complex analysis and calculation on the original electrocardiogram signal, and the extracted features lose the time domain characteristics and may lose some important information in the original electrocardiogram. In recent years, some methods have appeared to convert one-dimensional electrocardiogram signals into two-dimensional frequency domain images and use image processing methods for SA detection. Although this method can sometimes increase the prediction accuracy of the model, the front-end work is too complicated and the cost is high, and the practical application value is not great. Therefore, a sleep apnea detection method based on wavelet transform is proposed. SUMMARY

[0005] The purpose of the present application is to provide a sleep apnea detection method based on wavelet transform, which aims to solve the problems raised in the background.

[0006] To achieve the above object, the present application provides the following technical solutions:

[0007] A sleep apnea detection method based on wavelet transform, comprising the following steps:

[0008] Step 1, denoising the original ECG signal, using a Symlet4 wavelet function to perform 5-layer discrete wavelet decomposition on the denoised ECG signal, generating 5 high-frequency subbands and 1 low-frequency subband, and splicing the 6 signal components obtained by decomposition as input data for subsequent processing;

[0009] Step 2, inputting the spliced signal components into a CNN composed of 9 one-dimensional convolution layers for feature transformation and extraction; inputting the multi-channel time series features output by the CNN into a BiLSTM to extract time domain features of each channel; inputting the output features of the BiLSTM after flattening into a classifier network composed of two fully connected layers and an activation function layer, and finally calculating the probability of occurrence of a sleep apnea event through a Softmax function, and then determining which type of event according to the probability;

[0010] Step 3, constructing two kinds of contrast losses according to sample labels, namely noise contrast estimation loss and normalized temperature scale cross-entropy loss, and using the overall loss obtained by combining the cross-entropy loss with the two kinds of contrast losses as the final network loss function for model training.

[0011] Further, the decomposition process of the Symlet4 wavelet function for 5-layer discrete wavelet decomposition of the denoised ECG signal comprises:

[0012] Filtering, filtering the ECG signal through a set of low-pass filters and high-pass filters to obtain approximation coefficients and detail coefficients, i.e., low-frequency signal components and high-frequency signal components;

[0013] Downsampling, interval sampling of the filtered signal;

[0014] The calculation formula of the filtering process is as follows:

[0015]

[0016] Where A[n] is the low-frequency approximation coefficient of the nth layer, d[n] is the high-frequency approximation coefficient of the nth layer, A[n+1] and D[n+1] of the nth+1 layer are obtained by decomposing A[n] of the nth layer; x is the original signal sequence, f and g are the high-frequency and low-frequency filters respectively, and k is the serial number corresponding to each layer of decomposition.

[0017] After 5-layer decomposition, 6 signal components D[1], D[2], D[3], D[4], D[5] and A[5] are obtained.

[0018] Further, the channel number of the CNN is c = {16, 16, 32, 32, 64, 64, 128, 128, 256}, and the formula of each layer of convolution is as follows:

[0019] x c = ReLU (Conv1d (x c-1 ))

[0020] where x c-1 is the output feature of the previous layer, ReLU is the activation function used, Conv1d is a one-dimensional convolution sub, and the output of the last layer is defined as x c .

[0021] Further, the formula of the BiLSTM is as follows:

[0022] h n , c n ) = BiLSTM (x c )

[0023] where x c is the input feature, h is the output of all time steps, h n and c n are the hidden state and cell state of the last time step, respectively; finally, h is taken as the input of the classifier network.

[0024] Further, the formula for the classifier network to obtain the probability distribution of the classification result is as follows:

[0025] x out = Softmax (Linear (ReLU (Linear (Flatten (Dropout (h)))))

[0026] where h is the input feature, x out is the output prediction result, ReLU is the activation function used, Linear is a linear neural network, Flatten is a flattening function, and Dropout is a random neuron discarding function.

[0027] Further, the formula of the noise contrast estimation loss is as follows:

[0028] For a target sample x i and its positive sample K noise samples are constructed, where k ∈ {1, 2, …, K}:

[0029]

[0030]

[0031] wherein sim is the similarity of the sample pair, calculated by cosine similarity; x is the sample, is the corresponding positive sample, is the contrast loss of the positive sample pair, is the contrast loss of the noise sample pair, L i is the overall loss, L NCE is the overall loss of this batch, N is the total number of samples, and i is the corresponding sequence number.

[0032] Further, the normalized temperature scale cross-entropy loss formula is as follows:

[0033] for sample x i and its positive sample

[0034]

[0035] wherein sim is the similarity of the sample pair, calculated by cosine similarity; τ is the temperature coefficient, j is the corresponding sequence number, L NTXENT_i is the normalized temperature scale cross-entropy loss of the positive sample pair, L NTXENT is the overall loss of this batch, N is the total number of samples, and i is the corresponding sequence.

[0036] Further, the cross-entropy loss formula is as follows:

[0037]

[0038] wherein L CE is the average cross-entropy loss of this batch of samples, M is the number of samples in this batch; y i is the true label of the jth sample, taking the value of 0 or 1; is the probability of the jth sample belonging to the positive class predicted by the model, taking the value between 0 and 1.

[0039] Further, the overall loss formula is as follows:

[0040] L = αL CE + (λ1L NCE + λ2L NTXENT )

[0041] wherein L is the overall loss, L CE is the cross-entropy loss, L NCE is the noise contrast estimation loss, L NTXENTFor the normalized temperature scale cross-entropy loss, alpha, lambda1 and lambda2 are parameters for adjusting the proportion of the three losses respectively, alpha regulates the cross-entropy loss, the proportion is 90%, lambda1 and lambda2 regulate the noise contrast estimation loss and the normalized temperature scale cross-entropy loss respectively, the total proportion is 10%, the sum of alpha, lambda1 and lambda2 is 1.

[0042] Compared with the prior art, the present application has the following advantages:

[0043] The present application not only reduces the dependence on specific electrocardiogram features, but also improves the detection accuracy. The core is to use discrete wavelet decomposition technology to decompose the electrocardiogram signal into multi-scale information as model input, and to fuse the supervised contrast learning concept to construct an end-to-end deep learning network model, realizing efficient and accurate classification of sleep apnea. Compared with traditional methods that rely on expert manual analysis or require complex preprocessing steps, the present application improves detection efficiency, simplifies signal processing complexity, and makes the entire detection process more smooth and automated. More importantly, the present application directly uses time series signals as input, avoiding complex secondary feature extraction process, which makes the present application more easily converted into actual hardware equipment for implementation. In addition, in terms of performance verification, the ApneaECG dataset is used for testing, and the prediction accuracy of the present application exceeds the current electrocardiogram-based automatic sleep apnea detection model, proving its outstanding performance and great potential in practical application. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The algorithm implementation flowchart of the present application.

[0045] Figure 2 The comparison chart of the prediction results and the true labels of the same sample; (a) is the true label, and (b) is the prediction result. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0047] As shown in Figure 1 , a sleep apnea detection method based on wavelet transform provided by an embodiment of the present application includes the following steps:

[0048] 1. Wavelet decomposition; discrete wavelet decomposition (DWT) is a tool for analyzing signals at multiple resolutions. DWT decomposes the signal into sub-signals of different frequency bands, which have good localization characteristics in both time and frequency. This decomposition method is very suitable for processing non-stationary signals such as images, time series signals, etc. Among them, the common time series signals include audio signals and physiological time series signals such as electrocardiogram (ECG). The original sampling of the electrocardiogram (ECG) signal has certain noise, and there are baseline drift, power frequency interference and other phenomena, which are not convenient for neural network to learn and train directly. DWT can effectively realize the denoising processing of time series data, and extract low-frequency trend information and high-frequency detail information from it, while retaining effective information to a great extent. The wavelet function used in the present application is Symlet4 wavelet, which is a member of the Symlet wavelet family and a kind of wavelet with good symmetry. The Symlet series wavelet has symmetry and orthogonality, in addition, the Symlet4 wavelet can better fit the signals with peak waveform such as electrocardiogram, therefore, the present application first carries out denoising processing on the ECG signal, and the specific operation is to use a bandpass filter, set the high frequency to 45HZ and the low frequency to 0.05HZ. Then use a 5-layer Symlet4 wavelet function to decompose the denoised ECG signal into 5 high-frequency parts and 1 low-frequency part, and then splice the 6 signal components to serve as the input data of the model.

[0049] The decomposition process includes filtering and down-sampling two steps, filtering is to filter the denoised ECG signal through a set of low-pass filter and high-pass filter, to obtain the approximation coefficient and the detail coefficient, i.e. the low-frequency signal and the high-frequency signal component; down-sampling is to sample the filtered signal at intervals. The calculation formula of the filtering process is as follows:

[0050]

[0051] Among them, A[n] is the low-frequency approximation coefficient of the nth layer, D[n] is the high-frequency approximation coefficient of the nth layer, A[n+1] and D[n+1] of the nth+1 layer are obtained by decomposing A[n] of the nth layer; x is the original signal sequence, f and g are the high-frequency and low-frequency filters respectively, and k is the serial number corresponding to each layer of decomposition; after 5 layers of decomposition, D[1], D[2], D[3], D[4], D[5] and A[5] are obtained, which are 6 signal components.

[0052] 2. Hierarchical feature extraction; in order to extract the high-dimensional features of the sequence data, the obtained 6 signal components are spliced and then input into a convolutional neural network (CNN) composed of 9 one-dimensional convolutional layers for feature transformation and extraction, and finally a multi-channel time series feature is obtained.

[0053] where the number of channels is c = {16, 16, 32, 32, 64, 64, 128, 128, 256}, and the formula of each layer of convolution is as follows:

[0054] x c = ReLU(Conv1d(x c-1 ))

[0055] where x c-1 is the output feature of the previous layer, ReLU is the activation function used, Conv1d is a one-dimensional convolutional sub, and the output of the last layer is defined as x c .

[0056] 3. Time domain feature extraction; Bi-directional long short-term memory network (BiLSTM) is a recurrent neural network that can capture the dependencies of sequence data before and after. It processes the input sequence through two LSTM layers in the forward and backward directions, so it can better understand the context information of the sequence. The input of the BiLSTM layer is the multi-channel time sequence features extracted by CNN, and its main function is to further learn the time sequence correlation features in the data. The formula of BiLSTM is as follows:

[0057] h,(h n ,c n ) = BiLSTM(x c )

[0058] where x c is the input feature, h is the output of all time steps, h n and c n are the hidden state and cell state of the last time step, respectively. Finally, h is taken as the input of the classifier network.

[0059] 4. Classification; The output feature of BiLSTM is flattened and input into the classifier network composed of two fully connected layers and an activation function layer. Finally, the output is normalized by the Softmax function of the second fully connected layer to obtain the probability distribution of the two categories, and the probability distribution is then converted into the final prediction result, i.e. whether sleep apnea exists. The probability distribution of the classification result is as follows:

[0060] x out = Softmax9Linear9ReLU9Linear(Flatten(Dropout(h))))))

[0061] where h is the input feature, and x outFor the output prediction result, ReLU is the activation function used, Linear is the linear neural network, Flatten is the flatten function, and Dropout is the random neuron function.

[0062] 5. Contrastive loss;

[0063] (1) Noise contrast estimation (NCE) loss: NCE loss is originally designed for training unsupervised models by contrasting the target distribution with the noise distribution, which is converted into a binary classification problem. By constructing a binary classification problem, the target sample is compared with the noise sample to optimize the embedding representation of the model. The loss function is based on cosine similarity, the similarity distribution is adjusted by a temperature parameter, and the final loss is calculated using a logarithmic loss function.

[0064] The NCE loss formula used in the present application is as follows, for a target sample x i and its positive sample and K noise samples are constructed where k∈{1,2,…,K}:

[0065]

[0066] where sim represents the similarity of the sample pair, calculated by cosine similarity; x is the sample, is the corresponding positive sample, is the contrastive loss of the positive sample pair, is the contrastive loss of the noise sample pair, L i is the overall loss, L NCE is the total loss of this batch, N is the total number of samples, and i is the corresponding serial number.

[0067] (2) Normalized temperature scale cross-entropy (NTXent) loss: NTXent loss is a commonly used loss function for contrastive learning, which is used to learn effective feature representations. The main idea is to maximize the similarity of similar sample pairs (positive sample pairs) while minimizing the similarity of dissimilar sample pairs (negative sample pairs) to train the model. By making the positive sample pairs closer in the embedding space and the negative sample pairs further apart, the embedding representation of the model is optimized. The loss function is also based on cosine similarity, and the similarity distribution is adjusted by a temperature parameter.

[0068] The NTXent loss formula used in the present application is as follows, for a sample x i and its positive sample

[0069]

[0070] wherein, sim is the similarity of sample pairs, calculated by cosine similarity; τ is the temperature coefficient, j is the corresponding serial number, L NTXENT_i is the normalized temperature scale cross-entropy loss of positive sample pairs, L NTXENT is the total loss of this batch, N is the total number of samples, i is the corresponding sequence.

[0071] 6. Loss function; cross-entropy (CE) loss is a commonly used loss function in classification tasks, which aims to measure the difference between the predicted probability distribution and the true distribution. The cross-entropy loss formula used in this invention is as follows:

[0072]

[0073] wherein, L CE is the average cross-entropy loss of this batch of samples, M is the number of samples in this batch; y i is the true label of the jth sample, taking the value of 0 or 1; is the probability of the jth sample predicted by the model to belong to the positive class, taking the value between 0 and 1.

[0074] In the model training process, cross-entropy loss and two kinds of contrast loss are used simultaneously, and the total loss formula is as follows:

[0075] L = aL CE + λ1L NCE + λ2L NTXENT )

[0076] wherein, L is the total loss, L CE is the cross-entropy loss, L NCE is the noise contrast estimation loss, L NTXENT is the normalized temperature scale cross-entropy loss, a, λ1 and λ2 are parameters for adjusting the proportion of the three losses, respectively, a controls the cross-entropy loss, accounting for 90%, λ1 and λ2 control the noise contrast estimation loss and the normalized temperature scale cross-entropy loss, accounting for 10% in total, the sum of a, λ1 and λ2 is 1.

[0077] Example 1, ApneaECG is a public dataset in PhysioNet database. It contains 70 subjects' overnight sleep electrocardiogram records, each record is 7-10 hours long. PhysioNet divides the first 35 records of this dataset as the training set, and the last 35 records as the test set. Based on this division, the invention divides 20% of the segments in the training set as the validation set, trains the model, and stops training at 30 epochs. The prediction index results of the final model on the test set are shown in Table 1:

[0078] Table 1 Prediction index results

[0079]

[0080] As shown in Table 1, the accuracy, recall, specificity, and F1 score of this method reached 92.2%, 90.8%, 93.1%, and 0.899, respectively. Except for specificity, all other indicators are superior to other methods, demonstrating the accuracy and efficiency of this method in the sleep apnea detection task.

[0081] The comparison chart of the predicted results and the true labels for the same sample is shown below. Figure 2 As shown, (a) represents the true label, and (b) represents the prediction result. Figure 2 This indicates that the method can predict most sleep apnea events during sleep, and the overall trend is the same, with some differences in details.

[0082] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A wavelet transform based sleep apnea detection method, characterized by, The method comprises the following steps: Step 1, denoising the original ECG signal, using a Symlet4 wavelet function to perform 5-layer discrete wavelet decomposition on the denoised ECG signal, generating 5 high-frequency subbands and 1 low-frequency subband, splicing the 6 signal components obtained by decomposition as input data for subsequent processing; Step 2, inputting the spliced signal components into a CNN composed of 9 one-dimensional convolutional layers for feature transformation and extraction; inputting the multi-channel time series features output by the CNN into a BiLSTM to extract time domain features of each channel; flattening the output features of the BiLSTM and inputting them into a classifier network composed of two fully connected layers and an activation function layer, and finally calculating the probability of sleep apnea event occurrence through a Softmax function, and then determining the specific type of event according to the probability; Step 3, constructing two kinds of contrast losses, namely noise contrast estimation loss and normalized temperature scale cross-entropy loss, and using the total loss obtained by combining the cross-entropy loss and the two kinds of contrast losses as the final network loss function for model training.

2. The wavelet transform based sleep apnea detection method according to claim 1, characterized in that, The decomposition process of using the Symlet4 wavelet function to perform 5-layer discrete wavelet decomposition on the denoised ECG signal comprises: Filtering, filtering the ECG signal through a set of low-pass filters and high-pass filters to obtain approximation coefficients and detail coefficients, i.e. low-frequency signal components and high-frequency signal components; Downsampling, interval sampling of the filtered signal; The calculation formula of the filtering process is as follows: Wherein, A[n] is the low-frequency approximation coefficient of the nth layer, D[n] is the high-frequency approximation coefficient of the nth layer, A[n+1] and D[n+1] of the (n+1) th layer are obtained by decomposing A[n] of the nth layer; x is the original signal sequence, f and g are high-frequency and low-frequency filters respectively, and k is the serial number corresponding to each layer of decomposition; After 5-layer decomposition, 6 signal components D[1], D[2], D[3], D[4], D[5] and A[5] are obtained.

3. The wavelet transform based sleep apnea detection method of claim 1, wherein, The channel number of the CNN is c={16, 16, 32, 32, 64, 64, 128, 128, 256}, and the formula for each layer of convolution is as follows: x c = ReLU(Conv1d(x c-1 )) where x c-1 is the output of the previous layer, ReLU is the activation function used, Conv1d is a one-dimensional convolutional layer, and the output of the last layer is defined as x c .

4. The wavelet transform based sleep apnea detection method of claim 1, wherein, The formula of the BiLSTM is as follows: h,(h n ,c n )=BiLSTM(x c ) where x c are input features, h is the output at all time steps, h n and c n are the hidden state and cell state at the last time step, respectively; finally, h is taken as the input of a classifier network.

5. The wavelet transform based sleep apnea detection method according to claim 1, wherein, The probability distribution formula of the classification result obtained by the classifier network is as follows: x out = Softmax(Linear(ReLU(Linear(Flatten(Dropout(h)))))) where h is the input feature, x out is the output prediction result, ReLU is the activation function used, Linear is the linear neural network, Flatten is the flattening function, and Dropout is the random neuron dropping function.

6. The wavelet transform based sleep apnea detection method according to claim 1, wherein, The noise contrast estimation loss formula is as follows: For a target sample x i and its positive sample Construct K noisy samples where k ∈ {1, 2, …, K}: wherein sim is the similarity of the sample pair, calculated by cosine similarity; x is the sample, is the corresponding positive sample, is the contrast loss of the positive sample pair, is the contrast loss of the noise sample pair, L i is the overall loss, L NCE is the overall loss of this batch, N is the total number of samples, and i is the corresponding serial number.

7. The wavelet transform based sleep apnea detection method according to claim 1, wherein, The normalized temperature scale cross-entropy loss formula is as follows: For a sample x i and its positive sample wherein sim is the similarity of the sample pair, calculated by cosine similarity; τ is the temperature coefficient, j is the corresponding serial number, L NTXENT_i is the normalized temperature scale cross-entropy loss of the positive sample pair, L NTXENT is the total loss of this batch, N is the total number of samples, i is the corresponding sequence.

8. The wavelet transform based sleep apnea detection method of claim 1, wherein, The cross-entropy loss formula is as follows: wherein L CE is the average cross-entropy loss of the batch of samples, and M is the number of samples in the batch; y i is the true label of the jth sample, taking a value of 0 or 1; is the probability that the jth sample belongs to the positive class predicted by the model, taking a value between 0 and 1.

9. The wavelet transform based sleep apnea detection method of claim 1, wherein, The total loss formula is as follows: L = aL CE + (λ1L NCE + λ2L NTXENT ) where L is the total loss, L CE is the cross-entropy loss, L NCE is the noise contrast estimation loss, L NTXENT is the normalized temperature-scaled cross-entropy loss, and a, l1, and l2 are parameters that adjust the proportions of the three losses, respectively. a controls the cross-entropy loss, which accounts for 90%. l1 and l2 control the noise contrast estimation loss and the normalized temperature-scaled cross-entropy loss, respectively, which together account for 10%. The sum of a, l1, and l2 is 1.

Images