Electronic nose COVID-19 respiratory diagnosis device based on DWT-1DCNN

Abstract

The present invention discloses an electronic nose COVID-19 respiratory diagnostic device based on DWT-1DCNN, comprising an electronic nose device for receiving exhaled gas and generating a raw response signal dataset; a preprocessing module for preprocessing the raw response signal to obtain a preprocessed dataset; a signal decomposition and reconstruction module for decomposing and reconstructing the signal in the preprocessed dataset to obtain an expanded training set; a channel attention module for assigning weights to each feature channel in the expanded training set; and a one-dimensional convolutional neural network module for training the expanded training set to obtain normal or abnormal results for exhaled gas detection. The present invention solves the problem that it is difficult for electronic noses to collect a large number of COVID-19 datasets for model training. It uses DWT to enhance small sample datasets and achieves good results.

Description

Technical Field

[0001] The present invention belongs to the field of human respiratory gas detection devices, and specifically relates to an electronic nose COVID-19 respiratory diagnosis device based on DWT-1DCNN. Background Art

[0002] The current standard and preferred test for the coronavirus (COVID-19) caused by SARS-CoV-2 is based on real-time reverse transcription polymerase chain reaction (RT-PCR) of nasopharyngeal and / or oropharyngeal swabs. The specificity and sensitivity of this test are very high, but the absence of SARS-CoV-2 in the oropharyngeal environment negatively affects the true sensitivity of the test. The incidence of false-negative test results is relatively high, which has led to the desire for a new sampling device. Other diagnostic tests include chest CT scans and analysis of stool or saliva by reverse transcription polymerase chain reaction to detect current infection. However, all of these tests are expensive and time-consuming, require highly trained personnel, and in the case of chest CT scans, expose the patient to X-rays.

[0003] Breath analysis devices are a recently developed detection device. Due to their rapid, non-invasive nature, and ease of patient acceptance, they are gaining increasing attention in the healthcare system. Exhaled breath contains over 1,000 organic compounds. Detecting volatile organic compounds in exhaled breath is primarily accomplished through two methods: chemical composition analysis and pattern recognition-based methods. The most commonly used technique for the former is gas chromatography-mass spectrometry (GC / MS) to identify specific breath biomarkers. However, GC-MS has certain limitations, such as high cost, lack of portability, and difficulty in operation.

[0004] In recent years, deep artificial neural networks have achieved significant success in pattern recognition. These successes rely on big data to ensure good generalization. However, in the field of time series recognition, many datasets are often very small, a particularly prominent example being time series data in the medical field. An effective approach to addressing this issue is to use data augmentation to prevent overfitting. Data augmentation is widely used in the image field, with common methods such as cropping, rotation, and adding noise points. However, these methods do not generalize well to time series data. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: In view of the shortcomings of the existing technology, the present invention proposes an electronic nose COVID-19 respiratory diagnosis device based on DWT-1DCNN.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] The present invention proposes an electronic nose COVID-19 respiratory diagnosis device based on DWT-1DCNN, comprising an electronic nose device for receiving exhaled gas and generating a raw response signal dataset;

[0008] an electronic nose device for receiving exhaled gas and generating a raw response signal dataset;

[0009] The host computer includes a preprocessing module, a signal decomposition and reconstruction module, a channel attention module, and a one-dimensional convolutional neural network module; among them,

[0010] a preprocessing module, configured to receive a dataset of original response signals output by the electronic nose device, and preprocess the original response signals in the dataset to obtain a preprocessed dataset;

[0011] The signal decomposition and reconstruction module is used to decompose and reconstruct the signals in the preprocessed data set to obtain an expanded training set;

[0012] Channel attention module, used to assign weights to each feature channel in the expanded training set;

[0013] A one-dimensional convolutional neural network module is used to train the expanded training set to obtain normal or abnormal results of exhaled gas detection; the one-dimensional convolutional neural network module includes: a first convolution, and a first channel attention module, a first pooling layer, a second convolution, a second channel attention module, a second pooling layer, a flatten module, a fully connected layer, a Dense module, and a Softmax classifier connected in series with the first convolution.

[0014] Furthermore, the aforementioned electronic nose device includes a sensor array, a controller, and a transmission module. The sensor array is a plurality of gas sensors that respond to human exhalation; the controller receives the exhalation response signal output by the gas sensor array and generates an original response signal, which is then output to the preprocessing module through the wireless transmission module.

[0015] Furthermore, the aforementioned pre-processing module is configured to perform the following actions:

[0016] S101. Based on the original response signal dataset X formed by the original response signals of the exhaled breath collected by the electronic nose, the resistance change between the exhaled signal and the baseline signal is divided by the median of the baseline signal according to the following formula:

[0017]

[0018] in, is the response value of the gas sensor to the ambient gas at a certain moment in the baseline phase for the kth breath sample, is the response value of the sensor to the exhalation signal at a certain moment in the exhalation phase of the kth exhalation sample, is the characteristic signal of the gas sensor response to the kth breath sample at a certain moment.

[0019] S102, perform normalization processing according to the following formula to obtain the preprocessed data set X m :

[0020]

[0021] in, is the maximum value of the kth breath sample, is the minimum value of the kth breath sample, For the sample set X m The kth breath sample in .

[0022] Furthermore, the aforementioned signal decomposition and reconstruction module is configured to perform the following actions:

[0023] S201, the sample set X m According to the preset ratio, it is randomly divided into training sample sets X train And the test sample set X test , for the training sample set X train The single gas sensor response signal x(z) in the breath sample is decomposed by wavelet:

[0024] The signal x(z) is downsampled through the low-pass filter F0(Z) according to the following formula to obtain the average signal c(z) with half the scale and resolution, that is, the low-frequency part;

[0025]

[0026] The signal x(z) is downsampled through the high-pass filter F1(Z) according to the following formula to obtain the detail signal d(z) with half the scale and resolution, that is, the high-frequency component:

[0027]

[0028] S202. The average signal c(z) is upsampled according to the following formula, and then passed through a low-pass filter to obtain a large-scale low-resolution approximation, i.e., a low-pass output; the detail signal d(z) is upsampled and then passed through a high-pass filter to obtain a high-pass output. The low-pass output and the high-pass output are added together to obtain a reconstructed signal:

[0029]

[0030] S203. Using the correlation coefficient R as an evaluation index, the similarity between the original signal and the reconstructed signal is calculated according to the following formula:

[0031]

[0032] The range of R is (0, 1), and the closer R is to 1, the higher the similarity of the signals.

[0033] S204, reconstruct the signal and X train The new training set X′ is obtained by accumulating the signals in train .

[0034] Furthermore, the channel attention module is configured to perform the following actions:

[0035] S301. Build a channel attention mechanism, use global average pooling to compress global time information into channel descriptors according to the following formula, and generate a channel statistical vector:

[0036]

[0037] Among them, the input is Y = [y1, y2, ..., y c ], y i ∈R W×1 , channel statistics vector z∈R 1×c , c is the number of channels, w is the length of each channel;

[0038] S302: Use gate control to fully capture channel correlation and generate a channel recalibration vector z′ according to the following formula:

[0039] z′=σ(F″(δ(F′(Z))))

[0040] Where δ is the ReLU activation function, F′ and F″ represent convolution operations with a channel number of 1 and a convolution kernel size of 1*1, respectively, and σ is the Sigmoid function, which compresses the dynamic range of the input activation vector to [0, 1].

[0041] S303. Calibrate the input using the calibration vector z′ and calculate the feature M according to the following formula:

[0042] M=[m1,m2,...,m c ]=Yz′=[y1z′1, y2z′2,...,y c z′ c ].

[0043] Furthermore, the aforementioned one-dimensional convolutional neural network module is configured to perform the following actions:

[0044] Using a one-dimensional convolutional neural network and introducing a channel attention mechanism, the dataset X′ train Each sample data, and data set X' train The preset detection results corresponding to each sample data in the dataset are used as output to build a respiratory gas detection model; trainEach sample data is input, data set X′ train The output is whether the exhaled gas is normal or abnormal corresponding to each sample data. The respiratory gas detection model is trained to obtain the respiratory gas detection model to achieve classification detection of whether the exhaled gas is normal or not.

[0045] Furthermore, the above-mentioned breathing gas detection model is used to test the test set X test Perform classification detection, draw confusion matrix, and calculate classification accuracy, sensitivity, and specificity.

[0046] The present invention adopts the above technical solution, and the beneficial effects compared with the prior art are as follows:

[0047] The present invention proposes to use DWT reconstruction to enhance time series data sets. DWT is often used in the field of time series to eliminate noise signals in the high-frequency part while retaining the main information in the low-frequency part. The indicator R can be used to calculate the similarity between the reconstructed signal and the original signal. Under the condition that R is greater than a certain value, the reconstructed signal basically retains the characteristics of the original signal and can be used for data enhancement. The enhanced data is used to add a 1DCNN with an attention channel mechanism to automatically extract the features of the data, and achieves very good results. It solves the problem that it is difficult for electronic noses to obtain sufficient samples in the field of medical diagnosis, and avoids the difficulties in extracting features and selecting various machine learning classifiers. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic diagram of the overall structure of the electronic nose COVID-19 respiratory diagnosis device based on DWT-1DCNN of the present invention.

[0049] Figure 2 This is the training step for the respiratory gas detection model in the respiratory diagnostic device.

[0050] Figure 3 3 is a comparison diagram of the DWT reconstructed signal and the original signal in one embodiment.

[0051] Figure 4 is a channel attention module in one embodiment.

[0052] Figure 5 1DCNN module in one embodiment.

[0053] Figure 6 Result diagram of classification accuracy, loss and number of iterations in one embodiment, where (a) and (b) are the relationships between classification accuracy, loss and number of iterations before and after data enhancement, respectively. DETAILED DESCRIPTION

[0054] In order to better understand the technical content of the present invention, specific embodiments are given and described below with reference to the accompanying drawings.

[0055] Various aspects of the present invention are described herein with reference to the accompanying drawings, which show a number of illustrative embodiments. The embodiments of the present invention are not limited to those described in the accompanying drawings. It should be understood that the present invention can be implemented by any of the various concepts and embodiments described above, as well as the concepts and implementations described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. In addition, some aspects disclosed herein may be used alone or in any appropriate combination with other aspects disclosed herein.

[0056] like Figure 1 As shown, the electronic nose device includes a sensor array and a controller and a transmission module connected to the sensor array in sequence, which is used to collect the original response signal of human breathing. The sensor array is a number of gas sensors that respond to human exhalation; the transmission module then transmits the original response signal to the host computer, which includes a trained respiratory gas detection model, and uses the model to analyze and detect the original response signal and output the detection results.

[0057] like Figure 2 As shown, the present invention provides an electronic nose COVID-19 respiratory diagnosis device based on DWT-1DCNN, comprising:

[0058] an electronic nose device for receiving exhaled gas and generating a raw response signal dataset;

[0059] The host computer includes a preprocessing module, a signal decomposition and reconstruction module, a channel attention module, and a one-dimensional convolutional neural network module; among them,

[0060] a preprocessing module, configured to receive a dataset of original response signals output by the electronic nose device, and preprocess the original response signals in the dataset to obtain a preprocessed dataset;

[0061] The signal decomposition and reconstruction module is used to decompose and reconstruct the signals in the preprocessed data set to obtain an expanded training set;

[0062] Channel attention module, used to assign weights to each feature channel in the expanded training set;

[0063] A one-dimensional convolutional neural network module is used to train the expanded training set to obtain normal or abnormal results of exhaled gas detection; the one-dimensional convolutional neural network module includes: a first convolution, and a first channel attention module, a first pooling layer, a second convolution, a second channel attention module, a second pooling layer, a flatten module, a fully connected layer, a Dense module, and a Softmax classifier connected in series with the first convolution.

[0064] The pre-processing module is configured to perform the following actions:

[0065] S101. Based on the original response signal dataset X formed by the original response signals of the exhaled breath collected by the electronic nose, the resistance change between the exhaled signal and the baseline signal is divided by the median of the baseline signal according to the following formula:

[0066]

[0067] in, is the response value of the gas sensor to the ambient gas at a certain moment in the baseline phase for the kth breath sample, is the response value of the sensor to the exhalation signal at a certain moment in the exhalation phase of the kth exhalation sample, is the characteristic signal of the gas sensor response to the kth breath sample at a certain moment.

[0068] S102, perform normalization processing according to the following formula to obtain the preprocessed data set X m :

[0069]

[0070] in, is the maximum value of the kth breath sample, is the minimum value of the kth breath sample, For the sample set X m The kth breath sample in .

[0071] The dataset X m The training sample set X is randomly divided into train And the test sample set X test Where X train There are 28 COVID-19 samples and 39 healthy control samples; test There are 13 COVID-19 samples and 18 healthy control samples.

[0072] The signal decomposition and reconstruction module is configured to perform the following actions:

[0073] S201, the sample set X m The training sample set X is randomly divided into train And the test sample set X test Where X train There are 28 COVID-19 samples and 39 healthy control samples; test There are 13 COVID-19 samples and 18 healthy control samples. train The single gas sensor response signal x(z) in the breath sample is decomposed by wavelet:

[0074] The signal x(z) is downsampled through the low-pass filter F0(Z) according to the following formula to obtain the average signal c(z) with half the scale and resolution, that is, the low-frequency part;

[0075]

[0076] The signal x(z) is downsampled through the high-pass filter F1(Z) according to the following formula to obtain the detail signal d(z) with half the scale and resolution, that is, the high-frequency component:

[0077]

[0078] S202. The average signal c(z) is upsampled according to the following formula, and then passed through a low-pass filter to obtain a large-scale low-resolution approximation, i.e., a low-pass output; the detail signal d(z) is upsampled and then passed through a high-pass filter to obtain a high-pass output. The low-pass output and the high-pass output are added together to obtain a reconstructed signal:

[0079]

[0080] S203. Using the correlation coefficient R as an evaluation index, the similarity between the original signal and the reconstructed signal is calculated according to the following formula:

[0081]

[0082] The range of R is (0, 1), and the closer R is to 1, the higher the similarity of the signals.

[0083] S204, reconstruct the signal and X train The new training set X′ is obtained by accumulating the signals in train .

[0084] This study selected three wavelets, bior4.4, coif1, and sym3, to reconstruct the signal. Figure 3 As shown in Figure 3, the DWT reconstructed signal is highly consistent with the original signal.

[0085] Figure 4 Schematic diagram of the channel attention module. The channel attention module is configured to perform the following actions:

[0086] S301. Build a channel attention mechanism, use global average pooling to compress global time information into channel descriptors according to the following formula, and generate a channel statistical vector:

[0087]

[0088] Among them, the input is Y = [y1, y2, ..., y c ], y i ∈RW×1 , channel statistics vector z∈R 1×c , c is the number of channels, w is the length of each channel;

[0089] S302: Use gate control to fully capture channel correlation and generate a channel recalibration vector z′ according to the following formula:

[0090] z′=σ(F″(δ(F′(Z))))

[0091] Where δ is the ReLU activation function, F′ and F″ represent convolution operations with a channel number of 1 and a convolution kernel size of 1*1, respectively, and σ is the Sigmoid function, which compresses the dynamic range of the input activation vector to [0, 1].

[0092] S303. Calibrate the input using the calibration vector z′ and calculate the feature M according to the following formula:

[0093] M=[m1,m2,...,m c ]=Yz′=[y1z′1, y2z′2,...,y c z′ c ].

[0094] like Figure 5 The diagram below shows a schematic diagram of a 1DCNN module. This one-dimensional convolutional neural network (1DCNN) consists of the first convolution layer, followed by the first channel attention module, the first pooling layer, the second convolution layer, the second channel attention module, the second pooling layer, the flatten module, a fully connected layer, a dense module, and a softmax classifier. The convolution kernels used are 1D kernels, with the first convolution kernel size being 1*4 and the second being 1*8. During the convolution process, padding is set to "same" and strides is set to 1. A ReLU activation function is used to increase the nonlinearity of the network. The pooling layer uses max pooling, with the first and second pooling kernel sizes both being 2 and strides set to 2. Flattening is performed after the second pooling layer, followed by a fully connected layer with dropout set to 0.3. Finally, a softmax classifier generates the final labels. The model uses cross-entropy as the loss function, and employs the Adam adaptive momentum stochastic optimization algorithm to ensure rapid convergence of the loss function to the global minimum. During the training process, the learning rate is set to 0.01, the batch_sizes is set to 10, the epoch size is 50, and 20% of the data from the training set is divided into the validation set. The data of each process is shown in Table 1:

[0095] Table 1

[0096]

[0097] Use the trained respiratory gas detection model to test the test set X test Perform classification detection, draw confusion matrix, and calculate classification accuracy, sensitivity, and specificity.

[0098] X train and X′ train , input the respiratory gas detection model respectively, draw the accuracy and loss curves of its training set and validation set respectively, and compare the effects before and after enhancing the dataset.

[0099] like Figure 6 (a) Figure 6 As shown in (b), the validation set loss rate before data augmentation did not converge with the increase in epochs, while the accuracy curve achieved excellent results, indicating that the model had a serious overfitting problem. After data augmentation, the model gradually converged with the increase in epochs, and the overfitting problem was resolved.

[0100] Use the trained model to test the test set X test Classification detection was performed and the classification accuracy, sensitivity and specificity were calculated. The results are shown in Table 2.

[0101] Table 2

[0102] Accuracy Sensitivity Specificity 93.60％ 91.73％ 95.19％

[0103] It can be seen from the table that the DWT-1DCNN model of the present invention has achieved very good results.

[0104] While the present invention has been described above with reference to preferred embodiments, this is not intended to limit the present invention. Persons skilled in the art will readily appreciate that various modifications and variations can be made without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. The electronic nose COVID-19 respiratory diagnostic device based on DWT-1DCNN is characterized by: include: an electronic nose device for receiving exhaled gas and generating a raw response signal dataset; The host computer includes a preprocessing module, a signal decomposition and reconstruction module, a channel attention module, and a one-dimensional convolutional neural network module; among them, a preprocessing module, configured to receive a dataset of original response signals output by the electronic nose device, and preprocess the original response signals in the dataset to obtain a preprocessed dataset; The signal decomposition and reconstruction module is used to decompose and reconstruct the signals in the preprocessed data set to obtain an expanded training set; the signal decomposition and reconstruction module is configured to perform the following actions: S201, the sample set X m According to the preset ratio, it is randomly divided into training sample sets X train And the test sample set X test , for the training sample set X train The single gas sensor response signal x(z) in the breath sample is decomposed by wavelet: The signal x(z) is downsampled through the low-pass filter F0(Z) according to the following formula to obtain the average signal c(z) with half the scale and resolution, that is, the low-frequency part; c j+1 (z)=∑ m∈z c j (m)F0(m-2z), The signal x(z) is downsampled through the high-pass filter F1(Z) according to the following formula to obtain the detail signal d(z) with half the scale and resolution, that is, the high-frequency component: d j+1 (z)=∑ m∈z c j (m)F1(m-2z), S202. The average signal c(z) is upsampled according to the following formula, and then passed through a low-pass filter to obtain a large-scale low-resolution approximation, i.e., a low-pass output; the detail signal d(z) is upsampled and then passed through a high-pass filter to obtain a high-pass output. The low-pass output and the high-pass output are added together to obtain a reconstructed signal: c j (z)=∑ m∈z c j+1 (m)F0(z-2m)+d j+1 (m)F1(z-2m), S203. Using the correlation coefficient R as an evaluation index, the similarity between the original signal and the reconstructed signal is calculated according to the following formula: The range of R is (0, 1), and the closer R is to 1, the higher the similarity of the signals. S204, reconstruct the signal and X train The new training set X′ is obtained by accumulating the signals in train ; Channel attention module, used to assign weights to each feature channel in the expanded training set; A one-dimensional convolutional neural network module is used to train the expanded training set to obtain normal or abnormal ECG signal results for exhaled gas detection; the one-dimensional convolutional neural network module includes: a first convolution, and a first channel attention module, a first pooling layer, a second convolution, a second channel attention module, a second pooling layer, a flatten module, a fully connected layer, a dense module, and a softmax classifier connected in series with the first convolution; Use the trained respiratory gas detection model to test the test set X test Perform classification detection, draw confusion matrix, and calculate classification accuracy, sensitivity, and specificity.

2. The electronic nose COVID-19 respiratory diagnostic device based on DWT-1DCNN according to claim 1 is characterized in that: The electronic nose device includes a sensor array, a controller, and a transmission module. The sensor array is a number of gas sensors that respond to human exhalation. The controller receives the exhalation response signal output by the gas sensor array and generates an original response signal, which is then output to the preprocessing module via the wireless transmission module.

3. The electronic nose COVID-19 respiratory diagnostic device based on DWT-1DCNN according to claim 2 is characterized in that: The pre-processing module is configured to perform the following actions: S101. Based on the original response signal dataset X formed by the original response signals of the exhaled breath collected by the electronic nose, the resistance change between the exhaled signal and the baseline signal is divided by the median of the baseline signal according to the following formula: in, is the response value of the gas sensor to the ambient gas at a certain moment in the baseline phase for the kth breath sample, is the response value of the sensor to the exhalation signal at a certain moment in the exhalation phase of the kth exhalation sample, The characteristic signal of the gas sensor response to the k-th breath sample at a certain moment; S102, perform normalization processing according to the following formula to obtain the preprocessed data set X m : in, is the maximum value of the kth breath sample, is the minimum value of the kth breath sample, For the sample set X m The kth breath sample in .

4. The electronic nose COVID-19 respiratory diagnostic device based on DWT-1DCNN according to claim 3 is characterized in that: The channel attention module is configured to perform the following actions: S301. Build a channel attention mechanism, use global average pooling to compress global time information into channel descriptors according to the following formula, and generate a channel statistical vector: Among them, the input is Y=[y1,y2,…,y c ],y i ∈R W×1 , channel statistics vector z∈R 1×c ,c is the number of channels, w is the length of each channel; S302: Use gate control to fully capture channel correlation and generate a channel recalibration vector z′ according to the following formula: z′=σ(F″(δ(F′(Z)))) Where δ is the ReLU activation function, F′ and F″ represent convolution operations with 1 channel and 1*1 kernel size, respectively, and σ is the Sigmoid function, which compresses the dynamic range of the input activation vector to [0, 1]. S303. Calibrate the input using the calibration vector z′ and calculate the feature M according to the following formula: M=[m1,m2,…,m c ]=Yz′=[y1z′1,y2z′2,…,y c z′ c ]。 5. The electronic nose COVID-19 respiratory diagnostic device based on DWT-1DCNN according to claim 4 is characterized in that: The one-dimensional convolutional neural network module is configured to perform the following actions: Using a one-dimensional convolutional neural network and introducing a channel attention mechanism, the dataset X′ train Each sample data, and data set X' train The preset detection results corresponding to each sample data in the dataset are used as output to build a respiratory gas detection model; train Each sample data is input, data set X′ train The output is whether the exhaled gas is normal or abnormal corresponding to each sample data. The respiratory gas detection model is trained to obtain the respiratory gas detection model to achieve classification detection of whether the exhaled gas is normal or not.

Images