Regional pulmonary ventilation disorder disease identification and grade evaluation system

Abstract

The invention discloses a recognition and grade evaluation system for a regional pulmonary ventilation disorder disease. The recognition and grade evaluation system comprises a multi-channel human body impedance acquisition module, a signal processing module, a pulmonary function evaluation model and an evaluation result output module. The technical effect of the method is undoubted, and the method carries out the comprehensive analysis and training of the impedance values of multiple channels of the lung based on the deep neural network through the bioelectrical impedance detection technology, so as to obtain a classification result. And the diagnosis efficiency and accuracy of the lung ventilation disorder disease are improved. According to the invention, the lung ventilation disorder of a patient can be intelligently evaluated in a non-invasive and real-time manner, the change of an illness state can be found in time, and an important basis is provided for adjusting a treatment scheme in time; misdiagnosis possibly caused by manual evaluation is avoided, and the diagnosis efficiency and accuracy of the lung ventilation disorder disease are improved; the device is convenient to operate, low in use cost and capable of reducing the economic overhead of patients.

Description

Technical Field

[0001] The present invention relates to the technical field of disease diagnosis, and in particular to a system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases. Background Art

[0002] Respiratory diseases are a common and multiple disease that seriously affects people's life and health. Pulmonary function status assessment helps to understand the health of the human lungs, and pulmonary function parameters can be used to assess whether an individual has a respiratory disease. Pulmonary function status assessment has important reference value for the type and degree of airway obstruction, airway hyperresponsiveness, and reversibility of airway obstruction.

[0003] In the diagnosis of lung diseases, regional pulmonary ventilation disorder is a common pathological manifestation. Its accurate identification and evaluation are of great significance for the early diagnosis of the disease, the formulation of treatment plans, and the evaluation of patient prognosis. The current diagnosis and evaluation methods have many shortcomings. Traditional medical imaging examinations, such as X-rays and CT, can provide overall information on the lung structure, but they are expensive and lack convenience, real-time performance, and popularity. Although other diagnostic technologies such as the detection of molecular biological markers and genetic analysis have certain specificity, they still have many limitations in clinical application.

[0004] Secondly, although pulmonary function testing technologies, such as vital capacity testing and airway resistance testing, can evaluate lung ventilation function, they are also insufficient in the evaluation of regional ventilation disorders. These technologies mainly focus on changes in overall lung function and are difficult to accurately reflect the ventilation conditions of specific areas. At the same time, pulmonary function tests are complex to operate, require high patient cooperation, have a high risk of cross-infection, and the evaluation results are usually manually observed and concluded by doctors, which are easily affected by subjective factors and are likely to result in inaccurate diagnosis and evaluation results. Summary of the invention

[0005] The purpose of the present invention is to provide a regional pulmonary ventilation disorder disease identification and grade assessment system, including a multi-channel human body impedance acquisition module, a signal processing module, a lung function assessment module, and an assessment result output module.

[0006] The multi-channel human body impedance acquisition module is used to collect impedance change data of different lung regions of the human body. The lung regions include the right upper lobe region, the right middle lobe region, the right lower lobe region, the left upper lobe region, and the left lower lobe region.

[0007] The signal processing module is used to pre-process the lung area impedance change data collected by the multi-channel human body impedance collection module.

[0008] The pulmonary function assessment module stores a pulmonary function assessment model.

[0009] The lung function assessment model is used to perform feature extraction, feature analysis, feature integration and classification on the pre-processed lung regional impedance change data, and assess the type of lung disease and the corresponding severity.

[0010] The evaluation result output module is used to output the type of lung disease evaluated and the corresponding severity.

[0011] Furthermore, the multi-channel human body impedance acquisition module includes a constant current source, a biosensor analog front-end chip, and a plurality of excitation electrodes.

[0012] The biosensor analog front-end chip is equipped with a constant current source.

[0013] The constant current source is used to provide a constant current signal.

[0014] The biosensor analog front-end chip injects the current signal output by the constant current source into the human chest cavity through the excitation electrode, and then collects impedance change data of different lung areas.

[0015] The placement point of the excitation electrode is determined by the following steps:

[0016] Mark the center of the first rib on the midclavicular line of the subject to be tested as position 8, mark the center of the eighth rib as position 5, mark positions 7 and 6 equidistantly on the line between 8 and 5, mark positions 1 to 4 at the intersection of the horizontal positions of 5 to 8 and the right axillary midline, mark positions 9 to 12 at the intersection of the horizontal positions of 5 to 8 and the left axillary midline, and select positions 1, 2, 3, 6, 7, 8, 9, and 10 as electrode placement points.

[0017] Furthermore, the preprocessing includes wavelet filtering, wavelet decomposition, soft thresholding and inverse wavelet transform.

[0018] The wavelet filtering process is used to filter the lung area impedance change data to obtain filtered impedance data.

[0019] The wavelet decomposition process uses a multi-layer soft threshold method to set the threshold of each level, and uses the threshold of each level to decompose the filtered impedance data to obtain detail coefficients.

[0020] The soft threshold processing sets detail coefficients below a threshold to zero and shrinks detail coefficients above a threshold to remove noise.

[0021] The inverse wavelet transform process performs inverse wavelet transform on the detail coefficients after soft threshold processing to obtain the pre-processed lung area impedance change data.

[0022] Furthermore, the calculation formula of the wavelet filtering process is as follows:

[0023] Z=sign(x)×max(0,x-x0) (1)

[0024] Where Z is the impedance modulus after filtering, x is the input impedance modulus, x0 is the threshold, sign(x) represents the sign function of x, and max(0,x-x0) means that if x is greater than x0, the output is x, and if x is less than or equal to x0, the output is 0.

[0025] Furthermore, the method for extracting features from the preprocessed lung area impedance change data by the lung function assessment model includes a convolutional neural network, and the steps are as follows:

[0026] s11 constructs three convolutional neural network modules and sets s=1.

[0027] The convolution kernel size of the convolution layer of the first convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is d, and the number of output channels is e, where a, b, c, d, and e are all positive integers.

[0028] The convolution kernel size of the convolution layer of the second convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is e, and the number of output channels is 2e.

[0029] The convolution kernel size of the convolution layer of the third convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is 2e, and the number of output channels is 4e.

[0030] s12 convolves the data of different input channels with the corresponding channels of the sth convolutional neural network module, and adds the convolved data of different input channels to obtain the elements of the output feature map.

[0031] The calculation formula of the convolution is as follows:

[0032] Z st =W s ·X t:t+1 +b s (2)

[0033] Where t is the time step. st is the output feature vector of the sth convolutional layer at time step t. s is the weight matrix of the sth convolutional layer. t:t+1 is the local area of ​​the input data X from time step t to t+1. s is the bias term.

[0034] s13 activates the elements of the output feature map through the ReLU function, and the activated feature vector is as follows:

[0035] A st =max(0,Z st ) (3)

[0036] In the formula, A st is the activated feature vector.

[0037] s14 normalizes the activated feature vector to obtain the feature vector output by the convolution layer of the sth convolutional neural network module As shown below:

[0038]

[0039] In the formula, γ st is the scaling parameter. st is the offset parameter. ε is a constant.

[0040] Among them, the mean μ and variance σ of the current iteration batch 2 As shown below:

[0041]

[0042] Where i is the iteration sample number, and n0 is the number of iteration samples.

[0043] s15 takes the feature vector output by the convolutional layer of the sth convolutional neural network module Input the pooling layer to obtain the feature vector output by the sth convolutional neural network module, as shown below:

[0044]

[0045] Where J is the time step position. j is the time step. is the feature vector output by the convolutional layer of the sth convolutional neural network module at time step j. st [J] is the feature vector output by the sth convolutional neural network module at time step J.

[0046] S16 determines whether s=3 is established. If not, the feature vector output by the sth convolutional neural network module is used as input data, s=s+1, and returns to step s12. If yes, the local features of the input data are extracted.

[0047] Furthermore, the steps of characteristic analysis of the lung function assessment model are as follows:

[0048] s21 constructs two bidirectional LSTM layers, sets u=1, and uses the local features of the input data as input features.

[0049] The number of input channels of the first bidirectional LSTM layer is 4e, and the hidden state dimension is 2e.

[0050] The second bidirectional LSTM layer has an input channel number of 4e and a hidden state dimension of e.

[0051] s22 inputs the input features into the forward input gate of the u-th bidirectional LSTM layer and obtains the forward gate output as shown below:

[0052] z t =softmax(w h ·h t +b h ) (8)

[0053] In the formula, z t is the forward gate output of the u-th bidirectional LSTM layer. h is the coefficient of the u-th bidirectional LSTM layer. h is the weight of the u-th bidirectional LSTM layer.

[0054] Among them, the state of the hidden layer at time t h t As shown below:

[0055] h t =o t ⊙tanh(C t ),C t =f t ⊙(C t-1 )+i t ⊙(g t ) (9)

[0056] In the formula, C t , C t-1 are the intermediate parameters at time t and t-1 respectively. ⊙ is the Hadamard product of the matrix.

[0057] Among them, the weight matrices of the forget gate, input gate, new memory gate, and output gate are as follows:

[0058]

[0059] In the formula, f t 、i t , g t , o t are the weight matrices of the forget gate, input gate, new memory gate, and output gate respectively. σ is the Sigmoid function. p t is the input feature. t-1 is the state of the hidden layer at time t-1. f 、w i 、w g 、w oare the input coefficients of the forget gate, input gate, new memory gate, and output gate respectively. f 、v i 、v g 、v o are the hidden coefficients of the forget gate, input gate, new memory gate, and output gate respectively. f 、b i 、b g 、b o They are the weights of the forget gate, input gate, new memory gate, and output gate respectively.

[0060] s23 will output the positive gate z t Back propagate and update the parameters.

[0061] The calculation formula for updating parameters is as follows:

[0062]

[0063] Where E is the loss function. t+1 is the intermediate parameter at time t+1. t+1 is the weight matrix of the forget gate at time t+1. t is the input feature.

[0064] s24 determines whether u=2 holds. If not, the output feature of the u-th bidirectional LSTM layer is used as the input feature, u=u+1, and returns to step s22. If yes, feature analysis data is obtained.

[0065] Furthermore, the steps of feature integration of the lung function assessment model are as follows:

[0066] s31 is based on the forward gate output z of the bidirectional LSTM layer t , construct two fully connected layers.

[0067] s32 converts the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Input the first fully connected layer and use the Flatten function to convert the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Flatten it into a one-dimensional array, and then use the Linear function to expand the hidden layer state h t The number of features.

[0068] The data output by the first fully connected layer is input into the second fully connected layer. The Flatten function is used to flatten the data output by the first fully connected layer into a one-dimensional array. Then the Linear function is used to expand the feature number of the output data of the first fully connected layer to obtain the hidden layer state h after the expanded feature number. t .

[0069] s33 applies ReLU activation function and Dropout layer to the hidden layer state h after expanding the number of features t Perform feature integration.

[0070] The ReLU activation function is as follows:

[0071] H t =max(0,h t ) (12)

[0072] In the formula, H t To activate the feature.

[0073] The output of the Dropout layer is as follows:

[0074] D(H t )=H t ⊙m (13)

[0075] In the formula, D(H t ) is the output of the Dropout layer. ⊙ is the Hadamard product of the matrix.

[0076] Among them, the random number m is as follows:

[0077]

[0078] Where p is the probability value. i is the random mask function.

[0079] Further, the steps of classification of the lung function assessment model are as follows:

[0080] S41 classifies lung diseases into z1 Build a disease classifier.

[0081] The severity is classified into y z2 Build a severity classifier.

[0082] s42 Dropout layer output D(H t ) is input into the disease classifier and the Linear function is used to z1 The involved diseases are selected from the lung diseases, and then the Softmax activation function is used to obtain the probability distribution of the selected disease types.

[0083] s43 Dropout layer output D(H t ) is input into the severity classifier and the Linear function is used to z2 The involved severity information is selected from the severity levels, and then the Softmax activation function is used to obtain the probability distribution of the selected severity information.

[0084] s44 completes the classification of the patient's disease and severity based on the probability distribution of disease types and the probability distribution of severity information.

[0085] Further, the Softmax activation function is as follows:

[0086]

[0087] Where softmax(D) is the Softmax activation function. D is the output of the Dropout layer. n is the ordinal number representing the category, and n represents the number of categories.

[0088] Furthermore, the steps of constructing the lung function assessment model include:

[0089] a1 collects human lung impedance data of different disease types and severities, and constructs training sets and test sets.

[0090] a2 builds an evaluation model.

[0091] a3 takes human lung impedance data as input and disease type and severity as output, and uses the training set to train the evaluation model to obtain a trained evaluation model.

[0092] a4 inputs the human lung impedance data of the test set into the trained evaluation model, and adjusts the hyperparameters of the evaluation model according to the output data to obtain the lung function evaluation model.

[0093] The technical effect of the present invention is unquestionable. The present invention uses bioelectrical impedance detection technology to comprehensively analyze and train the impedance values ​​of multiple lung channels based on a deep neural network to obtain classification results, thereby improving the diagnostic efficiency and accuracy of pulmonary ventilation disorders.

[0094] The present invention can perform intelligent assessment of the patient's pulmonary ventilation disorder in real time and in a non-invasive manner, and can promptly detect changes in the condition, providing an important basis for timely adjustment of the treatment plan; it avoids misdiagnosis that may be caused by manual assessment, and improves the diagnostic efficiency and accuracy of pulmonary ventilation disorder diseases; the present invention is easy to operate, has a low cost of use, and can reduce the patient's economic expenses. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 It is a flowchart of the present invention;

[0096] Figure 2 It is a schematic diagram of the three-dimensional lung five-zone model of the present invention;

[0097] Figure 3 is a two-dimensional chest cavity model of the present invention;

[0098] Figure 4is a diagram of the configuration of the lung impedance acquisition electrodes of the present invention;

[0099] Figure 5 It is a structural diagram of the human body impedance acquisition system of the present invention;

[0100] Figure 6 It is a schematic diagram of the wavelet filtering principle of the present invention;

[0101] Figure 7 It is the image of the lung impedance waveform diagram and impedance value after wavelet filtering of the present invention; Figure 7 (a) is the waveform of lung impedance; Figure 7 (b) is the image of impedance value after wavelet filtering;

[0102] Figure 8 It is a schematic diagram of the convolutional neural network principle of the present invention;

[0103] Fig. 9 It is a schematic diagram of the principle of the long short-term memory network of the present invention;

[0104] Fig.10 It is a schematic diagram of the accuracy of the regional lung disease classification model of the present invention; DETAILED DESCRIPTION

[0105] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0106] Embodiment 1:

[0107] See also Figures 1 to 10 , a regional pulmonary ventilation disorder disease identification and grade assessment system, including a multi-channel human body impedance acquisition module, a signal processing module, a pulmonary function assessment module, and an assessment result output module.

[0108] The multi-channel human body impedance acquisition module is used to collect impedance change data of different lung regions of the human body. The lung regions include the right upper lobe region, the right middle lobe region, the right lower lobe region, the left upper lobe region, and the left lower lobe region.

[0109] The signal processing module is used to pre-process the lung area impedance change data collected by the multi-channel human body impedance collection module.

[0110] The pulmonary function assessment module stores a pulmonary function assessment model.

[0111] The lung function assessment model is used to perform feature extraction, feature analysis, feature integration and classification on the pre-processed lung regional impedance change data, and assess the type of lung disease and the corresponding severity.

[0112] The evaluation result output module is used to output the type of lung disease evaluated and the corresponding severity.

[0113] Embodiment 2:

[0114] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders. The main technical content is shown in Example 1. Furthermore, the multi-channel human body impedance acquisition module includes a constant current source, a biosensor analog front-end chip, and a plurality of excitation electrodes.

[0115] The biosensor analog front-end chip is equipped with a constant current source.

[0116] The constant current source is used to provide a constant current signal.

[0117] The biosensor analog front-end chip injects the current signal output by the constant current source into the human chest cavity through the excitation electrode, and then collects impedance change data of different lung areas.

[0118] The placement point of the excitation electrode is determined by the following steps:

[0119] Mark the center of the first rib on the midclavicular line of the subject to be tested as position 8, mark the center of the eighth rib as position 5, mark positions 7 and 6 equidistantly on the line between 8 and 5, mark positions 1 to 4 at the intersection of the horizontal positions of 5 to 8 and the right axillary midline, mark positions 9 to 12 at the intersection of the horizontal positions of 5 to 8 and the left axillary midline, and select positions 1, 2, 3, 6, 7, 8, 9, and 10 as electrode placement points.

[0120] Embodiment 3:

[0121] A system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases, the main technical content of which is shown in any one of Examples 1 to 2. Furthermore, the preprocessing includes wavelet filtering processing, wavelet decomposition processing, soft threshold processing, and inverse wavelet transform processing.

[0122] The wavelet filtering process is used to filter the lung area impedance change data to obtain filtered impedance data.

[0123] The wavelet decomposition process uses a multi-layer soft threshold method to set the threshold of each level, and uses the threshold of each level to decompose the filtered impedance data to obtain detail coefficients.

[0124] The soft threshold processing sets detail coefficients below a threshold to zero and shrinks detail coefficients above a threshold to remove noise.

[0125] The inverse wavelet transform process performs inverse wavelet transform on the detail coefficients after soft threshold processing to obtain the pre-processed lung area impedance change data.

[0126] Embodiment 4:

[0127] A system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases, the main technical content of which is shown in any one of Embodiments 1 to 3. Further, the calculation formula of the wavelet filtering process is as follows:

[0128] Z=sign(x)×max(0,x-x0) (1)

[0129] Where Z is the impedance modulus after filtering, x is the input impedance modulus, x0 is the threshold, sign(x) represents the sign function of x, and max(0,x-x0) means that if x is greater than x0, the output is x, and if x is less than or equal to x0, the output is 0.

[0130] Embodiment 5:

[0131] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of embodiments 1 to 4. Further, the method for extracting features from the preprocessed lung regional impedance change data by the pulmonary function evaluation model includes a convolutional neural network, and the steps are as follows:

[0132] s11 constructs three convolutional neural network modules and sets s=1.

[0133] The convolution kernel size of the convolution layer of the first convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is d, and the number of output channels is e, where a, b, c, d, and e are all positive integers.

[0134] The convolution kernel size of the convolution layer of the second convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is e, and the number of output channels is 2e.

[0135] The convolution kernel size of the convolution layer of the third convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is 2e, and the number of output channels is 4e.

[0136] s12 convolves the data of different input channels with the corresponding channels of the sth convolutional neural network module, and adds the convolved data of different input channels to obtain the elements of the output feature map.

[0137] The calculation formula of the convolution is as follows:

[0138] Zst =W s ·X t:t+1 +b s (2)

[0139] Where t is the time step. st is the output feature vector of the sth convolutional layer at time step t. s is the weight matrix of the sth convolutional layer. t:t+1 is the local area of ​​the input data X from time step t to t+1. s is the bias term.

[0140] s13 activates the elements of the output feature map through the ReLU function, and the activated feature vector is as follows:

[0141] A st =max(0,Z st ) (3)

[0142] In the formula, A st is the activated feature vector.

[0143] s14 normalizes the activated feature vector to obtain the feature vector output by the convolution layer of the sth convolutional neural network module As shown below:

[0144]

[0145] In the formula, γ st is the scaling parameter. st is the offset parameter. ε is a constant.

[0146] Among them, the mean μ and variance σ of the current iteration batch 2 As shown below:

[0147]

[0148] Where i is the iteration sample number, and n0 is the number of iteration samples.

[0149] s15 takes the feature vector output by the convolutional layer of the sth convolutional neural network module Input the pooling layer to obtain the feature vector output by the sth convolutional neural network module, as shown below:

[0150]

[0151] Where J is the time step position. j is the time step. is the feature vector output by the convolutional layer of the sth convolutional neural network module at time step j. st[J] is the feature vector output by the sth convolutional neural network module at time step J.

[0152] S16 determines whether s=3 is established. If not, the feature vector output by the sth convolutional neural network module is used as input data, s=s+1, and returns to step s12. If yes, the local features of the input data are extracted.

[0153] Embodiment 6:

[0154] A system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases, the main technical content of which is shown in any one of Examples 1 to 5. Further, the steps of characteristic analysis of the pulmonary function evaluation model are as follows:

[0155] s21 constructs two bidirectional LSTM layers, sets u=1, and uses the local features of the input data as input features.

[0156] The number of input channels of the first bidirectional LSTM layer is 4e, and the hidden state dimension is 2e.

[0157] The second bidirectional LSTM layer has an input channel number of 4e and a hidden state dimension of e.

[0158] s22 inputs the input features into the forward input gate of the u-th bidirectional LSTM layer and obtains the forward gate output as shown below:

[0159] z t =softmax(w h ·h t +b h ) (8)

[0160] In the formula, z t is the forward gate output of the u-th bidirectional LSTM layer. h is the coefficient of the u-th bidirectional LSTM layer. h is the weight of the u-th bidirectional LSTM layer.

[0161] Among them, the state of the hidden layer at time t h t As shown below:

[0162] h t =o t ⊙tanh(C t ),C t =f t ⊙(C t-1 )+i t ⊙(g t ) (9)

[0163] In the formula, C t , C t-1are the intermediate parameters at time t and t-1 respectively. ⊙ is the Hadamard product of the matrix.

[0164] Among them, the weight matrices of the forget gate, input gate, new memory gate, and output gate are as follows:

[0165]

[0166] In the formula, f t 、i t , g t , o t are the weight matrices of the forget gate, input gate, new memory gate, and output gate respectively. σ is the Sigmoid function. p t is the input feature. t-1 is the state of the hidden layer at time t-1. f 、w i 、w g 、w o are the input coefficients of the forget gate, input gate, new memory gate, and output gate respectively. f 、v i 、v g 、v o are the hidden coefficients of the forget gate, input gate, new memory gate, and output gate respectively. f 、b i 、b g 、b o They are the weights of the forget gate, input gate, new memory gate, and output gate respectively.

[0167] s23 will output the positive gate z t Back propagate and update the parameters.

[0168] The calculation formula for updating parameters is as follows:

[0169]

[0170] Where E is the loss function. t+1 is the intermediate parameter at time t+1. t+1 is the weight matrix of the forget gate at time t+1. t is the input feature.

[0171] s24 determines whether u=2 holds. If not, the output feature of the u-th bidirectional LSTM layer is used as the input feature, u=u+1, and returns to step s22. If yes, feature analysis data is obtained.

[0172] Embodiment 7:

[0173] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Examples 1 to 6. Further, the steps of feature integration of the pulmonary function evaluation model are as follows:

[0174] s31 is based on the forward gate output z of the bidirectional LSTM layer t , construct two fully connected layers.

[0175] s32 converts the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Input the first fully connected layer and use the Flatten function to convert the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Flatten it into a one-dimensional array, and then use the Linear function to expand the hidden layer state h t The number of features.

[0176] The data output by the first fully connected layer is input into the second fully connected layer. The Flatten function is used to flatten the data output by the first fully connected layer into a one-dimensional array. Then the Linear function is used to expand the feature number of the output data of the first fully connected layer to obtain the hidden layer state h after the expanded feature number. t .

[0177] s33 applies ReLU activation function and Dropout layer to the hidden layer state h after expanding the number of features t Perform feature integration.

[0178] The ReLU activation function is as follows:

[0179] H t =max(0,h t ) (12)

[0180] In the formula, H t To activate the feature.

[0181] The output of the Dropout layer is as follows:

[0182] D(H t )=H t ⊙m (13)

[0183] In the formula, D(H t ) is the output of the Dropout layer. ⊙ is the Hadamard product of the matrix.

[0184] Among them, the random number m is as follows:

[0185]

[0186] Where p is the probability value. i is the random mask function.

[0187] Embodiment 8:

[0188] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Embodiments 1 to 7. Further, the steps of classifying the pulmonary function evaluation model are as follows:

[0189] S41 classifies lung diseases into z1 Build a disease classifier.

[0190] The severity is classified into y z2 Build a severity classifier.

[0191] s42 Dropout layer output D(H t ) is input into the disease classifier and the Linear function is used to z1 The involved diseases are selected from the lung diseases, and then the Softmax activation function is used to obtain the probability distribution of the selected disease types.

[0192] s43 Dropout layer output D(H t ) is input into the severity classifier and the Linear function is used to z2 The involved severity information is selected from the severity levels, and then the Softmax activation function is used to obtain the probability distribution of the selected severity information.

[0193] s44 completes the classification of the patient's disease and severity based on the probability distribution of disease types and the probability distribution of severity information.

[0194] Embodiment 9:

[0195] A system for identifying and evaluating regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Embodiments 1 to 8. Further, the Softmax activation function is as follows:

[0196]

[0197] Where softmax(D) is the Softmax activation function. D is the output of the Dropout layer. n is the ordinal number representing the category, and n represents the number of categories.

[0198] Embodiment 10:

[0199] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Embodiments 1 to 9. Further, the steps of constructing the pulmonary function evaluation model include:

[0200] a1 collects human lung impedance data of different disease types and severities, and constructs training sets and test sets.

[0201] a2 builds an evaluation model.

[0202] a3 takes human lung impedance data as input and disease type and severity as output, and uses the training set to train the evaluation model to obtain a trained evaluation model.

[0203] a4 inputs the human lung impedance data of the test set into the trained evaluation model, and adjusts the hyperparameters of the evaluation model according to the output data to obtain the lung function evaluation model.

[0204] Embodiment 11:

[0205] See also Figures 1 to 10 , a regional pulmonary ventilation disorder disease identification and grade assessment system, including a multi-channel human body impedance acquisition module, a signal processing module, a pulmonary function assessment module, and an assessment result output module.

[0206] The multi-channel human body impedance acquisition module is used to collect impedance change data of different lung regions of the human body. The lung regions include the right upper lobe region, the right middle lobe region, the right lower lobe region, the left upper lobe region, and the left lower lobe region.

[0207] The signal processing module is used to pre-process the lung area impedance change data collected by the multi-channel human body impedance collection module.

[0208] The pulmonary function assessment module stores a pulmonary function assessment model.

[0209] The lung function assessment model is used to perform feature extraction, feature analysis, feature integration and classification on the pre-processed lung regional impedance change data, and assess the type of lung disease and the corresponding severity.

[0210] The evaluation result output module is used to output the type of lung disease evaluated and the corresponding severity.

[0211] Embodiment 12:

[0212] A system for identifying and grading regional pulmonary ventilation disorder diseases, the main technical content of which is shown in Example 11. Furthermore, the multi-channel human body impedance acquisition module includes a constant current source, a biosensor analog front-end chip, and 8 excitation electrodes.

[0213] The biosensor analog front-end chip is equipped with a constant current source.

[0214] The constant current source is used to provide a constant current signal.

[0215] The biosensor analog front-end chip injects the current signal output by the constant current source into the human chest cavity through the excitation electrode, and then collects impedance change data of different lung areas.

[0216] The placement point of the excitation electrode is determined by the following steps:

[0217] Mark the center of the first rib on the midclavicular line of the subject to be tested as position 8, mark the center of the eighth rib as position 5, mark positions 7 and 6 equidistantly on the line between 8 and 5, mark positions 1 to 4 at the intersection of the horizontal positions of 5 to 8 and the right axillary midline, mark positions 9 to 12 at the intersection of the horizontal positions of 5 to 8 and the left axillary midline, and select positions 1, 2, 3, 6, 7, 8, 9, and 10 as electrode placement points.

[0218] First, the bioelectrical impedance measurement technology is used to apply an excitation source to the chest of the subject to be tested, and the lung impedance data corresponding to the subject to be tested in the five measurement channels are obtained. The excitation electrode uses a commercially available disposable silver chloride electrocardiogram electrode with an electrode diameter of 51mm and an AC impedance of 3kΩ or less. The eight electrodes are placed according to Figure 3 Place the electrodes in the pattern shown in the figure, and find the intersection of the center of the first rib (marked as No. 8) and the center of the eighth rib (marked as No. 5) on the mid-clavicular line of the subject to be tested. Mark the positions of electrodes No. 7 and No. 6 equidistantly on the line between No. 8 and No. 5. The intersections of these four horizontal positions with the midline of the right axilla are marked from bottom to top as No. 1 to No. 4, and the intersections with the midline of the left axilla are marked from bottom to top as No. 9 to No. 12. Finally, No. 1, No. 2, No. 3, No. 6, No. 7, No. 8, No. 9, and No. 10 were selected as electrode placement points, and the upper, middle, and lower lobes of the right lung and the upper and lower lobes of the left lung were placed in the same way. Figure 2 As shown in the figure, they are named as one to five partitions, as shown in the figure. Figure 4 shown.

[0219] In order to collect the regional impedance change data of the five lung lobes at any breathing time, the following Figure 5 The multi-channel human body impedance acquisition system shown in the figure is mainly composed of a biosensor analog front end ADS1292, a constant current source, a microcontroller unit (MCU), a switch array module, a serial communication interface and a signal processing unit. The ADS1292 built-in constant current source injects a current excitation with a frequency of 64kHZ and an amplitude of 1mA into the human chest cavity through the excitation electrode, and the lung impedance data under five channels can be obtained.

[0220] Embodiment 13:

[0221] A system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases, the main technical content of which is shown in any one of Examples 11 to 12. Furthermore, the preprocessing includes wavelet filtering processing, wavelet decomposition processing, soft threshold processing, and inverse wavelet transform processing.

[0222] The wavelet filtering process is used to filter the lung area impedance change data to obtain filtered impedance data.

[0223] The wavelet decomposition process uses a multi-layer soft threshold method to set the threshold of each level, and uses the threshold of each level to decompose the filtered impedance data to obtain detail coefficients.

[0224] The soft threshold processing sets detail coefficients below a threshold to zero and shrinks detail coefficients above a threshold to remove noise.

[0225] The inverse wavelet transform process performs inverse wavelet transform on the detail coefficients after soft threshold processing to obtain the pre-processed lung area impedance change data.

[0226] The lung impedance data of the five channels are processed by wavelet filtering, as shown in Figure 6 .

[0227] The impedance data of the five channels were decomposed by wavelet respectively, and the db1 wavelet basis function was selected, and the level of wavelet decomposition was set to 5 layers; then the multi-layer soft threshold method was used to set the threshold, and the threshold of each layer was 800, 700, 600, 500, and 400 respectively. The pywt.wavedec function was used to perform wavelet decomposition on the data to obtain detail coefficients, and the decomposed coefficients were soft-thresholded to set the coefficients below the threshold (usually considered to be noise) to zero, and the coefficients greater than the threshold were shrunk to remove noise; the pywt.waverec function was used to perform inverse wavelet transform on the coefficients after threshold processing to obtain impedance data with the main features retained and the noise weakened, thereby obtaining a data set suitable for model input.

[0228] Embodiment 14:

[0229] A system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases, the main technical content of which is shown in any one of Embodiments 11 to 13. Further, the calculation formula of the wavelet filtering process is as follows:

[0230] Z=sign(x)×max(0,x-x0) (1)

[0231] Where Z is the impedance modulus after filtering, x is the input impedance modulus, x0 is the threshold, sign(x) represents the sign function of x, and max(0,x-x0) means that if x is greater than x0, the output is x, and if x is less than or equal to x0, the output is 0.

[0232] Embodiment 15:

[0233] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Examples 11 to 14. Further, the method for extracting features from the preprocessed lung regional impedance change data by the pulmonary function evaluation model includes a convolutional neural network, and the steps are as follows:

[0234] s11 constructs three convolutional neural network modules and sets s=1.

[0235] The convolutional layer of the first convolutional neural network module has a convolution kernel size of 3, a stride of 1, a padding of 1, 5 input channels, and 32 output channels.

[0236] The convolutional kernel size of the convolutional neural network module of the second convolutional neural network module is 3, the stride is 1, the padding is 1, the number of input channels is 32, and the number of output channels is 64.

[0237] The convolutional kernel size of the convolutional layer of the third convolutional neural network module is 3, the stride is 1, the padding is 1, the number of input channels is 64, and the number of output channels is 128.

[0238] s12 convolves the data of different input channels with the corresponding channels of the sth convolutional neural network module, and adds the convolved data of different input channels to obtain the elements of the output feature map.

[0239] The calculation formula of the convolution is as follows:

[0240] Z st =W s ·X t:t+1 +b s (2)

[0241] Where t is the time step. st is the output feature vector of the sth convolutional layer at time step t. s is the weight matrix of the sth convolutional layer. t:t+1 is the local area of ​​the input data X from time step t to t+1. s is the bias term.

[0242] s13 activates the elements of the output feature map through the ReLU function, and the activated feature vector is as follows:

[0243] A st =max(0,Z st ) (3)

[0244] In the formula, A st is the activated feature vector.

[0245] s14 normalizes the activated feature vector to obtain the feature vector output by the convolution layer of the sth convolutional neural network module As shown below:

[0246]

[0247] In the formula, γ st is the scaling parameter. st is the offset parameter. ε is a constant.

[0248] Among them, the mean μ and variance σ of the current iteration batch 2 As shown below:

[0249]

[0250] Where i is the iteration sample number, and n0 is the number of iteration samples.

[0251] s15 takes the feature vector output by the convolutional layer of the sth convolutional neural network module Input the pooling layer to obtain the feature vector output by the sth convolutional neural network module, as shown below:

[0252]

[0253] Where J is the time step position. j is the time step. is the feature vector output by the convolutional layer of the sth convolutional neural network module at time step j. st [J] is the feature vector output by the sth convolutional neural network module at time step J.

[0254] S16 determines whether s=3 is established. If not, the feature vector output by the sth convolutional neural network module is used as input data, s=s+1, and returns to step s12. If yes, the local features of the input data are extracted.

[0255] Embodiment 16:

[0256] A system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases, the main technical content of which is shown in any one of Examples 11 to 15. Further, the steps of characteristic analysis of the pulmonary function evaluation model are as follows:

[0257] s21 constructs two bidirectional LSTM layers, sets u=1, and uses the local features of the input data as input features.

[0258] The first bidirectional LSTM layer has an input channel number of 128 and a hidden state dimension of 64.

[0259] The second bidirectional LSTM layer has an input channel number of 128 and a hidden state dimension of 32.

[0260] s22 inputs the input features into the forward input gate of the u-th bidirectional LSTM layer and obtains the forward gate output as shown below:

[0261] z t =softmax(w h ·h t +b h ) (8)

[0262] In the formula, z t is the forward gate output of the u-th bidirectional LSTM layer. h is the coefficient of the u-th bidirectional LSTM layer. h is the weight of the u-th bidirectional LSTM layer.

[0263] Among them, the state of the hidden layer at time t h t As shown below:

[0264] h t =o t ⊙tanh(C t ),C t =f t ⊙(C t-1 )+i t ⊙(g t ) (9)

[0265] In the formula, C t , C t-1 are the intermediate parameters at time t and t-1 respectively. ⊙ is the Hadamard product of the matrix.

[0266] Among them, the weight matrices of the forget gate, input gate, new memory gate, and output gate are as follows:

[0267]

[0268] In the formula, f t 、i t , g t , o t are the weight matrices of the forget gate, input gate, new memory gate, and output gate respectively. σ is the Sigmoid function. p t is the input feature. t-1 is the state of the hidden layer at time t-1. f 、w i 、w g 、w o are the input coefficients of the forget gate, input gate, new memory gate, and output gate respectively. f 、v i 、vg 、v o are the hidden coefficients of the forget gate, input gate, new memory gate, and output gate respectively. f 、b i 、b g 、b o They are the weights of the forget gate, input gate, new memory gate, and output gate respectively.

[0269] s23 will output the positive gate z t Back propagate and update the parameters.

[0270] The calculation formula for updating parameters is as follows:

[0271]

[0272] Where E is the loss function. t+1 is the intermediate parameter at time t+1. t+1 is the weight matrix of the forget gate at time t+1. t is the input feature.

[0273] s24 determines whether u=2 holds. If not, the output feature of the u-th bidirectional LSTM layer is used as the input feature, u=u+1, and returns to step s22. If yes, feature analysis data is obtained.

[0274] Embodiment 17:

[0275] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Examples 11 to 16. Further, the steps of feature integration of the pulmonary function evaluation model are as follows:

[0276] s31 is based on the forward gate output z of the bidirectional LSTM layer t , construct two fully connected layers.

[0277] s32 converts the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Input the first fully connected layer and use the Flatten function to convert the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Flatten it into a one-dimensional array of 2048 features, and then use the Linear function to expand the hidden layer state h t The number of features obtained is 128.

[0278] The data output by the first fully connected layer is input into the second fully connected layer. The Flatten function is used to flatten the data output by the first fully connected layer into a one-dimensional array of 2048 features. The Linear function is then used to expand the number of features of the output data of the first fully connected layer to 128. The hidden layer state h after the expanded feature number is obtained. t .

[0279] s33 applies ReLU activation function and Dropout layer to the hidden layer state h after expanding the number of features t Perform feature integration.

[0280] The ReLU activation function is as follows:

[0281] H t =max(0,h t ) (12)

[0282] In the formula, H t To activate the feature.

[0283] The output of the Dropout layer is as follows:

[0284] D(H t )=H t ⊙m (13)

[0285] In the formula, D(H t ) is the output of the Dropout layer. ⊙ is the Hadamard product of the matrix.

[0286] Among them, the random number m is as follows:

[0287]

[0288] Where p is the probability value, which is set to 0.2. i is the random mask function.

[0289] Embodiment 18:

[0290] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Examples 11 to 17. Further, the steps of classifying the pulmonary function evaluation model are as follows:

[0291] S41 classifies lung diseases into z1 Build a disease classifier.

[0292] The severity is classified into y z2 Build a severity classifier.

[0293] The lung diseases include lung cancer, pneumonia, and pulmonary effusion.

[0294] The severity levels include mild, moderate, and severe.

[0295] s42 Dropout layer output D(H t ) is input into the disease classifier and the Linear function is used to z1 The involved diseases are selected from the lung diseases, and then the Softmax activation function is used to obtain the probability distribution of the selected disease types.

[0296] s43 Dropout layer output D(H t ) is input into the severity classifier and the Linear function is used to z2 The involved severity information is selected from the severity levels, and then the Softmax activation function is used to obtain the probability distribution of the selected severity information.

[0297] s44 completes the classification of the patient's disease and severity based on the probability distribution of disease types and the probability distribution of severity information.

[0298] Embodiment 19:

[0299] A system for identifying and evaluating the grade of regional pulmonary ventilation disorder diseases, the main technical content of which is shown in any one of Examples 11 to 18. Further, the Softmax activation function is as follows:

[0300]

[0301] Where softmax(D) is the Softmax activation function. D is the output of the Dropout layer. n is the ordinal number representing the category, and n represents the number of categories.

[0302] Embodiment 20:

[0303] A system for identifying and evaluating the grade of regional pulmonary ventilation disorders, the main technical content of which is shown in any one of Examples 11 to 19. Further, the steps of constructing the pulmonary function evaluation model include:

[0304] a1 collects human lung impedance data of different disease types and severities, and constructs training sets and test sets.

[0305] a2 builds an evaluation model.

[0306] a3 takes human lung impedance data as input and disease type and severity as output, and uses the training set to train the evaluation model to obtain a trained evaluation model.

[0307] a4 inputs the human lung impedance data of the test set into the trained evaluation model, and adjusts the hyperparameters of the evaluation model according to the output data to obtain the lung function evaluation model.

[0308] Embodiment 21:

[0309] See also Figures 1 to 10 , a regional lung ventilation disorder disease identification and grade assessment system, the main technical contents include:

[0310] First, the bioelectrical impedance measurement technology is used to apply an excitation source to the chest of the subject to be tested, and the lung impedance data corresponding to the subject to be tested in the five measurement channels are obtained. The excitation electrode uses a commercially available disposable silver chloride electrocardiogram electrode with an electrode diameter of 51mm and an AC impedance of 3kΩ or less. The eight electrodes are placed according to Figure 3 Place the electrodes in the pattern shown in the figure, and find the intersection of the center of the first rib (marked as No. 8) and the center of the eighth rib (marked as No. 5) on the mid-clavicular line of the subject to be tested. Mark the positions of electrodes No. 7 and No. 6 equidistantly on the line between No. 8 and No. 5. The intersections of these four horizontal positions with the midline of the right axilla are marked from bottom to top as No. 1 to No. 4, and the intersections with the midline of the left axilla are marked from bottom to top as No. 9 to No. 12. Finally, No. 1, No. 2, No. 3, No. 6, No. 7, No. 8, No. 9, and No. 10 were selected as electrode placement points, and the upper, middle, and lower lobes of the right lung and the upper and lower lobes of the left lung were placed in the same way. Figure 2 As shown in the figure, they are named as one to five partitions, as shown in the figure. Figure 4 As shown:

[0311] In order to collect the regional impedance change data of the five lung lobes at any breathing time, the following Figure 5 The multi-channel human body impedance acquisition system shown in the figure is mainly composed of the biosensor analog front end ADS1292, constant current source, microprocessor unit (MCU), switch array module, serial communication interface and signal processing unit. The built-in constant current source of ADS1292 injects the current excitation with a frequency of 64kHZ and an amplitude of 1mA into the human chest cavity through the excitation electrode, and the lung impedance data under five channels can be obtained.

[0312] The lung impedance data of the five channels are processed by wavelet filtering, as shown in Figure 6 , the principle of wavelet filtering of multi-channel impedance data is:

[0313] Z=sign(x)×max(0,xc)(1)

[0314] Where Z is the impedance modulus after filtering, x is the input impedance modulus, c is the threshold, sign(x) represents the sign function of x, and max(0,xc) means that if x is greater than c, the output is x, otherwise the output is 0.

[0315] First, the impedance data of the five channels were decomposed by wavelet, and the db1 wavelet basis function was selected, and the level of wavelet decomposition was set to 5 layers; then the multi-layer soft threshold method was used to set the threshold, and the threshold of each layer was 800, 700, 600, 500, and 400 respectively. The pywt.wavedec function was used to perform wavelet decomposition on the data to obtain detail coefficients. The decomposed coefficients were soft-thresholded to set the coefficients below the threshold (usually considered to be noise) to zero, and the coefficients greater than the threshold were shrunk to remove noise; the pywt.waverec function was used to perform an inverse wavelet transform on the coefficients after threshold processing to obtain impedance data with the main features retained and the noise weakened, thereby obtaining a data set suitable for model input.

[0316] The preprocessed data set is fed into the lung function assessment model for feature extraction, feature analysis, feature integration and classification. First, CNN is used as a filter and feature extractor. The process is as follows Figure 8 , CNN uses a series of convolutional layers and pooling layers to extract and learn features from the input data.

[0317] When passing through the first convolution layer, the Conv1d function is used to convolve the five channel data respectively, with the convolution kernel size of 3, the stride of 1, the padding of 1, the number of input channels of 5, and the number of output channels of 32. For the impedance data Xi of each channel (where i ranges from 1 to 5), we convolve it with the corresponding channel part of the filter, and then add the results of the five channels to obtain an element of the output feature map. This process is repeated at all positions of the input data, and the convolution operation can be expressed as:

[0318] Z 1t =W1·X t:t+1 +b1(2)

[0319] Where W1 is the weight matrix of the first convolutional layer, whose dimension is [32, 5, 3], the bias term b1 is a vector of length 32, and Z 1t is the output feature vector of the first convolutional layer at time step t, X t:t+1 It is the local area of ​​the input data X from time step t to t+1. And the convolution layer output Z 1t After activation by the ReLU function, the activated feature vector is output, and the mathematical expression is:

[0320] A 1t =max(0,Z 1t )(3)

[0321] Then to A 1t Use the BatchNorm1d function for batch normalization, mathematically expressed as:

[0322]

[0323] where μ and σ 2 is the mean and variance of the current batch, ε is a very small constant, its main purpose is to prevent the denominator from being zero, thereby ensuring the stability of numerical calculations, γ 1t and β 1t are learnable scaling and offset parameters. The number of samples used in each iteration is 128, then μ and σ 2 The mathematical expression is:

[0324]

[0325] Finally, the feature vector obtained by passing the data through the first convolution layer Input to the pooling layer, select the maximum value in each pooling window, the window size is 2, the stride is 2, and the mathematical expression is:

[0326]

[0327] Where i is the time step position.

[0328] The output vector is passed through the second convolutional neural network module, the convolutional kernel size of the convolutional layer is 3, the stride is 1, the padding is 1, the number of input channels is 32, and the number of output channels is 64; the convolutional kernel size of the convolutional layer of the third convolutional neural network module is 3, the stride is 1, the padding is 1, the number of input channels is 64, and the number of output channels is 128. The principle is the same as above. The data set is passed through the three convolutional neural network modules mentioned above in turn to extract the local features of the input data.

[0329] The obtained partition impedance data x1 to x5 are used as input features and passed through the first LSTM layer. The specific process is as follows Fig. 9 At time t, the feature is input into the LSTM forward input gate, the number of input channels is 128, and the hidden state dimension of the previous time step is 64, thus obtaining four gates (forget gate w f , input gate w i 、New Memory Gate g and output gate w o ) is (128+64)×64=212×64. According to the following principle formula:

[0330]

[0331] Where σ is the Sigmoid function, ⊙ is the Hadamard product of the matrix, w, v, b are the coefficients and weights of the linear relationship of each gate, and the hidden layer state h at time t is obtained. t The dimension is 64, and the calculation principle is:

[0332] ht =o t ⊙tanh(C t ),C t =f t ⊙(C t-1 )+i t ⊙(g t )(9)

[0333] The final forward gate output is:

[0334] z t =softmax(w h ·h t +b h )(10)

[0335] Since the direction is bidirectional, back propagation begins. The main purpose is to solve the partial derivatives of the loss function E (the function of h0, h1, h2...) with respect to the above weights and update all parameters. The principle is:

[0336]

[0337] After the bidirectional LSTM, the final output channel number is 128 and the hidden state dimension is 64.

[0338] The output vector is passed through the second bidirectional LSTM layer, whose input channel number is 128 and the hidden state dimension is set to 32. According to the above principles and steps, the final output channel is 64 and the hidden state dimension is 32. The data set is passed through the two long short-term memory network modules mentioned above in turn to analyze the characteristics of the data.

[0339] The data z t The feature integration part of the model consists of two fully connected layers, which will receive the hidden state h of the last time step of the LSTM layer. t As input and expand the number of input features. Use the Flatten function to flatten it into a one-dimensional array of 2048 features, and then use the Linear function to get 128 data features.

[0340] Apply ReLU activation function and Dropout layer:

[0341] H t =max(0,h t )(12)

[0342] The probability of dropout p is set to 0.2, and some outputs in the network are randomly set to zero to prevent overfitting. The mathematical principle is:

[0343] D(H t )=Ht ⊙m(13)

[0344]

[0345] Among them, r i It is the random mask function rand().

[0346] The feature D(H) after the Dropout layer t ) are input into the disease classifier and severity classifier respectively, and the Linear function is applied again to obtain the tensors of the three features, and the Softmax activation function is used to obtain the probability distribution of the three features, namely:

[0347]

[0348] The output is two tensors, where y1, y2, and y3 represent the probabilities of lung cancer, pneumonia, and pulmonary effusion, respectively; y4, y5, and y6 represent the three severity classification results, with 1 to 3 representing mild, moderate, and severe. Doctors can obtain more accurate information about the type and severity of the disease from the output results compared to manual evaluation, and they can further determine the treatment plan for the patient based on this information.

[0349] Before the partitioned impedance dataset is fed into the pulmonary function assessment model, the model must be trained first. Following the above steps, a large amount of pulmonary impedance data from five channels of patients is collected and processed in a tertiary hospital. The dataset contains enough cases to cover three types of diseases and three levels of each disease type, and a healthy control group should also be included to provide a baseline reference. The dataset is divided into training set, validation set, and test set in a ratio of 7:1.5:1.5. The training set is used to train the model, the validation set is used to adjust hyperparameters and prevent overfitting, and the test set is used to evaluate the generalization ability of the model. When the training loss is below 0.1, the training accuracy is above 0.95, and the test accuracy is above 0.90, the pulmonary function assessment model is considered to be effective.

[0350] 36,838 samples were randomly collected from tertiary hospitals to train the above lung function assessment model. The training loss, training accuracy, and test accuracy of the classification results are shown in Figure 2. Fig.10 As shown in the figure, when the data set is traversed 20 times, the learning rate is 0.0005, the number of samples used in each iteration is 128, and the running device is CPU, the model training time is 20 minutes, and the training loss, training accuracy, and test accuracy of the classification results are 0.075, 0.973, and 0.912 respectively, which meet the above good effect standards.

[0351] The above-mentioned lung function assessment model is trained and saved, and is used for subsequent intelligent assessment of the patient's lung ventilation disorder according to the above steps, thereby obtaining the identification and grade assessment method of regional lung ventilation disorder diseases of this patent. The process of the whole method is as follows: Figure 1 shown.

Claims

1. A system for identifying and evaluating regional pulmonary ventilation disorders, characterized in that: It includes a multi-channel human body impedance acquisition module, a signal processing module, a lung function evaluation module, and an evaluation result output module; The multi-channel human body impedance acquisition module is used to collect impedance change data of different lung areas of the human body; the lung areas include the right upper lobe area, the right middle lobe area, the right lower lobe area, the left upper lobe area, and the left lower lobe area. The signal processing module is used to pre-process the lung area impedance change data collected by the multi-channel human body impedance collection module; The pulmonary function assessment module stores a pulmonary function assessment model. The lung function assessment model is used to perform feature extraction, feature analysis, feature integration and classification on the pre-processed lung regional impedance change data, and assess the type of lung disease and the corresponding severity; The evaluation result output module is used to output the type of lung disease evaluated and the corresponding severity.

2. A regional pulmonary ventilation disorder disease identification and grade assessment system according to claim 1, characterized in that: The multi-channel human body impedance acquisition module includes a constant current source, a biosensor analog front-end chip, and a plurality of excitation electrodes; The biosensor analog front-end chip has a built-in constant current source; The constant current source is used to provide a constant current signal; The biosensor analog front-end chip injects the current signal output by the constant current source into the human chest cavity through the excitation electrode, and then collects impedance change data of different lung areas; The placement point of the excitation electrode is determined by the following steps: Mark the center of the first rib on the midclavicular line of the subject to be tested as position 8, mark the center of the eighth rib as position 5, mark positions 7 and 6 equidistantly on the line between 8 and 5, mark positions 1 to 4 at the intersection of the horizontal positions of 5 to 8 and the right axillary midline, mark positions 9 to 12 at the intersection of the horizontal positions of 5 to 8 and the left axillary midline, and select positions 1, 2, 3, 6, 7, 8, 9, and 10 as electrode placement points.

3. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 1, characterized in that: The preprocessing includes wavelet filtering, wavelet decomposition, soft thresholding and inverse wavelet transform. The wavelet filtering process is used to filter the lung area impedance change data to obtain filtered impedance data; The wavelet decomposition process uses a multi-layer soft threshold method to set the threshold of each layer, and uses the threshold of each layer to decompose the filtered impedance data to obtain detail coefficients; The soft threshold processing sets detail coefficients below the threshold to zero and shrinks detail coefficients above the threshold to remove noise; The inverse wavelet transform process performs inverse wavelet transform on the detail coefficients after soft threshold processing to obtain the pre-processed lung area impedance change data.

4. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 3, characterized in that: The calculation formula of the wavelet filtering process is as follows: Z=sign(x)×max(0,x-x0) (1) Where Z is the impedance modulus after filtering, x is the input impedance modulus, x0 is the threshold, sign(x) represents the sign function of x, and max(0,x-x0) means that if x is greater than x0, the output is x, and if x is less than or equal to x0, the output is 0.

5. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 1, characterized in that: The method for extracting features from the preprocessed lung area impedance change data by the lung function assessment model includes a convolutional neural network, and the steps are as follows: s11 builds three convolutional neural network modules and sets s=1; The convolution kernel size of the convolutional layer of the first convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is d, and the number of output channels is e, where a, b, c, d, and e are all positive integers; The convolution kernel size of the convolutional layer of the second convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is e, and the number of output channels is 2e; The convolution kernel size of the convolution layer of the third convolutional neural network module is a, the stride is b, the padding is c, the number of input channels is 2e, and the number of output channels is 4e; s12 convolves the data of different input channels with the corresponding channels of the sth convolutional neural network module, and adds the convolved data of different input channels to obtain the elements of the output feature map; The calculation formula of the convolution is as follows: Z st =W s ·X t:t+1 +b s (2) Where t is the time step; Z st is the output feature vector of the sth convolutional layer at time step t; W s is the weight matrix of the sth convolutional layer; X t:t+1 is the local area of ​​the input data X from time step t to t+1; b s is the bias term; s13 activates the elements of the output feature map through the ReLU function, and the activated feature vector is as follows: A st =max(0,Z st ) (3) In the formula, A st is the activated feature vector; S14 normalizes the activated feature vector to obtain the feature vector output by the convolution layer of the sth convolutional neural network module As shown below: In the formula, γ st is the scaling parameter; β st is the offset parameter; ε is a constant; Among them, the mean μ and variance σ of the current iteration batch 2 As shown below: In the formula, i is the iterative sample number, n0 is the number of iterative samples; s15 takes the feature vector output by the convolutional layer of the sth convolutional neural network module Input the pooling layer to obtain the feature vector output by the sth convolutional neural network module, as shown below: Where J is the time step position; j is the time step; is the feature vector output by the convolutional layer of the sth convolutional neural network module at time step j; P st [J] is the feature vector output by the sth convolutional neural network module at time step J; s16 determines whether s=3 is true. If not, the feature vector output by the sth convolutional neural network module is used as the input data, s=s+1, and returns to step s12; if so, the local features of the input data are extracted.

6. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 5, characterized in that: The steps of characteristic analysis of the lung function assessment model are as follows: s21 constructs two bidirectional LSTM layers, and sets u = 1, and uses the local features of the input data as input features; The number of input channels of the first bidirectional LSTM layer is 4e, and the hidden state dimension is 2e; The second bidirectional LSTM layer has an input channel number of 4e and a hidden state dimension of e; s22 inputs the input features into the forward input gate of the u-th bidirectional LSTM layer and obtains the forward gate output as shown below: z t =softmax(w h ·h t +b h ) (8) In the formula, z t is the forward gate output of the u-th bidirectional LSTM layer; w h is the coefficient of the u-th bidirectional LSTM layer; b h is the weight of the u-th bidirectional LSTM layer; Among them, the state of the hidden layer at time t h t As shown below: h t =o t ⊙tanh(C t ),C t =f t ⊙(C t-1 )+i t ⊙(g t ) (9) In the formula, C t , C t-1 are the intermediate parameters at time t and t-1 respectively; ⊙ is the Hadamard product of the matrix; Among them, the weight matrices of the forget gate, input gate, new memory gate, and output gate are as follows: In the formula, f t 、i t , g t , o t are the weight matrices of the forget gate, input gate, new memory gate, and output gate respectively; σ is the Sigmoid function; p t is the input feature; h t-1 is the state of the hidden layer at time t-1; w f 、w i 、w g 、w o are the input coefficients of the forget gate, input gate, new memory gate, and output gate respectively; v f 、v i 、v g 、v o are the hidden coefficients of the forget gate, input gate, new memory gate, and output gate respectively; b f 、b i 、b g 、b o They are the weights of the forget gate, input gate, new memory gate, and output gate respectively; s23 will output the positive gate z t Back propagation and update parameters; The calculation formula for updating parameters is as follows: Where E is the loss function; C t+1 is the intermediate parameter at time t+1; f t+1 is the weight matrix of the forget gate at time t+1; p t is the input feature; s24 determines whether u=2 is true. If not, the output feature of the u-th bidirectional LSTM layer is used as the input feature, u=u+1 is set, and the process returns to step s22. If yes, the feature analysis data is obtained.

7. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 6, characterized in that: The steps of feature integration of the lung function assessment model are as follows: s31 is based on the forward gate output z of the bidirectional LSTM layer t , construct two fully connected layers; s32 converts the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Input the first fully connected layer and use the Flatten function to convert the state h of the hidden layer of the last time step of the bidirectional LSTM layer t Flatten it into a one-dimensional array, and then use the Linear function to expand the hidden layer state h t The number of features; The data output by the first fully connected layer is input into the second fully connected layer. The Flatten function is used to flatten the data output by the first fully connected layer into a one-dimensional array. Then the Linear function is used to expand the feature number of the output data of the first fully connected layer to obtain the hidden layer state h after the expanded feature number. t ; s33 applies ReLU activation function and Dropout layer to the hidden layer state h after expanding the number of features t Perform feature integration; The ReLU activation function is as follows: H t =max(0,h t ) (12) In the formula, H t To activate the feature; The output of the Dropout layer is as follows: D(H t )=H t ⊙m (13) In the formula, D(H t ) is the output of the Dropout layer; ⊙ is the Hadamard product of the matrix; Among them, the random number m is as follows: In the formula, p is the probability value; r i is the random mask function.

8. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 7, characterized in that: The steps of the lung function assessment model for classification are as follows: S41 classifies lung diseases into z1 ,Build a disease classifier; The severity is classified into y z2 ,construct severity classifier; s42 Dropout layer output D(H t ) is input into the disease classifier and the Linear function is used to z1 Select the involved diseases from the lung diseases, and then use the Softmax activation function to get the probability distribution of the selected disease types; s43 Dropout layer output D(H t ) is input into the severity classifier and the Linear function is used to z2 Select the severity information involved from the severity levels, and then use the Softmax activation function to obtain the probability distribution of the selected severity information; s44 completes the classification of the patient's disease and severity based on the probability distribution of disease types and the probability distribution of severity information.

9. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 8, characterized in that: The Softmax activation function is as follows: Where softmax(D) is the Softmax activation function; D is the output of the Dropout layer; i n is the ordinal number representing the category, and n represents the number of categories.

10. A system for identifying and evaluating regional pulmonary ventilation disorders according to claim 1, characterized in that: The steps of constructing the lung function assessment model include: a1 Collect human lung impedance data of different disease types and severity levels to build training sets and test sets; a2 build an evaluation model; a3 uses human lung impedance data as input, disease type and severity as output, and uses the training set to train the evaluation model to obtain a trained evaluation model; a4 inputs the human lung impedance data of the test set into the trained evaluation model, and adjusts the hyperparameters of the evaluation model according to the output data to obtain the lung function evaluation model.

Images