A lung cancer electronic nose data classification method

Abstract

This invention discloses a lung cancer electronic nose data classification method. The method performs linear discriminant analysis on gas sensors collecting different gas data, selects sensors based on their contributions, converts the corresponding sensor's univariate time series into a maximum trajectory matrix image, reconstructs the sample's maximum trajectory matrix image using a spatial information reorganization method, and inputs the reconstructed sample image sequence into a classifier to obtain classification results. The lung cancer electronic nose data classification method of the present invention sorts the gas sensor array by contribution, removing data collected by redundant channels, which helps improve the accuracy of gas sensor data classification while reducing classification time. The method also determines the optimal arrangement of the maximum trajectory images in a sample by maximizing the sum of the inter-frame mutual information of the images, and reconstructs the image sequence in the sample according to the optimal arrangement, thereby improving the classification accuracy of the gas sensor data.

Description

Technical Field

[0001] The present invention relates to sensor data classification technology, and in particular to a classification method for lung cancer electronic nose data. Background Art

[0002] Lung cancer (LC) is one of the most common cancers that threatens human life. It causes 28% of all cancer-related deaths each year, and over 1.6 million people die from lung cancer worldwide each year. However, current lung cancer diagnostic technologies struggle to meet the requirements for reliability, simplicity, affordability, rapidity, and non-invasiveness. Recently, electronic nose technology has been introduced for lung cancer diagnosis. This technology utilizes volatile organic compounds (VOCs) in human breath for lung cancer screening and is characterized by ease of use, high sensitivity, affordability, and non-invasiveness.

[0003] Multivariate time series (MTS) classification is one of the fundamental tasks of lung cancer electronic nose (LC detection E-nose). Over the years, researchers have proposed numerous MTS classification methods. These methods can be categorized into two types: traditional methods and deep learning (DL) methods. Commonly used traditional methods include distance-based methods (e.g., dynamic time warping (DTW)) and feature-based methods (e.g., symbolic Fourier approximation (SFA)). However, these methods require extensive manual feature engineering; more importantly, manual feature engineering often fails to capture the underlying characteristics of MTS. To overcome these shortcomings, numerous DL methods for MTS classification have been proposed, such as TapNet, which uses an attention prototype network to learn latent features. HIVE-COTE 2.0 replaces three of the four classifiers that comprised HIVE-COTE 1.0. Its building blocks include a shapelet-based transform classifier, a convolution-based ROCKET classifier ensemble, a dictionary-based representation (TDE), and a margin-based DrCIF. Among them, DrCIF is an extension of CIF. CIF is a collection of time series tree classifiers built using standard time series features, Catch22 features, and simple summary statistics extracted from phase-related intervals. ROCKET applies a large number of randomly parameterized convolution kernels. When each kernel is applied to a sequence, the maximum value and the proportion of positive values ​​are recorded and concatenated into a feature vector. These DL methods have shown good performance. Nevertheless, the above networks still have limited ability to extract spatiotemporal information in MTS, which may affect the discriminability of the underlying features.

[0004] In recent years, phase-space reconstruction (PSR)-based DL methods have been considered promising approaches for improving the ability of networks to extract spatiotemporal information, as they allow for more detailed nonlinear characterization of signal features. Since time series data is typically nonlinear, nonlinear processing of this data is necessary. PSR-based methods precisely meet this requirement. Consequently, they have been successfully applied to time series classification. These methods use phase-space reconstruction to embed univariate time series (UTS) into phase space to form phase space trajectories. These trajectories are then projected onto a two-dimensional plane to generate a two-dimensional reconstructed phase space image (RPSI). Finally, convolutional neural networks are used to classify the RPSI.

[0005] However, to our knowledge, current PSR-based methods have only been applied to the classification of UTSs. In reality, they can also be applied to MTSs. Because MTSs are composed of UTSs, we only need to transform the UTSs one by one. However, this simplistic approach may result in the loss of spatiotemporal information in MTSs, thus failing to accurately capture the underlying characteristics of MTSs. This is because discriminative latent features may exist between UTSs, not just within a single UTS. A core additional complexity in multivariate time series classification (MTS classification, MTSC) lies in the fact that discriminative features may reside in the interactions between dimensions, not just in the autocorrelations within a single series. Features in UTSs may or may not be phase-correlated; whereas features in MTSs must also consider dimensional correlations; MTSs may or may not be dimensionally correlated. Furthermore, unlike UTSs, MTSs are often very high-dimensional, resulting in generally low classification accuracy for the aforementioned classification methods on MTS datasets. The massive amount of data can obscure discriminative features. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a lung cancer electronic nose data classification method to solve the technical problem of classifying lung cancer electronic nose data and improving the classification accuracy and classification efficiency.

[0007] The lung cancer electronic nose data classification method of the present invention comprises the following steps:

[0008] (1) Perform linear discriminant analysis on D gas sensors that collect different gas data:

[0009] g(F)=W T F (1)

[0010] where F=[f1,f2,…,f i ,…,f D ] T , 1≤i≤D,f i represents the characteristics of gas sensor i, Sample M l =[X1,X2,…,X D ] is a multivariate time series. Sample M l X in i =[x1,x2,…,x d ] T For a univariate time series, the univariate time series X i It is composed of d data collected by the i-th gas sensor in sequence according to the set sampling time interval; the function h(M l ,i) Return sample Ml A univariate time series X in i , the function max(·) returns X i The maximum value among all elements in W; W=[w1,w2,…,w i ,…,w D ] T is the weight vector, where w i The absolute value of represents the contribution of the i-th gas sensor to the linear discriminant analysis; |w i The larger | is, the greater the contribution of the i-th gas sensor; according to |w i |Sort the gas sensors by size.

[0011] (2) Prioritize Contribution The univariate time series collected by each gas sensor is converted into a maximum trajectory matrix image, thus obtaining Zhang maximum trajectory matrix image.

[0012] (3) Reorganize the information described in step (2) by using the spatial information reorganization method The maximum trajectory matrix image includes:

[0013] 1) Calculate the normalized mutual information matrix of the sample:

[0014] Let Θ s ={θ1,θ2,…,θ H×L} is a univariate time series X s The maximum trajectory matrix image obtained by transformation, where θ1, θ2, …, θ H×L Represent the pixel values ​​in the maximum trajectory matrix image, H and L represent Θ s The height and width of Among the maximum trajectory matrix images, any two maximum trajectory matrix images Θ s and Θ t The normalized mutual information of is expressed as:

[0015]

[0016] MI(Θ s ,Θ t ) represents Θ s and Θ t Mutual information between:

[0017]

[0018] Among them, p(θ s ,θ t ) represents θ s and θ t The joint probability mass function between them, and p(θs ) and p(θ t ) represent their marginal probability mass functions respectively; in formula (2), H(Θ s ) and H(Θ t ) represent Θ respectively s and Θ t , entropy; the value of normalized mutual information ranges from 0 to 1, values ​​close to 0 indicate two relatively independent distributions, while values ​​close to 1 indicate consistency of the distributions.

[0019] All the normalized mutual information calculated are formed into a matrix, and the matrix is ​​defined as the normalized mutual information matrix U of the sample j , 1≤j≤n, normalized mutual information matrix U j Included samples All possible NMI values:

[0020]

[0021] 2) Calculate the mutual information matrix of the training set, including:

[0022] The sample set The normalized mutual information matrix of all samples in is added together to obtain a matrix U:

[0023]

[0024] Where n represents the number of samples, For the sample set D tr A sequence pair in , 1≤j≤n, Represents a sample, Represents a sample The label vector of ; For a sample set containing c categories, the label vector is a vector of length c; each element u in the matrix U s,t satisfy:

[0025]

[0026] u s,t The value range is [0,n], and the matrix U is called the mutual information matrix of the training set.

[0027] 3) Calculate the sum of the mutual information between frames, including:

[0028] The sample of The maximum trajectory matrix image constitutes a permutation p i =(Θ r ,Θ s ,…,Θ t), 1≤i≤num, num is the number of permutations, and the channel permutation corresponding to this permutation is (r, s, …, t); the calculation formula for the sum of the mutual information between frames is:

[0029]

[0030] Among them, ψ(q) will p i The qth element in is mapped to its subscript. For example, for the above p i , ψ(1)=r.

[0031] 4) Arrangement filtering:

[0032] The definition above The set of maximum trajectory images is perm_list. The maximum trajectory images are taken from perm_list and arranged. The number of full arrangements is Filter out the inverse of each permutation in the full permutation (i.e., only one of the two permutations that are inverses of each other is retained), and obtain the permutation set filter_P consisting of the remaining permutations. A permutation p i =(Θ r ,Θ s ,…,Θ t ) is defined as:

[0033]

[0034] 5) Solve the optimal arrangement:

[0035] Calculate the sum of the inter-frame mutual information of each permutation in the permutation set filter_P to form a set CI = [I1, I2, ..., I i ,…,I Ium ]; Find the The optimal arrangement of p c The maximum sum of the mutual information between frames is obtained; finally, the maximum trajectory matrix images of all samples are rearranged in the order of the optimal arrangement to obtain a new information representation.

[0036] (4) Input the rearranged sample image sequence into the classifier to obtain the classification result.

[0037] Furthermore, the classifier is a 3D convolutional neural network.

[0038] Beneficial effects of the present invention:

[0039] 1. The lung cancer electronic nose data classification method of the present invention sorts the gas sensor array by contribution, removes the data collected by redundant channels, and is conducive to improving the accuracy of gas sensor data classification while reducing classification time.

[0040] 2. The lung cancer electronic nose data classification method of the present invention determines the optimal arrangement of the maximum trajectory images in the sample by maximizing the sum of the mutual information between image frames, and reorganizes the image sequence in the sample according to the optimal arrangement. Since more mutual information between adjacent frames is helpful in collecting more information related to the classification task, the present invention improves the classification accuracy of gas sensor data by reorganizing spatial information through image sequences.

[0041] 3. The experiments in the examples show that the lung cancer electronic nose data classification method of the present invention outperforms other state-of-the-art methods on lung cancer datasets: its accuracy, AUC, and specificity are the best at 0.8824, 0.9601, and 0.7333, respectively, and the test time is the shortest at 1.0545×10 3 seconds, while sensitivity and training time achieved suboptimal results. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the flow chart of OptiSIR-3D;

[0043] Figure 2 It is a schematic diagram of the structure of the classifier C3D;

[0044] Figure 3 is a parameter The accuracy on the LC data validation set when changing;

[0045] Figure 4 is the mutual information matrix. Figure (a) is the mutual information matrix U of a sample j ,(b) is the mutual information matrix U of the training set. DETAILED DESCRIPTION

[0046] The present invention will be further described below with reference to the accompanying drawings and examples.

[0047] The lung cancer electronic nose data classification method in this embodiment includes the following steps:

[0048] 1) Perform linear discriminant analysis on D gas sensors collecting different gas data:

[0049] g(F)=W T F (1)

[0050] where F=[f1,f2,…,f i ,…,f D ] T , 1≤i≤D,f i represents the characteristics of gas sensor i, Sample M l =[X1,X2,…,X D ] is a multivariate time series. Sample M l X ini =[x1,x2,…,x d ] T For a univariate time series, the univariate time series X i It is composed of d data collected by the i-th gas sensor in sequence according to the set sampling time interval; the function h(M l ,i) Return sample M l A univariate time series X in i , the function max(·) returns X i The maximum value among all elements in W; W=[w1,w2,…,w i ,…,w D ] T is the weight vector, where w i The absolute value of represents the contribution of the i-th gas sensor to the linear discriminant analysis; |w i The larger the | is, the greater the contribution of the i-th gas sensor. Therefore, we can calculate the value of the element w in W. i The absolute value of selects the sensors in descending order.

[0051] A characteristic of sensor arrays is cross-sensitivity, which often results in redundant channels within the sensor array. Data collected by redundant channels typically do not contribute to multivariate time series (MTS) classification and may even reduce classification accuracy. More importantly, the presence of redundant channels significantly increases the dimensionality of the data, leading to a significant increase in the time complexity of classification. Therefore, this step optimizes the sensor array algorithm to remove data collected by redundant channels, which can improve classification accuracy while reducing time complexity.

[0052] 2) Put Contribution First The univariate time series collected by each gas sensor is converted into a maximum trajectory matrix image, thus obtaining The maximum trajectory matrix images are generated. Combining these maximum trajectory matrix images to form an image sequence is a key issue, as different combinations yield different representations of the samples, significantly impacting the classification performance of the 3D convolutional neural network model. To address this, a spatial information reconstructing method for reconstructing the image sequence is proposed in step 3 below.

[0053] 3) Recombining the information in step 2) by using the Spatial Information Recombination (SIR) method The maximum trajectory matrix image includes:

[0054] 1) Calculate the normalized mutual information matrix of the sample:

[0055] Let Θ s ={θ1,θ2,…,θ H×L} is a univariate time series X s The maximum trajectory matrix image obtained by transformation, where θ1, θ2, …, θ H×L Represent the pixel values ​​in the maximum trajectory matrix image, H and L represent Θ s The height and width of Among the maximum trajectory matrix images, any two maximum trajectory matrix images Θ s and Θ t The normalized mutual information of is expressed as:

[0056]

[0057] MI(Θ s ,Θ t ) represents Θ s and Θ t Mutual information between:

[0058]

[0059] Among them, p(θ s ,θ t ) represents θ s and θ t The joint probability mass function between them, and p(θ s ) and p(θ t ) represent their marginal probability mass functions respectively; in formula (2), H(Θ s ) and H(Θ t ) represent Θ respectively s and Θ t , entropy; the value of normalized mutual information ranges from 0 to 1, values ​​close to 0 indicate two relatively independent distributions, while values ​​close to 1 indicate consistency of the distributions.

[0060] All the normalized mutual information calculated are formed into a matrix, and the matrix is ​​defined as the normalized mutual information matrix U of the sample j , 1≤j≤n, normalized mutual information matrix U j Included samples All possible NMI values:

[0061]

[0062] because You can know U j is a symmetric matrix with all elements on its main diagonal being 1. Therefore, we only need to calculate elements to obtain Uj .

[0063] 2) Calculate the mutual information matrix of the training set, including:

[0064] The sample set The normalized mutual information matrix of all samples in is added together to obtain a matrix U:

[0065]

[0066] Where n represents the number of samples, For the sample set D tr A sequence pair in , 1≤j≤n, Represents a sample, Represents a sample The label vector of ; For a sample set containing c categories, the label vector is a vector of length c; each element u in the matrix U s,t satisfy:

[0067]

[0068] u s,t The value range is [0,n], and the matrix U is called the mutual information matrix of the training set.

[0069] 3) Calculate the sum of the mutual information between frames, including:

[0070] The sample of The maximum trajectory matrix image constitutes a permutation p i =(Θ r ,Θ s ,…,Θ t ), 1≤i≤num, num is the number of permutations, and the channel permutation corresponding to this permutation is (r, s, …, t); the calculation formula for the sum of the mutual information between frames is:

[0071]

[0072] Among them, ψ(q) will p i The qth element in is mapped to its subscript. For example, for the above p i , ψ(1)=r.

[0073] 4) Arrangement filtering:

[0074] From step 1) we know U j is a symmetric matrix, Therefore, if two permutations are inverse permutations of each other, their corresponding inter-frame mutual information is the same.

[0075] The definition above The set of maximum trajectory images is perm_list. The maximum trajectory images are taken from perm_list and arranged. The number of full arrangements is In order to reduce the amount of calculation, we can filter out the inverse of each permutation in perm_list (that is, only one of the two permutations that are inverses of each other is retained), and obtain the permutation set filter_P consisting of the remaining permutations. i =(Θ r ,Θ s ,…,Θ t ) is defined as:

[0076]

[0077] 5) Solve the optimal arrangement:

[0078] Calculate the sum of the inter-frame mutual information of each permutation in the permutation set filter_P to form a set CI = [I1, I2, ..., I i ,…,I num ]; Find the The optimal arrangement of p c The maximum sum of the mutual information between frames is obtained; finally, the maximum trajectory matrix images of all samples are rearranged in the order of the optimal arrangement to obtain a new information representation.

[0079] 4) Input the rearranged sample image sequence into a 3D convolutional neural network model, which then outputs the classification results. 3D convolution has the advantage of preserving spatial and temporal information between multiple frames, and the output feature maps also have spatial and temporal relationships.

[0080] The following experiments are conducted to verify the effectiveness of the lung cancer electronic nose data classification method described in this embodiment.

[0081] The lung cancer (LC) dataset collected for the experiment contains 169 breath samples from 169 subjects, 107 of whom are lung cancer patients and 62 are healthy individuals. Therefore, our classification task is a binary classification problem. Each sample in the LC dataset (i.e., a multidimensional time series) contains data collected from 22 channels (i.e., 22 sensors, D = 22, as shown in Table 1). Each channel contains 360 data points (X i =[x1,x2,…,x N ] T,N=360,1≤i≤D), these data points are collected during the sampling phase. We use these 360 ​​data points for our experiments. Therefore, the length of all multidimensional time series is 360.

[0082] The index of all gas sensors (channels) of the electronic nose in the experiment is defined as Index1, see Table 1 for details.

[0083] Table 1 Information of gas sensors in electronic nose

[0084]

[0085]

[0086] After optimizing the sensor array in step 1, the number of sensors is reduced. The remaining sensors are renumbered in ascending order according to Index 1 to obtain Index 2. Table 2 provides an example of the corresponding relationship between Index 1 and Index 2.

[0087] Table 2 Index1 and Index2 corresponding to the sensors selected after optimization

[0088]

[0089] In addition, the sensors in Index2 are related to U j There is a corresponding relationship between the subscripts of the elements in U and the normalized mutual information values ​​of the sensors. For example, Represents the normalized mutual information value of gas sensor 1 and gas sensor 2 of the jth sample, where 1 and 2 are the index values ​​in Index2. s,t is the sum of the NMI between channels s and t for all samples in the test set. In this example, ψ(q) will be arranged into p i The qth element in is mapped to the index of this element.

[0090] Table 3 lists the training set of LC dataset (D tr ) and the test set (D te ) sample size. The training set contains samples used for training and validation. In this experiment, the validation set is divided into a certain proportion (about 10%) from the training set.

[0091] Table 3 LC dataset

[0092]

[0093] For fair comparison, the experiments used the average accuracy of five runs to reduce the bias caused by randomly initialized weights. All experiments were implemented based on the Pytorch deep learning architecture and Sklearn library in Python and run on an NVIDIA TITAN Xp graphics card with 3584 cores and 11GB of global memory. For binary classification tasks, the evaluation indicators include sensitivity (Sens), specificity (Spec), accuracy (Acc), and area under the ROC (Receiver Operating Characteristic) curve (Area Under the ROC Curve, AUC). The experimental results in the following sections are the average of five runs.

[0094] The experiment is conducted on the training set D of the LC dataset. tr The sensor contribution ranking based on the linear discriminant function is performed on the . Table 4 lists the degree of contribution (DoC) of all sensors in descending order.

[0095] Table 4 DoC table calculated by LDA

[0096]

[0097]

[0098] Assume that the number of sensors selected is Then D tr The optimal value of this parameter is determined by the validation set accuracy on . Unlike direct or stepwise selection methods, our method starts with an empty set and then adds sensors one by one in ascending order of "Rank" in Table 4 until all sensors are added. Each time a sensor is added, the features of all selected sensors are input into a classifier (Support Vector Machine (SVM) can be used here because of its low time complexity) for classification. This method records the change in validation set accuracy when adding sensors and selects the number of sensors as the parameter when the first validation accuracy reaches the highest value. The optimal value of . Figure 3 The figure shows the change in accuracy on the LC data validation set when different numbers of sensors are selected. Figure 3 ,we will If set to 8, only the top 8 sensors in "Rank" will be selected for subsequent classification.

[0099] We evaluated the effectiveness of sensor contribution ranking based on the linear discriminant function on the LC dataset through ablation experiments. The sensors used in the ablation experiments were selected based on their contribution ranking, and their indexes were reset to Index2. Details are shown in Table 2. The classifier used was C3D. The classifier hyperparameter settings are listed in Table 5.

[0100] Table 5 Classifier hyperparameter settings

[0101]

[0102] Table 6 compares the classification performance of the algorithm for optimizing the sensor array proposed in step (1) (named the SAO algorithm) and the algorithm without the SAO algorithm. It should be noted that for convenience, the optimized sensors are still arranged in ascending order according to Index2. The test time is generally evaluated based on the entire test set rather than a single sample. The experimental results show that after applying the SAO algorithm, the classification accuracy is improved by more than 2 percentage points, while the training and test time are reduced by 58% and 43% respectively. This improvement in accuracy indicates that the data collected by redundant channels has been successfully removed. Therefore, due to the removal of redundant data, the training and test time of the classifier are shortened. This conclusion further shows that our SAO algorithm can significantly improve the classification performance of MTS.

[0103] Table 6 Performance comparison of applying SAO algorithm and not applying SAO algorithm

[0104]

[0105] Experiments were designed to evaluate the impact of the SIR algorithm on MTS classification performance. In these experiments, the original LC dataset (Raw MTS) was first processed by the SAO algorithm. The optimized MTS was then converted into a maximum trajectory matrix image sequence (MTMIs). Finally, the SIR method was used to reconstruct the MTMIs into an image sequence for classification. Figure 4 Shows the normalized mutual information matrix U of a sample calculated by the SIR algorithm j And the mutual information matrix U of the training set. The numbers on the main diagonal of the matrix represent the number of samples, U j =1 and U = n (n is the number of samples in the training set, n = 152). In other words, the NMI value of an image with itself is 1. The darker the color, the larger the NMI value of the maximum trajectory matrix image (MTMI) of the two sensors.

[0106] We choose C3D as the 3DCNN classifier. As a general 3DCNN, C3D is efficient, compact, and easy to use. Given the limitations of computing resources, we chose the lightweight C3D model. The network structure of this model is as follows: Figure 2 As shown, the network consists of five convolutional layers, five max pooling layers, and two fully connected layers, culminating in a softmax output layer. All 3D convolution kernels are 3×3×3 in size, with a stride of 1 in both the spatial and temporal dimensions. The number of filters is indicated in each box. The 3D pooling layers from Pooling Layer 1 to Pooling Layer 3 are shown. All pooling kernels are 2×2×2, except for Pooling Layer 1, which is 1×2×2. Each fully connected layer has 2048 output units.

[0107] In order to evaluate the effectiveness of the lung cancer electronic nose data classification method described in the embodiment on the LC dataset, the hyperparameters of C3D remain unchanged in all experiments, as shown in Table 5. Since the different arrangements of the maximum trajectory matrix images are the only determining factor for the representation of different information in this lung cancer electronic nose data classification method, it is only necessary to evaluate the impact of different arrangements on the MTS classification performance. This experiment uses three typical arrangements C1, C2 and C3. C1 represents the sorting according to Rank in the DoC table (Table 3), C2 represents the sorting in ascending order of Index2 (that is, C2 is the result of sorting the C1 index in ascending order of Index1), and C3 is the optimal sorting. Table 7 shows the three arrangements (the elements in the arrangement here are all Index2) and their corresponding SIFMI values ​​and classification accuracy Acc.

[0108] Table 7 Comparison of classification performance using different arrangements

[0109]

[0110] Different permutations represent different information. We used SIFMI as a metric to compare three different permutations. As shown in Table 7, C1 achieved the lowest accuracy, 0.7529. This corresponds to a permutation with a relatively low SIFMI. C3, on the other hand, achieved the highest accuracy, 0.8824, corresponding to a SIFMI of 899.38, the highest of all permutations. This experiment confirms that increasing the sum of inter-frame mutual information can improve model accuracy.

[0111] To further demonstrate the effectiveness of the lung cancer electronic nose data classification method described in this example, we compared the method (named OptiSIR-3D) with other advanced MTSC methods, such as TapNet, CIF, DrCIF, HIVECOTE 2.0, and ROCKET. These models have demonstrated excellent performance on many datasets, so we selected them for comparison. The comparative experiments were conducted on the LC dataset. To reduce bias introduced by random initialization and optimization, we ran each method five times. Table 8 shows the average results of these five runs.

[0112] Table 8 Classification performance comparison with other different MTSC methods

[0113]

[0114] As can be seen from Table 8 above, many multivariate time series classification (MTS Classification, MTSC) methods have also achieved very competitive results. For example, ROCKET and HIVECOTE 2.0 achieved the best and second best accuracy of 0.8824 and 0.8353 respectively on the LC dataset, and their sensitivity was 1. However, the higher accuracy of HIVECOTE 2.0 comes at the cost of training time, which is 13 times that of OptiSIR-3D mentioned in this embodiment and 175 times that of the shortest method ROCKET. Despite this, OptiSIR-3D performed best in test time (inference time on the entire test set), taking less than 20% of the second-ranked DrCIF, while HIVECOTE 2.0 had the longest test time of all methods, which was 9.4720×10 4 seconds. OptiSIR-3D was able to respond quickly after model training, although its training time was slower than ROCKET. Both TapNet and OptiSIR-3D achieved the best specificity, while CIF and DrCIF performed similarly on the LC dataset. Furthermore, OptiSIR-3D tied with ROCKET for first place in accuracy and achieved the best AUC value among all MTSC methods, while its sensitivity was second best at 0.9636.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A lung cancer electronic nose data classification method, characterized by: Including steps: (1) Perform linear discriminant analysis on D gas sensors that collect different gas data: g(F)=W T F (1) where F=[f1,f2,…,f i ,…,f D ] T , 1≤i≤D,f i represents the characteristics of gas sensor i, Sample M l =[X1,X2,…,X D ] is a multivariate time series; sample M l X in i =[x1,x2,…,x d ] T For a univariate time series, the univariate time series X i It is composed of d data collected by the i-th gas sensor in sequence according to the set sampling time interval; the function h(M l ,i) Return sample M l A univariate time series X in i , the function max(·) returns X i The maximum value among all elements in W; W=[w1,w2,…,w i ,…,w D ] T is the weight vector, where w i The absolute value of represents the contribution of the i-th gas sensor to the linear discriminant analysis; |w i The larger | is, the greater the contribution of the i-th gas sensor; according to |w i |Sort the gas sensors by size; (2) Prioritize Contribution The univariate time series collected by each gas sensor is converted into a maximum trajectory matrix image, thus obtaining Zhang maximum trajectory matrix image; (3) Reorganize the information described in step (2) by using the spatial information reorganization method The maximum trajectory matrix image includes: 1) Calculate the normalized mutual information matrix of the sample: Let Θ s ={θ1,θ2,…,θ H×L } is a univariate time series X s The maximum trajectory matrix image obtained by transformation, where θ1, θ2, …, θ H×L Represent the pixel values ​​in the maximum trajectory matrix image, H and L represent Θ s The height and width of Among the maximum trajectory matrix images, any two maximum trajectory matrix images Θ s and Θ t The normalized mutual information of is expressed as: MI(Θ s ,Θ t ) represents Θ s and Θ t Mutual information between: Among them, p(θ s ,θ t ) represents θ s and θ t The joint probability mass function between them, and p(θ s ) and p(θ t ) represent their marginal probability mass functions respectively; in formula (2), H(Θ s ) and H(Θ t ) represent Θ respectively s and Θ t, The entropy of the normalized mutual information ranges from 0 to 1, with values ​​close to 0 indicating two relatively independent distributions and values ​​close to 1 indicating consistency of the distributions. All the normalized mutual information calculated are formed into a matrix, and the matrix is ​​defined as the normalized mutual information matrix U of the sample j , 1≤j≤n, normalized mutual information matrix U j Included samples All possible NMI values: 2) Calculate the mutual information matrix of the training set, including: The sample set The normalized mutual information matrix of all samples in is added together to obtain a matrix U: Where n represents the number of samples, For the sample set D tr A sequence pair in Represents a sample, Represents a sample The label vector of ; For a sample set containing c categories, the label vector is a vector of length c; each element u in the matrix U s,t satisfy: u s,t The value range is [0,n], and the matrix U is called the mutual information matrix of the training set; 3) Calculate the sum of the mutual information between frames, including: The sample of The maximum trajectory matrix image constitutes a permutation p i =(Θ r ,Θ s ,…,Θ t ), 1≤i≤num, num is the number of permutations, and the channel permutation corresponding to this permutation is (r, s, …, t); the calculation formula for the sum of the mutual information between frames is: Among them, ψ(q) will p i The qth element in is mapped to its subscript; 4) Arrangement filtering: The definition above The set of maximum trajectory images is perm_list. The maximum trajectory images are taken from perm_list and arranged. The number of full arrangements is Filter out the inverse of each permutation in the full permutation, that is, only keep one of the two permutations that are inverses of each other, and obtain the permutation set filter_P consisting of the remaining permutations; a permutation p i =(Θ r ,Θ s ,…,Θ t ) is defined as: 5) Solve the optimal arrangement: Calculate the sum of the inter-frame mutual information of each permutation in the permutation set filter_P to form a set CI = [I1, I2, ..., I i ,…,I num ]; Find the The optimal arrangement of p c The maximum sum of the mutual information between frames is obtained; finally, the maximum trajectory matrix images of all samples are rearranged in the order of the optimal arrangement to obtain a new information representation; (4) Input the rearranged sample image sequence into the classifier to obtain the classification result.

2. The lung cancer electronic nose data classification method according to claim 1, characterized in that: The classifier is a 3D convolutional neural network.

Images