Early esophageal cancer expiration characteristic information acquisition method and system based on sensor array

Abstract

The invention provides an early esophageal cancer expiration characteristic information acquisition method and system based on a sensor array, and relates to the technical field of machine learning, and the method comprises the steps: collecting response data of each volatile organic compound component in an expiration sample, and carrying out curve fitting; extracting a characteristic value corresponding to the response data, a curve fitting parameter and an area under the curve as response characteristics, and performing screening; acquiring physiological features corresponding to the subject, encoding the physiological features, and determining a sample feature data set according to the encoded physiological features and the screened response features; constructing an esophageal cancer prediction model; and performing Gradient SHAP solution on each feature of the sample feature data set to obtain a Shapley value of each feature, determining a contribution degree of each feature in the esophageal cancer prediction model, and determining expiration feature information representing the early esophageal cancer. According to the method, the expiration characteristics representing the early esophageal cancer can be accurately obtained, and the method is used for assisting in judging the early esophageal cancer.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and system for acquiring breath characteristic information of early esophageal cancer based on a sensor array. Background Art

[0002] At present, the commonly used diagnostic method for esophageal cancer is endoscopy supplemented by imaging examination. Endoscopic examination can obtain lesion tissue for pathological diagnosis, which is the "gold standard" for diagnosing esophageal cancer. It also has a high detection ability for early esophageal cancer and can achieve early screening of esophageal cancer to a certain extent. However, this method of esophageal cancer diagnosis also has the following defects: 1) Some patients may feel obvious discomfort due to the entry of the endoscope into the throat. 2) The accuracy of the examination depends on the doctor's experience and technical level, and complications such as puncture and bleeding may occur due to the doctor's misoperation. Therefore, at this stage, there is an urgent need for an auxiliary method for esophageal cancer screening that is more comfortable, easy to operate and has higher detection accuracy. Summary of the invention

[0003] To solve the above problems, the present invention provides a method and system for acquiring breath characteristic information of early esophageal cancer based on a sensor array. The method can accurately acquire breath characteristics characterizing early esophageal cancer to assist in the diagnosis of early esophageal cancer.

[0004] To achieve the above object, the present invention provides the following technical solutions.

[0005] The present invention provides a method for acquiring breath characteristic information of early esophageal cancer based on a sensor array, comprising the following steps: Acquiring response data of various volatile organic compound components in breath samples of esophageal cancer patients and healthy people; the response data is acquired through sensor array collection; Performing curve fitting on the response data, extracting characteristic values ​​corresponding to the response data, curve fitting parameters and the area under the curve as response characteristics; obtaining the corresponding physiological characteristics of the subjects, encoding them, and determining the sample characteristic data set according to the encoded physiological characteristics and the screened response characteristics; A deep neural network model is constructed, and an esophageal cancer prediction model is obtained by training with a sample feature data set; wherein the deep neural network model includes an input layer, a convolutional layer, a maximum pooling layer, a residual block, a global average pooling layer, a fully connected layer, and an output layer, and ResNet-18 is used as a residual block; The Gradient SHAP method was used to solve each feature of the sample feature data set, and the Shapley value of each feature was obtained to determine the contribution of each feature in the esophageal cancer prediction model. The exhaled breath feature information that mainly characterizes early esophageal cancer was determined based on the contribution of each feature.

[0006] Preferably, the step of obtaining the response data of each volatile organic compound component in the breath samples of esophageal cancer patients and healthy people comprises the following steps: Breath samples were collected from esophageal cancer patients and healthy people, and introduced into a gas chromatography-mass spectrometer to obtain data on the content of multiple organic compounds in the breath samples; The statistical Mann-Whitney U test and the omics OPLS-DA analysis method were combined to screen potential esophageal cancer biomarkers with significant differences; Based on esophageal cancer biomarkers, various sensors were selected to construct a sensor array for collecting volatile organic compounds; The response data of each volatile organic compound component in the exhaled breath is collected through the sensor array.

[0007] Preferably, the acquisition of the sample feature data set includes the following steps: Based on esophageal cancer biomarkers, various sensors were selected, including semiconductor sensors and PID sensors; Passing the breath sample into the sensor array and collecting the response data of each sensor, and performing curve fitting on the response data; According to the response characteristics of various sensors, the initial resistance value Ra, resistance change rate RCR, curve fitting parameter CFP and area under the curve AUC of the semiconductor sensor are extracted as characteristic values; the maximum response value R of the PID sensor is extracted max , growth steps Step, curve fitting parameter CFP and area under the curve AUC as characteristic values; The particle swarm optimization algorithm is used to repeatedly screen the feature values ​​extracted by each sensor in the array, and the feature represented by the global optimal example is selected each time; the number of times each feature is selected during the multiple repeated screening processes is recorded, and the feature with a selection number greater than the preset number is taken as the response feature; Physiological characteristics were obtained, including gender, age, smoking status, BMI index, and family medical history; gender and family medical history were considered as binary cases, using 0-1 coding; smoking status included smokers, former smokers, and non-smokers, using one-hot coding; age and BMI index were used as features based on the original values; A sample feature data set was constructed based on the encoded physiological characteristics and the screened response characteristics, and outliers were removed based on the Local Outlier Factor algorithm.

[0008] Preferably, the method of repeatedly screening the characteristic values ​​extracted by each sensor in the array by using a particle swarm optimization algorithm comprises the following steps: A group of particles are randomly initialized, each particle represents a feature group, and each feature group includes multiple feature values ​​extracted by each sensor; each particle has its own position and speed, the position represents the selection of the feature, and the speed represents the change in the feature selection; Evaluate the fitness of each particle to determine the quality of the feature group; For each particle, compare its current local optimal solution with the existing global optimal solution and update its local optimal solution; For the entire group, the solution corresponding to the particle with the best fitness is selected as the global optimal solution; According to the position and velocity of the particle, iteratively update the velocity and position of the particle; Repeat the process until the fitness value converges. The feature group represented by the global optimal particle is the result of feature selection.

[0009] Preferably, the outlier removal based on the Local Outlier Factor algorithm comprises the following steps: A critical value of a local outlier factor is set in advance. For each sample point in the sample feature data set, the average density of the surrounding samples divided by the density of the sample point is calculated as the local outlier factor. Based on the local outlier factor, if it is greater than the set local outlier factor critical value, the sample point is considered to be an abnormal sample point, and if it is less than the set local outlier factor critical value, the sample point is considered to be a normal sample point.

[0010] Preferably, the construction of the deep neural network model comprises the following steps: Build a deep convolutional neural network model, including input layer, convolution layer, maximum pooling layer, residual block, global average pooling layer, fully connected layer and output layer; ResNet-18 is used as the residual block, and skip connections are added after the convolution layer, batch normalization layer, and activation function layer; After the residual block is introduced, the residual mapping between input and output is expressed as follows: ; in, is the residual function, is the target mapping function, is the input; and Add together, the resulting is the final output of the residual block.

[0011] Preferably, a binary cross entropy loss function is used as a loss function during training of the deep neural network model.

[0012] Preferably, obtaining the Shapley value of each feature comprises the following steps: For any feature , the Shapley value is expressed as: ; In the formula, Representative function Medium Features The contribution of is the set of all features, represents a subset of features, Representatives in the collection Remove features , Representation Subset Add features , Representing a collection The number of elements in .

[0013] The present invention also provides a sensor array-based early esophageal cancer breath characteristic information acquisition system, the system comprising: processor; a memory having stored thereon a computer program executable on the processor; Wherein, when the computer program is executed by the processor, the steps of the method for acquiring breath characteristic information of early esophageal cancer based on a sensor array are implemented.

[0014] The present invention also provides a computer-readable storage medium, on which a data processing program is stored. When the data processing program is executed by a processor, the steps of the method for acquiring breath characteristic information of early esophageal cancer based on a sensor array are implemented.

[0015] Beneficial effects of the present invention: The present invention proposes a method for obtaining breath characteristic information of early esophageal cancer based on a sensor array. The present invention detects breath samples by gas chromatography-mass spectrometry and screens esophageal cancer biomarkers based on the statistical Mann-Whitney U test and omics OPLS-DA analysis method. This makes the selection of sensors in the present invention more targeted and the accuracy of esophageal cancer detection higher. The present invention performs data processing on the response value of the sensor array, including feature extraction, normality test, feature selection, and collects and encodes the physiological characteristics of the subject, which makes the features input into the machine learning model have higher specificity. In addition, the present invention constructs a model with high precision and high sensitivity for esophageal cancer diagnosis by introducing a residual block based on ResNet-18 and hyperparameter adjustment, adding regularization and other methods in the convolutional neural network, greatly improving the accuracy of the model and assisting medical staff in diagnosis. Finally, the present invention innovatively uses the SHAP analysis method to accurately and quantitatively describe the contribution and importance of each feature in the model decision-making process, and accurately obtains breath characteristic information for characterizing early esophageal cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of a method for acquiring breath characteristic information of early esophageal cancer based on a sensor array according to an embodiment of the present invention; Figure 2 is a schematic diagram of the response phase of the semiconductor sensor used in the embodiment of the present invention in the test process; Figure 3 is a fitting curve diagram of the response value of the semiconductor sensor response phase adopted in the embodiment of the present invention; Figure 4 It is an overall structure diagram of the Resnet model of an embodiment of the present invention; Figure 5 is a diagram showing the internal structure of a residual block according to an embodiment of the present invention; Figure 6 This is a comparison result diagram of the accuracy of the algorithm training set and test set before and after the introduction of the attention mechanism and residual block in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] Example The present invention proposes a method for obtaining breath characteristic information of early esophageal cancer based on a sensor array. Figure 1 As shown, the following steps are included: Step 1: Based on the gas chromatography analysis results of breath samples from esophageal cancer patients and healthy people, determine the volatile organic compound components in the breath and construct a sensor array for collecting volatile organic compounds.

[0019] Step 2: Pass the breath sample into the sensor array and collect the response data of each sensor, perform curve fitting on the response data, extract the characteristic values ​​corresponding to the response data, and extract the curve fitting parameters and the area under the curve as the response characteristics, and screen them; obtain the physiological characteristics corresponding to the breath sample, and encode them, and determine the sample feature data set based on the encoded physiological characteristics and the screened response characteristics.

[0020] Step 3: Build a deep neural network model and obtain the esophageal cancer prediction model through training with sample feature data sets; the deep neural network model includes input layer, convolution layer, maximum pooling layer, residual block, global average pooling layer, fully connected layer and output layer, and ResNet-18 is used as the residual block.

[0021] Step 4: Perform Gradient SHAP on each feature of the sample feature data set to obtain the Shapley value of each feature and determine the contribution of each feature in the esophageal cancer prediction model; determine the exhaled feature information characterizing early esophageal cancer based on the contribution of each feature, and update the sensor array based on the contribution of each feature.

[0022] In step 1 of the present invention, the volatile organic components in the breath are determined according to the gas chromatography analysis results of the breath samples of esophageal cancer patients and healthy people, and a sensor array for collecting volatile organic compounds is constructed, which includes the following steps: The breath samples of esophageal cancer were collected from patients diagnosed in the Department of Thoracic Surgery of a certain hospital, and the breath samples of healthy people were collected from the patients' accompanying family members, hospital medical staff, etc. A total of 50 breath samples of esophageal cancer patients and 50 breath samples of healthy people were collected.

[0023] Gas chromatography-mass spectrometry was used to measure the content of volatile organic compounds in breath samples. Mann-Whitney U test and omics OPLS-DA analysis method were combined to screen variables and obtain biomarkers with high correlation with esophageal cancer.

[0024] The selection of sensors needs to comprehensively consider factors such as their response performance to VOCs in exhaled breath, the ease of acquisition, and cost. In addition, different types of sensors have their own advantages and disadvantages. Semiconductor sensors are low-cost and respond quickly, but have low response accuracy and obvious cross-response characteristics; optical PID sensors have high response accuracy, rapid response and recovery, but are expensive and require frequent maintenance. The present invention comprehensively selects the above sensors to form an array to achieve the effect of complementary advantages, and at the same time can greatly enhance the array's ability to detect various components in exhaled breath.

[0025] The sample feature data set construction in step 2 of the present invention includes the following steps: S1: To ensure the air tightness of the air chamber and completely exhaust the gas residue of the previous test sample, high-purity air is continuously introduced for 5 minutes before the sample is tested, and the sensor response value is stabilized.

[0026] S2: Connect the test gas sample to the gas chamber inlet through a silicone tube, wait for 50 seconds, then turn on the gas sampling pump to deliver the sample into the test chamber at a flow rate of 1L / min.

[0027] S3: Continue to wait for 70 seconds, and the breath sample will all enter the test chamber.

[0028] S4: High-purity air is introduced at a flow rate of 2 L / min to clean the test chamber. When the sensor response stabilizes to the initial value, the next set of sample tests is performed.

[0029] Based on the actual response situation, the features of each sensor are extracted; the extracted features are optimized and screened through the particle swarm optimization (PSO) algorithm, and then the physiological characteristics of the subjects are encoded. Finally, the Local OutlierFactor (LOF) algorithm is used to analyze and eliminate outliers in the cleaned data.

[0030] In the specific implementation process, a scatter plot of the response value of each sensor is drawn with the response time as the horizontal axis and the response value as the vertical axis. Taking the n-type semiconductor sensor selected in the present invention as an example, the reduced VOCs in the exhaled breath reduce the sensor resistance value, which is manifested as a decrease in the response value. The scatter plot of the response value during the test is shown in FIG. Figure 2 Among them, 0~50s is the initial stage, which reflects the response of the sensor in high-purity air; 50~120s is the response stage, which reflects the response of the sensor to the introduced sample gas; 120s onwards is the recovery stage, which reflects the response recovery of the sensor after the high-purity air is introduced again.

[0031] The initial resistance value (Ra) and the resistance change rate (RCR) are extracted as response characteristics. Ra is the response value of the sensor in the initial stage. In the present invention, the average value of the sensor response value within 0 to 50 seconds is taken. The response value of the n-type semiconductor sensor decreases after the breath sample is passed through, and the minimum value in the response process is recorded as Rg. RCR can be calculated from Ra and Rg by the following formula: ; Considering that the differences in the responses of different breath samples are mainly reflected in the response stage, the response values ​​of the response stage are mapped to the range of 1 to 70 seconds and the scatter plot is redrawn. Figure 3 After comparison, the index parameter R of the third-order multivariate function fitting the scatter plot is2 Generally, the value exceeds 0.99 and the p-value is less than 0.01, so it is considered that the third-order multivariate function It has good fitting for the response value in the response phase. In order to fully reflect the differences between the responses of different breath samples, the curve fitting parameter (CFP) and the area under the curve (AUC) are additionally extracted as features. CFP is the third-order multivariate function Parameters in a , b , c and d ; AUC is the area enclosed by the fitting curve and the horizontal axis.

[0032] The PSO algorithm is used to optimize and screen the extracted features. The specific process is as follows: The PSO algorithm randomly initializes a group of particles, each particle has its own position and speed, and uses a binary 0-1 string with a length equal to the number of input features to represent the position and speed of the particle. The particle position represents the state of the particle, that is, the selection of each input feature. If the value of each binary bit is 0, it means that the feature is not selected, and if it is 1, it means that the feature is selected. The particle speed represents the subsequent movement direction and movement amplitude of the particle, that is, the change in the particle's selection of each feature. If the value of each binary bit is 0, it means that the corresponding feature will be abandoned, and if it is 1, it means that the corresponding feature will be selected. The particle's speed will be affected by the individual optimal solution and the group optimal solution, and its update formula can be expressed as: ; in, It is a particle In the The velocity vector of the generation; is the inertia weight, which is used to adjust the influence of the previous generation velocity of the particle on the current velocity; and are learning factors, representing the influence strength of individual particle cognition (i.e., the influence of individual optimality) and social cognition (i.e., the influence of global optimality); and is A random number between is used to introduce randomness; It is a particle In the The best historical position of the generation (i.e., the individual optimality); The whole particle group is The global best position of the generation (i.e., global optimum).

[0033] The PSO algorithm is used to repeatedly screen the input features 30 times, and the feature selection group represented by the global optimal particle is recorded after each screening. After 30 screenings, the number of times each feature vector is selected in all the input feature vectors is counted. The feature importance is evaluated based on the frequency of feature vector selection.

[0034] In addition, the relevant physiological characteristics of the subjects were considered, including gender, age, smoking status, BMI index and family medical history. Gender and family medical history were considered as binary categories and coded in 0-1; smoking status included smokers, former smokers (quitters), and non-smokers, using one-hot coding; age and BMI index were used as features based on the original values.

[0035] The outlier processing is performed based on the Local Outlier Factor (LOF) algorithm, which includes: pre-setting the neighbor number parameter k, calculating the distance to the k nearest neighbor points for each sample data point, and using this to calculate the density in the neighborhood of the point, that is, the local reachable density; finally, by comparing the local reachable density of each data point with the local reachable density of its neighboring points, the local outlier factor is calculated to evaluate its abnormality.

[0036] Based on the Local Outlier Factor (LOF) algorithm, outliers are processed and the threshold of the local outlier factor is set to divide the data points into normal points and outliers. It is worth noting that the threshold of the local outlier factor here is usually not a fixed value, but a reference value obtained by sorting all input data by local outlier factors or through multiple cross-validation. The Local Outlier Factor (LOF) algorithm treats the marked outliers as abnormal points and removes them to ensure the reliability of the data.

[0037] The data after outlier processing uses 0-1 coding to encode category information, with esophageal cancer patients coded as 1 and healthy people coded as 0. The encoded data set is divided into training set and test set in proportion, of which 75% of the data is the training set for model learning, and 25% of the data is the test set for model parameter tuning and overall performance evaluation.

[0038] In step 3 of the present invention, a deep neural network model based on ResNet-18 is built to evaluate the diagnostic performance of the model for esophageal cancer and to optimize the model, including the following steps: The present invention adopts ResNet-18 as the main model and integrates the self-attention mechanism module. The overall model architecture and diagnostic algorithm logic are as follows: Figure 4 As shown in the figure, the model consists of the following from top to bottom: initial convolutional layer, four residual layers, global average pooling layer, fully connected layer and output layer. The following is a detailed description of each layer:

[0039] The number of input channels of the initial convolution layer is 1, the number of output channels is 64, the convolution kernel size is 7x7, the step size is 1, and there is no padding by default. The batch normalization layer is introduced to accelerate model training and stabilize the model, and ReLU is used as the activation function.

[0040] The initial convolutional layer is followed by a maximum pooling layer with a pooling kernel size of 3x3, a stride of 2, and a padding of 1 to retain important features and effectively prevent overfitting.

[0041] The four residual layers are the key parts to realize the residual neural network. The present invention integrates the self-attention mechanism layer in the residual layer. The detailed internal structure is as follows: Figure 5 As shown in the figure. Each residual layer contains two residual blocks, and each residual block contains two convolutional layers and a self-attention mechanism layer. The following is a description of the specific internal architecture of each residual layer: The two residual blocks in the first residual layer have the same structure, with 64 input and output channels and a stride of 1. The two convolutional layers in the residual block also have the same structure, with a kernel size of 3x3, a stride of 1, and a padding of 1. Both convolutional layers are followed by a batch normalization layer to accelerate training and stabilize the model, using ReLU as the activation function. The skip connection layer connected afterwards is one of the core concepts in the ResNet structure. The skip connection allows the gradient to propagate directly from the subsequent layers to the previous layers, thereby alleviating the gradient vanishing problem in deep neural networks. In addition, since the two previous convolutional layers have the same structure and the same number of input and output channels, the skip connection is an identity mapping that directly adds the input to the output. Finally, a self-attention mechanism layer is added and ReLU is used as the activation function.

[0042] The first residual block inside the second residual layer has 64 input channels, 128 output channels, a 3x3 convolution kernel size, a stride of 2, and a padding of 1. The first residual block still consists of two convolution layers, a batch normalization layer, a jump connection layer, and an attention mechanism layer. Since the number of input channels of the first residual block is not equal to the number of output channels, the jump connection layer uses a convolution with a kernel size of 1x1 and a stride of 2 to adjust the dimension, and then batch normalizes the 128 channels. Finally, the self-attention mechanism layer is still added and ReLU is used as the activation function. The second residual block has 128 input and output channels, and the structure is similar to the first residual block, except that the jump connection layer is an identity mapping.

[0043] The structures of the third and fourth residual layers are similar to those of the second residual layer. The only difference is that the number of input channels of the third residual layer is 128, the number of output channels is 256, and the number of input channels of the fourth residual layer is 256, the number of output channels is 512, so they will not be repeated here.

[0044] The global average pooling layer calculates the average value of all elements in each feature map, thereby converting each feature map into a single value, reducing the number of parameters to reduce the risk of overfitting. The fully connected layer uses the Sigmoid activation function, which is calculated as: ; in, x is the output of the network before the fully connected layer, is the value after activation by the Sigmoid function. From the activation function formula, we can see that the corresponding activation value is in the range of 0 to 1, which can be regarded as the probability value of the binary classification problem.

[0045] In order to make full use of hardware computing resources, smooth gradients, and reduce the impact of single sample noise on parameter updates, a batch update strategy is adopted, and the batch size is set to 20. The AdamW optimizer with an initial learning rate of 0.001 is used, combined with the momentum method and weight decay (L2 regularization) to prevent overfitting, and the learning rate is adaptively adjusted during training. The binary cross entropy loss function is selected to better adapt to the binary classification problem, and the cross entropy between the predicted value and the true label is used to measure the prediction performance of the model. The model continuously adjusts and optimizes the hyperparameters during the gradient update process to achieve better diagnostic performance.

[0046] According to the standard gas collection process, breath samples were collected from 50 esophageal cancer patients and 50 healthy people in a hospital, and the response data was obtained through the sensor array for neural network diagnosis model training. Combined with the clinical needs of esophageal cancer diagnosis, the performance evaluation of the diagnostic method uses five indicators: specificity, sensitivity, precision, accuracy, and F1 score. Figure 6 The figure shows the comparison of the algorithm training set and test set accuracy before and after the introduction of the attention mechanism and residual block. After the introduction of the attention mechanism and residual block, the diagnostic effect is improved accordingly, and the test set accuracy increases from 81% to 86%.

[0047] In step 4 of the present invention, each feature of the sample feature data set is solved by Gradient SHAP to obtain the Shapley value of each feature, and the analysis model is used to calculate the contribution of each feature in the esophageal cancer diagnosis model to clarify the role of each sensor in the overall decision-making process, which specifically includes: The SHAP algorithm assumes that the output of the model is an additive function, and its formula can be expressed as: ; in, is the output of the model, is the benchmark value, usually the average predicted value of all sample data. It is a feature i SHAP value of It is a binary variable. A value of 1 indicates that the calculated feature is in the feature subset and the impact of the feature on the output is considered. A value of 0 indicates that the calculated feature is not in the feature subset and the impact of the feature on the corresponding output is not considered.

[0048] Gradient SHAP is a variant of the SHAP algorithm, which is particularly suitable for explaining deep learning models. The core idea of ​​this method is to calculate the gradient of the input features to estimate the contribution of each feature to the model output. This calculation method combines gradient information with SHAP value theory to provide an efficient and accurate explanation method for complex models. The calculation steps of the GradientSHAP method are:

[0049] Select reference value: Usually the mean, median or a randomly selected portion of the training data is used as the reference value to calculate the marginal contribution of the feature under different states.

[0050] Linear interpolation: Linear interpolation is performed between the input data point and the reference value to generate multiple interpolation samples. The interpolation samples are used to simulate the evolution of the feature from the reference state to the actual input state. The calculation formula for the interpolation samples is as follows: ; in, is the interpolation parameter, and its value range is [0,1].

[0051] Gradient calculation: For each interpolation sample , calculate the gradient of the model output to the input features. Indicated in The gradient at is the model function.

[0052] Calculate the expected gradient: Take the average of the gradients of all the above interpolated samples to get the expected gradient of the feature.

[0053] Calculate the SHAP value: Perform weighted summation on the expected gradient and calculate the SHAP value of the feature. The calculation formula is: ; in, is the gradient of sample interpolation, integrate the gradient, It changes from 0 to 1, indicating the cumulative gradient contribution during the entire interpolation process. The result of the integration is the expected gradient, which is multiplied by the difference between the input feature value and the reference value to obtain the corresponding feature. i In actual calculations, due to the high complexity of gradient integration, the calculation of the expected gradient usually uses the Monte Carlo method for approximate integration.

[0054] Calculate the SHAP value distribution corresponding to each feature to obtain the contribution and importance of each feature in the model prediction process. The absolute value of the SHAP value indicates the contribution to the corresponding prediction result, and the positive or negative value indicates the positive or negative effect on the prediction result. Sort the importance of the features based on the SHAP value, retain the features at the top of the importance ranking, discard the features at the bottom of the ranking and the corresponding sensors, or replace them with new sensors for a second test to obtain the best sensor combination.

[0055] The above is a method for acquiring breath characteristic information of early esophageal cancer based on a sensor array provided by one embodiment of this embodiment. Based on the same idea, this embodiment also provides a corresponding system for acquiring breath characteristic information of early esophageal cancer based on a sensor array. The specific definition of the system for acquiring breath characteristic information of early esophageal cancer based on a sensor array can be found in the definition of the method for acquiring breath characteristic information of early esophageal cancer based on a sensor array above, which will not be repeated here. Each module in the above-mentioned system for acquiring breath characteristic information of early esophageal cancer based on a sensor array can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0056] This embodiment also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A method for acquiring breath characteristic information of early esophageal cancer based on a sensor array is provided.

[0057] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), tape, floppy disk, flash memory or optical memory, etc. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0058] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for acquiring breath characteristic information of early esophageal cancer based on a sensor array, characterized in that: The following steps are involved: Acquiring response data of various volatile organic compound components in breath samples of esophageal cancer patients and healthy people; the response data is acquired through sensor array collection; Perform curve fitting on the response data, extract characteristic values ​​corresponding to the response data, curve fitting parameters and the area under the curve as response characteristics; Acquire and encode the physiological characteristics corresponding to the subject, and determine the sample characteristic data set according to the encoded physiological characteristics and the screened response characteristics; A deep neural network model is constructed, and an esophageal cancer prediction model is obtained by training with a sample feature data set; wherein the deep neural network model includes an input layer, a convolutional layer, a maximum pooling layer, a residual block, a global average pooling layer, a fully connected layer, and an output layer, and ResNet-18 is used as a residual block; The Gradient SHAP method was used to solve each feature of the sample feature data set, and the Shapley value of each feature was obtained to determine the contribution of each feature in the esophageal cancer prediction model. The exhaled breath feature information that mainly characterizes early esophageal cancer was determined based on the contribution of each feature.

2. The method for acquiring breath characteristic information of early esophageal cancer based on sensor array according to claim 1, characterized in that: The step of obtaining the response data of each volatile organic compound component in the breath samples of esophageal cancer patients and healthy people comprises the following steps: Breath samples were collected from esophageal cancer patients and healthy people, and introduced into a gas chromatography-mass spectrometer to obtain data on the content of multiple organic compounds in the breath samples; The statistical Mann-Whitney U test and the omics OPLS-DA analysis method were combined to screen potential esophageal cancer biomarkers with significant differences; Based on esophageal cancer biomarkers, various sensors were selected to construct a sensor array for collecting volatile organic compounds; The response data of each volatile organic compound component in the exhaled breath is collected through the sensor array.

3. The method for acquiring breath characteristic information of early esophageal cancer based on sensor array according to claim 2, characterized in that: The acquisition of the sample feature data set includes the following steps: Based on esophageal cancer biomarkers, various sensors were selected, including semiconductor sensors and PID sensors; Passing the breath sample into the sensor array and collecting the response data of each sensor, and performing curve fitting on the response data; According to the response characteristics of various sensors, the initial resistance value Ra, resistance change rate RCR, curve fitting parameter CFP and area under the curve AUC of the semiconductor sensor are extracted as characteristic values; the maximum response value R of the PID sensor is extracted max , growth steps Step, curve fitting parameter CFP and area under the curve AUC as characteristic values; The particle swarm optimization algorithm is used to repeatedly screen the feature values ​​extracted by each sensor in the array, and the feature represented by the global optimal example is selected each time; the number of times each feature is selected during the multiple repeated screening processes is recorded, and the feature with a selection number greater than the preset number is taken as the response feature; Physiological characteristics were obtained, including gender, age, smoking status, BMI index, and family medical history; gender and family medical history were considered as binary cases, using 0-1 coding; smoking status included smokers, former smokers, and non-smokers, using one-hot coding; age and BMI index were used as features based on the original values; A sample feature data set was constructed based on the encoded physiological characteristics and the screened response characteristics, and outliers were removed based on the Local Outlier Factor algorithm.

4. The method for acquiring breath characteristic information of early esophageal cancer based on sensor array according to claim 3, characterized in that: The method of using the particle swarm optimization algorithm to repeatedly screen the characteristic values ​​extracted by each sensor in the array includes the following steps: A group of particles are randomly initialized, each particle represents a feature group, and each feature group includes multiple feature values ​​extracted by each sensor; each particle has its own position and speed, the position represents the selection of the feature, and the speed represents the change in the feature selection; Evaluate the fitness of each particle to determine the quality of the feature group; For each particle, compare its current local optimal solution with the existing global optimal solution and update its local optimal solution; For the entire group, the solution corresponding to the particle with the best fitness is selected as the global optimal solution; According to the position and velocity of the particle, iteratively update the velocity and position of the particle; Repeat the process until the fitness value converges. The feature group represented by the global optimal particle is the result of feature selection.

5. The method for acquiring breath characteristic information of early esophageal cancer based on sensor array according to claim 3, characterized in that: The outlier elimination based on the Local Outlier Factor algorithm includes the following steps: A critical value of a local outlier factor is set in advance. For each sample point in the sample feature data set, the average density of the surrounding samples divided by the density of the sample point is calculated as the local outlier factor. Based on the local outlier factor, if it is greater than the set local outlier factor critical value, the sample point is considered to be an abnormal sample point, and if it is less than the set local outlier factor critical value, the sample point is considered to be a normal sample point.

6. The method for acquiring breath characteristic information of early esophageal cancer based on sensor array according to claim 1, characterized in that: The construction of the deep neural network model includes the following steps: Build a deep convolutional neural network model, including input layer, convolution layer, maximum pooling layer, residual block, global average pooling layer, fully connected layer and output layer; ResNet-18 is used as the residual block, and skip connections are added after the convolution layer, batch normalization layer, and activation function layer; After the residual block is introduced, the residual mapping between input and output is expressed as follows: ； in, is the residual function, is the target mapping function, is the input; and Add together, the resulting is the final output of the residual block.

7. The method for acquiring breath characteristic information of early esophageal cancer based on sensor array according to claim 1, characterized in that: The binary cross entropy loss function is used as the loss function during training of the deep neural network model.

8. The method for acquiring breath characteristic information of early esophageal cancer based on sensor array according to claim 1, characterized in that: The acquisition of the Shapley value of each feature comprises the following steps: For any feature , the Shapley value is expressed as: ； In the formula, Representative function Medium Features The contribution of is the set of all features, represents a feature subset, Representatives in the collection Remove features , Representation Subset Add features , Representing a collection The number of elements in .

9. A sensor array-based early esophageal cancer breath characteristic information acquisition system, characterized in that: The system comprises: processor; a memory having stored thereon a computer program executable on the processor; Wherein, when the computer program is executed by the processor, the steps of the method for acquiring breath characteristic information of early esophageal cancer based on a sensor array are implemented as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a data processing program, and when the data processing program is executed by the processor, the steps of the method for acquiring breath characteristic information of early esophageal cancer based on a sensor array are implemented as described in any one of claims 1 to 8.

Images