A method and system for early screening of asthma based on a two-way airflow sensor

Abstract

An asthma early screening method and system based on a two-way airflow sensor, first collect airflow signals and construct multiple example time sequence data packets, then calculate an adaptive collapse angle based on global error, then construct a double-flow feature input, extract and aggregate features from the flow rate-volume curve based on a gated attention mechanism, finally perform feature fusion and classification prediction, output the abnormal confidence score of the subject's breathing waveform, and set a classification threshold, at which time, output an early warning label, and intuitively display the classification results through an interactive display module. The problem of incomplete physical feature extraction and low signal classification reliability caused by discarding signal variability information and ignoring microscopic morphological changes when processing respiratory airflow signals in the prior art is solved.

Description

Technical Field

[0001] This invention belongs to the field of respiratory signal processing technology, and in particular relates to a method and system for early asthma screening based on a dual-path airflow sensor. Background Technology

[0002] In traditional physiological parameter testing, lung function testing using airflow sensors is a crucial method for obtaining respiratory dynamics data. During the testing process, subjects are typically required to perform multiple forced exhalations to obtain multiple respiratory airflow curves. However, current technologies still have the following shortcomings in automating the processing and feature extraction of these physical signals: 1. The utilization rate of associated time-series signals is low, and there is a lack of ability to characterize signal variability. Existing technologies often only select the airflow curve with the best numerical performance for analysis, or simply average the results of multiple exhalations, ignoring the differences between multiple exhalations of the same subject. In fact, the signal instability during the respiratory process itself contains important physical significance. Traditional processing methods result in the loss of a large number of physical parameters of variability between multiple sets of associated time-series signals, making it impossible to fully characterize the dynamic respiratory features of the subject.

[0003] 2. Limited ability to identify microscopic morphological anomalies. Existing technologies mainly focus on macroscopic statistical indicators such as airflow magnitude and population morphology, while paying insufficient attention to subtle morphological changes such as local collapses and bends in the descending branch of the airflow-volume curve, resulting in limited accuracy in identifying the evolution of respiratory tract physical morphology.

[0004] 3. Lack of robust processing mechanisms for non-standard signals: In actual data acquisition scenarios, differences in subject cooperation often lead to non-standard waveforms or interference signals containing a large amount of noise. Traditional deep learning models often directly perform feature mapping on the entire dataset, lacking the ability to automatically identify effective signal segments and suppress low-quality waveforms. This can easily cause the final feature extraction results to be skewed by abnormal interference signals, thereby reducing the reliability of the results.

[0005] Therefore, there is an urgent need for a signal analysis and feature extraction method that can make full use of multiple respiratory airflow data, take into account both microscopic morphology and overall stability, and possess high robustness. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for early asthma screening based on a dual-path airflow sensor, in order to solve the problems of incomplete physical feature extraction and low signal classification reliability caused by discarding signal variability information and ignoring microscopic morphological changes when processing respiratory airflow signals in the prior art.

[0007] The technical solution adopted in this invention is: 1. A method for early asthma screening based on a dual-path airflow sensor, the steps of which are as follows: Step S1: Acquire airflow signals and construct a multi-sample time-series data package; Step S2: Calculate the adaptive collapse angle based on the global error; Step S3: Construct a dual-stream feature input, which includes a waveform stream input and a statistical stream input; Step S4: Based on the gating attention mechanism, feature extraction and aggregation are performed on the velocity-volume curve; Step S5: Feature fusion and classification prediction, outputting the abnormal confidence score of the subject's respiratory waveform. And set classification thresholds. ,when At that time, a warning label will be output.

[0008] Furthermore, the specific steps of S1 are as follows: S11, Collect data continuously from the subject during a single testing task. The airflow signal generated by the forceful exhalation action is collected, and the raw airflow signal is preprocessed and spatially aligned. S12 transforms the airflow signal processed by S11 from the time domain to the velocity-volume domain. First, the start and end points of exhalation are located. Then, the velocity during the exhalation phase is integrated and accumulated to obtain the volume, thus obtaining the forced vital capacity (FVC) of a single exhalation. Then, the accumulated volume is divided by FVC to construct a normalized volume axis, obtaining a volume percentage sequence in the range [0,1]. A value is then set on this normalized volume axis. The flow velocity value at each of the three equally spaced sampling points is obtained by linear interpolation. Divide the flow rate value by the preset physiological limit flow rate benchmark value to obtain the flow rate-volume curve of a single exhalation action; S13, the same subject The velocity-volume curve is encapsulated as a multiple instance set, denoted as , ;in, It is the first A flow velocity-volume curve, Indicates the total number of sampling points. It represents the set of real numbers.

[0009] Furthermore, the specific steps of S2 are as follows: S21, define the search interval in the descending branch of the velocity-capacity curve. ; S22, traverse each sampling point within the search interval. This is considered as a potential morphological inflection point for each sampling point. Using this sampling point as the dividing point, the descent branch data is divided into the left data segment. and the right data segment Linear regression fitting was performed on each, and the total fitting error was calculated. Then select the option that minimizes the total fitting error. The point with the minimum value is taken as the global optimal shape inflection point, and the slope of the best fitted line on both sides of that point is obtained. , The collapse angle is calculated, and the formulas for calculating the total fitting error and the collapse angle are as follows:

[0010]

[0011] In the formula, Left data segment The Middle The actual flow velocity value at each sampling point The right-hand data segment The Middle The actual flow velocity value at each sampling point Indicates the left data segment Index of sampling points Indicates the right-hand data segment Index of the sampling points; This indicates that the fitted line on the left is in the left data segment. The predicted flow velocity values ​​at each sampling point, where the subscript 1 indicates the fitted straight line on the left; This indicates that the fitted line on the right side is in the data segment on the right side. The predicted flow velocity values ​​at each sampling point, where the subscript 2 indicates the second fitted line on the right; Indicates the angle of collapse. Represents the arctangent function. This indicates taking the absolute value. It represents pi (π).

[0012] Furthermore, the method for determining the dual-stream feature input described in S3 is as follows: 1) Waveform stream input: A multi-instance set constructed by S1 ; 2) Statistical Stream Input: Constructing Statistical Feature Vectors It is composed of dynamic physiological statistical features and static demographic features in sequence; the dynamic physiological statistical features include the mean, standard deviation, minimum and maximum values ​​of peak expiratory flow rate, small airway tangent angle, mid-segment average flow rate angle and collapse angle, totaling 20 dimensions; the static demographic features include age, sex, height and weight, totaling 4 dimensions.

[0013] Furthermore, the specific method for extracting the dynamic physiological statistical features is as follows: a) Peak expiratory flow rate: the maximum value in the flow-volume curve; b) Collapse angle: calculated from S2; c) Small airway tangent angle: Locate the 50% and 75% positions of the normalized volume axis respectively, set a local micro window, perform linear regression fitting on the flow velocity points within the window, and calculate the inclination angle of the fitted line. d) Mid-section average velocity angle: the angle corresponding to the slope of the line connecting the 25% and 75% points of the volume.

[0014] Furthermore, in S4, the multiple instance set is first... Each velocity-volume curve in Independent input to a one-dimensional convolutional neural network, outputting the corresponding high-dimensional instance feature vectors. Subsequently, a gated attention mechanism was introduced to calculate the contribution weight of each velocity-volume curve. Then, based on the calculated contribution weights, a weighted patient-level waveform feature representation is obtained. Simultaneously, the statistical feature vector constructed from the statistical flow in S3 The input is a multilayer perceptron (MLP), and the macroscopic physiological feature representation is obtained through feature mapping. The calculation formulas for contribution weight and patient-level waveform feature representation are as follows:

[0015]

[0016] in, , This represents two distinct learnable projection matrices. Used to map high-dimensional instance feature vectors to a low-dimensional space. Used to map high-dimensional instance feature vectors to a gated space. Represents the weight vector. For element-wise multiplication, This represents the hyperbolic tangent activation function. Indicates activation function This indicates transpose.

[0017] Furthermore, the one-dimensional convolutional neural network includes three one-dimensional convolutional layers with kernel sizes of 7, 5, and 3, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation layer. The first and second one-dimensional convolutional layers are followed by a max pooling layer after their ReLU activation layers, and the third one-dimensional convolutional layer is followed by an adaptive global average pooling layer after its ReLU activation layer. The output is the same as the first... Flow velocity-volume curve Corresponding high-dimensional instance feature vector .

[0018] Furthermore, in S5, the patient-level waveform features are represented. Representation of macroscopic physiological characteristics The data are spliced ​​together to form a joint feature representation that includes morphological, physiological, and variability information. This joint feature representation is then input into the classification network.

[0019] Furthermore, the classification network consists of a three-layer fully connected structure, wherein the first fully connected layer is followed by a batch normalization layer, a ReLU activation layer and a random deactivation layer in sequence, the second fully connected layer is followed by a ReLU activation layer, and the third fully connected layer is followed by an output layer.

[0020] This invention also provides an early asthma screening system based on a dual-path airflow sensor, used to implement an early asthma screening method based on a dual-path airflow sensor, comprising the following modules: Data acquisition module: used to acquire respiratory airflow signals generated by the subject's multiple forced exhalations during a single testing task, and send the airflow signals to the processing unit; Processor: Used to store and execute computer programs to implement the above-mentioned early asthma screening method based on dual-path airflow sensors; Interactive display module: Used to intuitively display classification results.

[0021] The beneficial effects of this invention are: 1. This invention, through the construction of multiple instance sets and dual-flow collaborative analysis, can effectively capture the weak and unstable feature evolution in respiratory signals, thereby improving the utilization rate of weak physical features.

[0022] 2. This invention utilizes a gating attention mechanism, which can automatically weaken the impact of low-quality data caused by poor operation coordination on classification results, improve the robustness of the system in real data collection scenarios, and has strong anti-interference and fault tolerance capabilities.

[0023] 3. This invention reduces the reliance on external auxiliary parameters. Even without inputting demographic parameters such as height and weight, it can still maintain extremely high recognition accuracy by relying solely on the automated analysis of waveform geometric features. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1This is the algorithm flowchart of the present invention.

[0026] Figure 2 This is a structural block diagram of the early asthma screening system of the present invention.

[0027] Figure 3 This is a comparison chart of the ROC performance of the model of this invention and the baseline model of the prior art on the test set.

[0028] Figure 4 This is the MIL attention heatmap output by this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] An embodiment of the present invention provides a method for early asthma screening based on a dual-path airflow sensor, the flowchart of which is shown below. Figure 1 As shown, the steps are as follows: Step S1: Acquire airflow signals and construct a multi-sample time-series data package to obtain standardized respiratory data, preserving variability information between multiple breaths by the subject, as detailed below: S11, using a portable airflow sensor to collect data continuously collected by the subject during a single testing task. One forceful exhalation action ( The generated airflow signal is preprocessed by using a Gaussian filter for smoothing and adjusting the standard deviation of the filter kernel. Preferably, [1.0, 2.0]) is spatially aligned to suppress high-frequency electromagnetic noise and airflow turbulence from the sensor.

[0031] S12 transforms the airflow signal processed by S11 from the time domain to the velocity-volume domain. First, it locates the start and end points of exhalation. Then, it integrates and accumulates the velocity during the exhalation phase to obtain the volume, thus obtaining the forced vital capacity (FVC) of a single exhalation. Then, it divides the accumulated volume by the FVC to construct a normalized volume axis, obtaining a volume percentage sequence in the range [0,1], i.e., 0 to 100% FVC. A setting is then set on this normalized volume axis. In this embodiment, a set of equally spaced sampling points is preferred. =400, the flow velocity value at each point is obtained through linear interpolation. The flow rate value is divided by a preset physiological limit flow rate benchmark to achieve uniform scaling and mapping to a dimensionless numerical range, thereby obtaining the flow rate-volume curve for a single exhalation. In this invention, the number of flow rate-volume curves is the same as the number of consecutive exhalation actions performed by the subject in a single testing task.

[0032] The method for determining the exhalation start point and end point is as follows: First, the pre-processed airflow signal is baseline corrected, and the average airflow value in the resting segment is used as the zero flow velocity reference. When multiple consecutive sampling points of the corrected airflow signal are greater than a preset positive threshold (in this embodiment, the preset positive threshold is preferably 0.1), it is determined to be the exhalation start point. When the corrected airflow signal falls back and enters the negative value range after exhalation is completed, it is determined to be the exhalation end point.

[0033] The physiological limit flow rate benchmark value is a preset global normalization benchmark, which is set based on the sensor range, the empirical upper bound of the maximum expiratory flow rate of the target population, and the model's unified normalization requirements. In this embodiment, it is preferably 14.0, and the model training stage is consistent with the actual deployment and use stage.

[0034] S13, the same subject The velocity-volume curve is encapsulated as a multiple instance set, denoted as , .in, It is the first The flow velocity-volume curve represents the first... The two-dimensional feature tensor of each exhalation is composed of the flow-volume curve of a single exhalation after normalization resampling, and the normalized volume position code. Each feature tensor contains two channels: channel 1 is the normalized flow velocity value, and channel 2 is the normalized volume position code. Indicates the total number of sampling points. Representing the set of real numbers, this step fully preserves the morphological consistency and operational variability of the subject in multiple breath-blowing attempts.

[0035] Step S2: Calculate the adaptive collapse angle based on global error. For the subtle morphological changes in the descending branch of the velocity-volume curve, an adaptive search method based on minimizing global fitting error is used to calculate the collapse angle, replacing the traditional local tangent method, thereby improving noise resistance. The specific steps are as follows: S21, define the search interval in the descending branch of the velocity-capacity curve. The preferred range is from 15% of the volume after the peak flow point (PEF) to 90% of the volume before the end of exhalation, in order to eliminate interference from the nonlinear regions at the beginning and end.

[0036] S22, traverse each sampling point within the search interval. This is considered as a potential morphological inflection point for each sampling point. Using this sampling point as the dividing point, the descent branch data is divided into the left data segment. and the right data segment Linear regression fitting was performed on each, and the total fitting error was calculated. Then select the option that minimizes the total fitting error. The point with the minimum value is taken as the global optimal shape inflection point, and the slope of the best fitted line on both sides of that point is obtained. , The collapse angle is calculated, and the smaller the value of this angle, the more obvious the collapse trend in the middle section of the descending airflow. The formulas for calculating the total fitting error and the collapse angle are as follows:

[0037]

[0038] In the formula, Left data segment The Middle The actual flow velocity value at each sampling point The right-hand data segment The Middle The actual flow velocity value at each sampling point Indicates the left data segment Index of sampling points Indicates the right-hand data segment Index of the sampling points; This indicates that the fitted line on the left is in the left data segment. The predicted flow velocity values ​​at each sampling point, where the subscript 1 indicates the fitted straight line on the left; This indicates that the fitted line on the right side is in the data segment on the right side. The predicted flow velocity values ​​at each sampling point, where the subscript 2 indicates the second fitted line on the right; Indicates the angle of collapse. Represents the arctangent function. This indicates taking the absolute value. It represents pi (π).

[0039] Step S3: Construct a dual-stream feature input to take into account both microscopic waveform morphology and macroscopic physiological benchmarks. The dual-stream feature input includes waveform stream input and statistical stream input, as detailed below: 1) Waveform stream input: directly using the multi-instance set constructed using S1 This input preserves the complete geometric shape of the flow-volume curve, which is used by the subsequent neural network to extract the microscopic pathological features of the curve trend.

[0040] 2) Statistical Stream Input: Constructing Statistical Feature Vectors The statistical feature vector is a one-dimensional concatenated vector, constructed by sequentially concatenating dynamic physiological statistical features and static demographic features. The dynamic physiological statistical features include the mean, standard deviation, minimum, and maximum values ​​of peak expiratory flow (Calc_PEF), small airway tangent angles (Angle_MEF50, Angle_MEF25), mid-segment mean flow angle (Angle_MMEF), and collapse angle (Calc_AC), totaling 20 dimensions. The static demographic features include age, gender, height, and weight, totaling 4 dimensions. The total dimensions are 24. The specific extraction method is as follows: a) Calc_PEF: The maximum value in the velocity-volume curve; b) Calc_AC: Calculated from S2; c) Angle_MEF50 and Angle_MEF25: Locate the 50% (corresponding to MEF50) and 75% (corresponding to MEF25) positions of the normalized volume axis, respectively, set a local small window, perform linear regression fitting on the velocity points within the window, and calculate the tilt angle of the fitted line. d) Angle_MMEF: The angle corresponding to the slope of the line connecting the 25% and 75% volume points.

[0041] based on For each exhalation, the mean, standard deviation, minimum, and maximum values ​​of the aforementioned dynamic physiological statistical features were calculated. The mean characterizes the overall physiological level, the standard deviation quantifies airway instability between exhalations, and the minimum and maximum values ​​characterize the range of feature fluctuations and extreme cases. The constructed statistical feature vectors are then input into a multilayer perceptron for feature mapping and fused with waveform flow features for classification and prediction.

[0042] Step S4: Based on the gating attention mechanism, feature extraction and aggregation are performed on the velocity-volume curve. First, the multi-example set... Each velocity-volume curve in Independent input to a one-dimensional convolutional neural network, outputting the corresponding high-dimensional instance feature vectors. Subsequently, a gated attention mechanism was introduced to calculate the contribution weight of each velocity-volume curve. This allows for the automatic differentiation between high-quality pathological waveforms and low-quality interference waveforms; then, based on the calculated contribution weights, a weighted patient-level waveform feature representation is obtained. Simultaneously, the statistical feature vector constructed from the statistical flow in S3 The input is a multilayer perceptron (MLP), and the macroscopic physiological feature representation is obtained through feature mapping. The calculation formulas for contribution weights and patient-level waveform feature representations are as follows:

[0043]

[0044] in, , This represents two distinct learnable projection matrices. Used to map high-dimensional instance feature vectors to a low-dimensional space. Used to map high-dimensional instance feature vectors to a gated space. Represents the weight vector. For element-wise multiplication, This represents the hyperbolic tangent activation function, which provides nonlinear transformation capabilities. This represents the activation function, which acts as a gating switch to regulate the information flow. This indicates transpose.

[0045] The one-dimensional convolutional neural network includes multiple convolutional modules. Preferably, the one-dimensional convolutional neural network in this invention consists of three one-dimensional convolutional layers with kernel sizes of 7, 5, and 3, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation layer. The ReLU activation layers after the first and second one-dimensional convolutional layers are followed by a max pooling layer, and the ReLU activation layer after the third one-dimensional convolutional layer is followed by an adaptive global average pooling layer, outputting the corresponding high-dimensional instance feature vector. .

[0046] Step S5, Feature Fusion and Classification Prediction. Represent the patient-level waveform features. Representation of macroscopic physiological characteristics The data are spliced ​​together to form a joint feature representation containing morphological, physiological, and variability information. This joint feature representation is then input into a classification network, which consists of three fully connected layers. The first fully connected layer is followed by a batch normalization layer, a ReLU activation layer, and a random inactivation layer. The second fully connected layer is followed by a ReLU activation layer, and the third fully connected layer is followed by an output layer. The output layer shows the abnormality confidence score of the subject's respiratory waveform. Set classification thresholds. ,when When the probability of an anomaly exceeds a threshold, a warning label is output; when... At that time, it was classified as normal.

[0047] The above classification results can effectively identify the risk of respiratory abnormalities in examinees and are suitable for early auxiliary screening of asthma.

[0048] Embodiments of the present invention also provide an early asthma screening system based on a dual-path airflow sensor, used to implement the above-described early asthma screening method based on a dual-path airflow sensor. The system structure is as follows: Figure 2 As shown, it includes: Data acquisition module: used to acquire respiratory airflow signals generated by the subject's multiple forced exhalations during a single testing task, and send the airflow signals to the processing unit, including an airflow sensor and a data transmission unit; Processor: Used to store and execute computer programs to implement the above-mentioned early asthma screening method based on dual-path airflow sensors; Interactive display module: Used to intuitively display the classification results. This module not only displays the final classification conclusion, but also draws the flow-volume curve of the subject on the screen, allowing for an intuitive view of the respiratory waveform morphology.

[0049] Experimental verification This experiment was conducted on a physical signal dataset containing 2953 samples. By comparing the impact of different feature combinations on classification performance, the effectiveness of the invention was verified. The experimental setup was as follows: Experimental Group 1 (Exp 1): Baseline scheme. A single-instance learning architecture was used, with only traditional flow velocity metrics as input, without incorporating demographic parameters; Experimental Group 2 (Exp 2): Morphological approach. Employing the multi-instance learning architecture of this invention, only the geometric morphological angles (Angle) and global collapse angles (AC) features extracted by this invention are input, without introducing demographic parameters; Experimental Group 3 (This Invention): Fully Fusion Scheme. Employing a multi-instance learning architecture, it fuses morphological features, flow velocity features, and demographic parameters.

[0050] The experimental results are shown in Table 1: Table 1 Performance comparison results of different schemes

[0051] like Figure 3The ROC curve comparison chart shows that the AUC (Area Under the Curve) value of the proposed model (solid line) is as high as 0.921, significantly better than the existing baseline model (dashed line, AUC=0.824) that relies solely on the PEF index. The rapid increase of the solid line in the upper left corner indicates that the proposed model significantly improves classification sensitivity while maintaining high specificity. The ACC (Accuracy) and Spec (Specificity) of the proposed model are also higher than those of experimental groups 1 and 2. Although the Sensitivity (Sensitivity) is slightly lower than that of experimental group 1 (80.70%), it is significantly higher than that of experimental group 2 (74.10%). Furthermore, the proposed model achieves the best results in both Spec and ACC, and its overall classification performance is the most balanced and stable.

[0052] like Figure 4 The attention heatmap shown further validates the interpretability of the model. The darker colored and thicker lines in the figure correspond to the waveform segments that the model assigns high weights to. As can be seen from the figure, the model can automatically focus on the subtle tremors or collapses that appear in the expiratory descending limb.

[0053] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for early asthma screening based on a dual-path airflow sensor, characterized in that, The steps are as follows: Step S1: Acquire airflow signals and construct a multi-sample time-series data package; Step S2: Calculate the adaptive collapse angle based on the global error; Step S3: Construct a dual-stream feature input, which includes a waveform stream input and a statistical stream input; Step S4: Based on the gating attention mechanism, feature extraction and aggregation are performed on the velocity-volume curve; Step S5: Feature fusion and classification prediction, outputting the abnormal confidence score of the subject's respiratory waveform. And set classification thresholds. ,when At that time, a warning label will be output.

2. The method for early asthma screening based on a dual-path airflow sensor according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Collect data continuously from the subject during a single testing task. The airflow signal generated by the forceful exhalation action is collected, and the raw airflow signal is preprocessed and spatially aligned. S12 transforms the airflow signal processed by S11 from the time domain to the velocity-volume domain. First, the start and end points of exhalation are located. Then, the velocity during the exhalation phase is integrated and accumulated to obtain the volume, thus obtaining the forced vital capacity (FVC) of a single exhalation. Then, the accumulated volume is divided by FVC to construct a normalized volume axis, obtaining a volume percentage sequence in the range [0,1]. A value is then set on this normalized volume axis. The flow velocity value at each of the three equally spaced sampling points is obtained by linear interpolation. Divide the flow rate value by the preset physiological limit flow rate benchmark value to obtain the flow rate-volume curve of a single exhalation action; S13, the same subject The velocity-volume curve is encapsulated as a multiple instance set, denoted as , ;in, It is the first A flow velocity-volume curve, Indicates the total number of sampling points. It represents the set of real numbers.

3. The method for early asthma screening based on a dual-path airflow sensor according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, define the search interval in the descending branch of the velocity-capacity curve. ; S22, traverse each sampling point within the search interval. This is considered as a potential morphological inflection point for each sampling point. Using this sampling point as the dividing point, the descent branch data is divided into the left data segment. and the right data segment Linear regression fitting was performed on each, and the total fitting error was calculated. Then select the option that minimizes the total fitting error. The point with the minimum value is taken as the global optimal shape inflection point, and the slope of the best fitted line on both sides of that point is obtained. , The collapse angle is calculated, and the formulas for calculating the total fitting error and the collapse angle are as follows: In the formula, Left data segment The Middle The actual flow velocity value at each sampling point The right-hand data segment The Middle The actual flow velocity value at each sampling point Indicates the left data segment Index of sampling points Indicates the right-hand data segment Index of the sampling points; This indicates that the fitted line on the left is in the left data segment. The predicted flow velocity values ​​at each sampling point, where the subscript 1 indicates the fitted straight line on the left; This indicates that the fitted line on the right side is in the data segment on the right side. The predicted flow velocity values ​​at each sampling point, where the subscript 2 indicates the second fitted line on the right; Indicates the angle of collapse. Represents the arctangent function. This indicates taking the absolute value. It represents pi (π).

4. The method for early asthma screening based on a dual-path airflow sensor according to claim 1, characterized in that, The method for determining the dual-stream feature input described in S3 is as follows: 1) Waveform stream input: A multi-instance set constructed by S1 ; 2) Statistical Stream Input: Constructing Statistical Feature Vectors It is composed of dynamic physiological statistical features and static demographic features in sequence; the dynamic physiological statistical features include the mean, standard deviation, minimum and maximum values ​​of peak expiratory flow rate, small airway tangent angle, mid-segment average flow rate angle and collapse angle, totaling 20 dimensions; the static demographic features include age, sex, height and weight, totaling 4 dimensions.

5. The method for early asthma screening based on a dual-path airflow sensor according to claim 4, characterized in that, The specific method for extracting the dynamic physiological statistical features is as follows: a) Peak expiratory flow rate: the maximum value in the flow-volume curve; b) Collapse angle: calculated from S2; c) Small airway tangent angle: Locate the 50% and 75% positions of the normalized volume axis respectively, set a local micro window, perform linear regression fitting on the flow velocity points within the window, and calculate the inclination angle of the fitted line. d) Mid-section average velocity angle: the angle corresponding to the slope of the line connecting the 25% and 75% points of the volume.

6. The method for early asthma screening based on a dual-path airflow sensor according to claim 1, characterized in that, In S4, the multi-instance set is first... Each velocity-volume curve in Independent input to a one-dimensional convolutional neural network, outputting the corresponding high-dimensional instance feature vectors. ; Subsequently, a gated attention mechanism is introduced to calculate the contribution weight of each velocity-volume curve. ; Then, based on the calculated contribution weights, the weighted patient-level waveform feature representation is obtained. Simultaneously, the statistical feature vector constructed from the statistical flow in S3 The input is a multilayer perceptron (MLP), and the macroscopic physiological feature representation is obtained through feature mapping. The calculation formulas for contribution weight and patient-level waveform feature representation are as follows: in, , This represents two distinct learnable projection matrices. Used to map high-dimensional instance feature vectors to a low-dimensional space. Used to map high-dimensional instance feature vectors to a gated space. Represents the weight vector. For element-wise multiplication, This represents the hyperbolic tangent activation function. This represents the activation function. This indicates transpose.

7. The method for early asthma screening based on a dual-path airflow sensor according to claim 6, characterized in that, The one-dimensional convolutional neural network comprises three one-dimensional convolutional layers with kernel sizes of 7, 5, and 3, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation layer. The first and second one-dimensional convolutional layers are followed by a max pooling layer after their ReLU activation layers, and the third one-dimensional convolutional layer is followed by an adaptive global average pooling layer after its ReLU activation layer. The output is the same as the first... Flow velocity-volume curve Corresponding high-dimensional instance feature vector .

8. The method for early asthma screening based on a dual-path airflow sensor according to claim 1, characterized in that, In S5, the patient-level waveform features are represented. Representation of macroscopic physiological characteristics The data are spliced ​​together to form a joint feature representation that includes morphological, physiological, and variability information. This joint feature representation is then input into the classification network.

9. A method for early asthma screening based on a dual-path airflow sensor according to claim 8, characterized in that, The classification network consists of a three-layer fully connected structure, where the first fully connected layer is followed by a batch normalization layer, a ReLU activation layer, and a random deactivation layer, the second fully connected layer is followed by a ReLU activation layer, and the third fully connected layer is followed by an output layer.

10. An early asthma screening system based on a dual-path airflow sensor, used to implement the early asthma screening method based on a dual-path airflow sensor as described in any one of claims 1 to 9, characterized in that, Includes the following modules: Data acquisition module: used to acquire respiratory airflow signals generated by the subject's multiple forced exhalations during a single testing task, and send the airflow signals to the processing unit; Processor: Used to store and execute computer programs to implement the above-mentioned early asthma screening method based on dual-path airflow sensors; Interactive display module: Used to intuitively display classification results.

Images