Cough sound-based screening auxiliary system for chronic obstructive pulmonary disease and electronic equipment

Abstract

The invention relates to the technical field of medical data assistance, in particular to a screening assistance system and electronic equipment for chronic obstructive pulmonary disease based on cough sound, and the screening assistance system comprises a data acquisition module which is used for acquiring to-be-processed audio data; the cough sound extraction module is used for acquiring a Mel-frequency spectrogram of the audio data and extracting cough sound fragments from the Mel-frequency spectrogram by using a pre-trained Conformer-BiLSTM network; the chronic obstructive pulmonary disease screening module is used for preprocessing the cough sound segment and converting the cough sound segment into a frequency domain signal, and acquiring an FBank characteristic parameter corresponding to the frequency domain signal through a Mel filter; and inputting the FBank characteristic parameters into a pre-trained classification model with dynamic convolution and a multi-scale attention mechanism, and outputting classification results of the chronic obstructive pulmonary cough sound and the non-chronic obstructive pulmonary cough sound. According to the invention, the possibility of classification errors is reduced, and the classification efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical data auxiliary technology, and in particular to a cough sound-based screening auxiliary system and electronic equipment for chronic obstructive pulmonary disease. Background Art

[0002] Chronic obstructive pulmonary disease (COPD) is a respiratory disease that affects hundreds of millions of people worldwide. The current standard diagnostic method for COPD is pulmonary function testing, but this method is costly and time-consuming, making it difficult to meet the needs of large-scale population screening. In contrast, cough sounds, as a biosignal that can reflect abnormal lung function, have the advantages of being non-invasive and easy to collect.

[0003] However, current COPD screening using cough sounds still relies on manual labeling and simple models, which are easily affected by environmental noise and have high missed detection and false detection rates. As a result, the extracted cough sounds cannot be accurately classified or are incorrectly classified during subsequent processing, affecting the system's classification accuracy. Summary of the Invention

[0004] (1) Technical issues to be resolved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a COPD screening auxiliary system and electronic device based on cough sounds, which solves the technical problem in the prior art that cough sounds cannot be accurately extracted and graded.

[0005] (2) Technical solution In order to achieve the above objectives, the main technical solutions adopted by the present invention include: In a first aspect, a cough sound-based COPD screening auxiliary system proposed in an embodiment of the present invention includes a data acquisition module for acquiring audio data to be processed; a cough sound extraction module, configured to obtain a mel-spectrogram of the audio data and extract cough sound segments from the mel-spectrogram using a pre-trained Conformer-BiLSTM network; The COPD screening module is used to preprocess the cough sound segment and convert it into a frequency domain signal, obtain the FBank feature parameters corresponding to the frequency domain signal through a Mel filter; input the FBank feature parameters into a pre-trained classification model with dynamic convolution and multi-scale attention mechanism, and output the classification results of COPD cough sound and non-COPD cough sound.

[0006] Optionally, the system also includes a COPD grading module, which is used to input the FBank feature parameters into a pre-trained screening model to obtain the level of the COPD cough sound when the classification result is the result of COPD cough sound; the level output by the pre-trained screening model includes at least five levels.

[0007] Optionally, the screening model uses a softmax function to obtain the level of COPD cough sounds; wherein the softmax function is an activation function for five-level classification, responsible for outputting probability; The structure of the screening model except the cross entropy loss function is the same as the structure of the non-cross entropy loss function in the classification model.

[0008] Optionally, the cough sound extraction module includes: a conversion unit for performing frequency domain conversion on audio data with a sampling rate of 16000 Hz and a frame length greater than 32 ms, and obtaining a mel-spectrogram with time-frequency features; a detection unit for extracting local time-frequency features with spatial dimension compression performance from the mel-spectrogram through the CNN layer of the Conformer-BiLSTM network; and obtaining the temporal dependency of the mel-spectrogram through the BiLSTM layer of the Conformer-BiLSTM network; the Conformer layer of the Conformer-BiLSTM network performs a residual connection based on the local time-frequency features and the temporal dependency to obtain the position of the cough frame corresponding to the audio data; and an extraction unit for extracting cough sound segments from the audio data based on the position of the cough frame.

[0009] Optionally, the Conformer-BiLSTM network includes: a three-layer CNN layer, a two-layer BiLSTM, and three stacked Conformer structures; each Conformer structure includes: a multi-head attention branch for capturing global contextual relationships, and a convolution branch for local feature enhancement; the Conformer structure is also used to fuse the residual connections output by the multi-head attention branch and the convolution branch respectively; in the training stage, the Adam optimizer is used, the initial learning rate is 0.001, the batch size is 16, the epoch is set to 100, and the training stops after reaching the maximum epoch number. The loss function is Focal Loss.

[0010] Optionally, the Conformer-BiLSTM network further includes a BatchNorm layer, which is used to normalize the output of the Conformer structure to obtain and output a frame-level cough probability sequence.

[0011] Optionally, the COPD screening module includes: a preprocessing unit for pre-emphasis on cough sound segments, and dividing the pre-emphasized signal into short time frames of specified duration, and performing windowing processing on each frame, and converting the windowed frames into frequency domain signals through fast Fourier transform; an FBank feature parameter acquisition unit for inputting the frequency domain signal into a Mel filter to obtain FBank feature parameters; a classification unit for inputting the FBank feature parameters into a pre-trained classification model with dynamic convolution and multi-scale attention mechanism, and outputting the classification results of COPD cough sounds and non-COPD cough sounds.

[0012] Optionally, the classification model includes: a full-dimensional dynamic convolution module with a four-branch parallel architecture, a residual network and a statistical pooling layer; The full-dimensional dynamic convolution module is used to perform cross-dimensional dynamic weighting processing on the number of convolution kernels, input channels, feature information of output channels, and spatial dimensions of the FBank feature parameters to generate an adaptive weight matrix to generate preliminary features; The residual network includes multiple Res2Net residual modules, each of which realizes cross-group feature interaction through group convolution and channel reorganization technology, and uses a multi-scale attention mechanism in the residual path to achieve weighted fusion of multi-scale features in multiple residual modules, strengthen pathology-related features, and output the first feature; The statistical pooling layer is used to normalize the first feature and output the classification results of COPD cough sounds and non-COPD cough sounds through a sigmoid function; wherein the sigmoid function is a binary classification activation function.

[0013] Optionally, the statistical pooling layer is a one-dimensional pooling layer.

[0014] In a second aspect, an electronic device proposed in an embodiment of the present invention includes any of the above-mentioned cough sound-based COPD screening auxiliary systems.

[0015] (3) Beneficial effects The beneficial effects of the present invention are as follows: the present invention provides a COPD screening auxiliary system and electronic device based on cough sounds. The present invention uses a data acquisition module to obtain audio data to be processed, and uses a cough sound extraction module to obtain a mel-spectrogram of the audio data, and uses a pre-trained Conformer-BiLSTM network to extract cough sound segments from the mel-spectrogram. The COPD screening module then pre-processes and converts the cough sound segments into a frequency domain to obtain feature parameters, and the FBank feature parameters are input into a pre-trained classification model with dynamic convolution and multi-scale attention mechanism to output the classification results of COPD cough sounds and non-COPD cough sounds. In this way, the classification of COPD can be achieved.

[0016] Furthermore, the present invention utilizes mel-spectrograms to locate cough sounds in both the time and frequency domains, improving cough sound localization accuracy. Furthermore, the present invention utilizes modules for both classification and cough sound extraction, rather than relying on human experience. This reduces the likelihood of classification errors and improves classification efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a COPD screening assistance system based on cough sounds provided by an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a cough sound extraction module provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of another cough sound extraction module provided by an embodiment of the present invention; Figure 4 A schematic structural diagram of a COPD grading module provided by an embodiment of the present invention; Figure 5 A schematic flow chart of another cough sound-based COPD screening assistance system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] Chronic obstructive pulmonary disease (COPD) is a common chronic disease characterized by chronic bronchitis and / or emphysema with airflow obstruction, which can further develop into cor pulmonale and respiratory failure. Currently, COPD has become a respiratory disease that affects the health of hundreds of millions of people worldwide. The current standard diagnostic method for COPD is pulmonary function testing, but this method is costly and time-consuming, making it difficult to meet the needs of large-scale population screening. In contrast, cough sounds, as a biological signal that can reflect abnormal lung function, have the advantages of being non-invasive and easy to collect.

[0019] In recent years, research on cough sound-assisted diagnosis of COPD has gradually attracted attention. However, traditional methods rely on manual labeling or simple models, are easily affected by environmental noise, and have high missed detection and false detection rates. Existing technologies are mostly limited to COPD screening, and are unable to capture pathological features such as abnormal low-frequency energy distribution and spectrum discretization in cough sounds, and cannot meet the clinical needs for accurate grading of the severity of COPD.

[0020] In order to solve the above problems, the present invention provides a COPD screening auxiliary system based on cough sounds. The present invention uses a data acquisition module to obtain audio data to be processed, and uses a cough sound extraction module to obtain a mel-spectrogram of the audio data, and uses a pre-trained Conformer-BiLSTM network to extract cough sound segments from the mel-spectrogram. The COPD screening module then pre-processes and converts the cough sound segments into a frequency domain to obtain feature parameters, and the FBank feature parameters are input into a pre-trained classification model with a dynamic convolution and multi-scale attention mechanism to output the classification results of COPD cough sounds and non-COPD cough sounds. In this way, the classification of COPD can be achieved, and the present invention uses a mel-spectrogram to locate cough sounds from both the time domain and the frequency domain, thereby improving the accuracy of cough sound positioning. In addition, the classification and cough sound extraction of the present invention are both completed using modules, rather than relying on the work experience of the staff. In this way, the possibility of classification errors is reduced to a certain extent, and the classification efficiency is improved.

[0021] For a better explanation of the present invention, please refer to Figure 1 ,like Figure 1 As shown, the system can be applied to any electronic device in medical field or medical computer system, including: The data acquisition module is used to acquire the audio data to be processed. In this embodiment, the audio data of the patient to be tested can be collected and stored in advance by a mobile phone or other audio acquisition means, and then the data acquisition module can acquire the audio data for processing.

[0022] The cough sound extraction module is used to obtain a mel-spectrogram of the audio data and extract cough sound segments from the mel-spectrogram using a pre-trained Conformer-BiLSTM network.

[0023] The COPD screening module is used to preprocess the cough sound segment and convert it into a frequency domain signal, obtain the FBank feature parameters corresponding to the frequency domain signal through a Mel filter; input the FBank feature parameters into a pre-trained classification model with dynamic convolution and multi-scale attention mechanism, and output the classification results of COPD cough sound and non-COPD cough sound.

[0024] Specifically, the system includes at least one of the following: a data acquisition module, a cough sound extraction module, and a COPD screening module. The data acquisition module is primarily used to acquire audio data collected by a collection device. The collection device includes, but is not limited to, at least one of the following: a smartphone, a microphone, a voice recorder, and a lung acoustics detector. The audio data collected by the data acquisition module has time-domain characteristics.

[0025] During the specific collection process, to maintain the purity of the sound signal and the clarity of the speech, participants were asked to remove their masks during the recording process. The microphone of the collection equipment was carefully disinfected after each use to ensure hygiene and safety. During recording, the microphone was placed 30-40 cm from the participant's mouth, at a 45-degree angle upward from the horizontal. This distance and angle were chosen to capture the details of the cough sound to the greatest extent possible. Participants coughed according to a standardized method, namely taking a deep breath first, then coughing out forcefully.

[0026] Due to varying environments during the collection process, the audio data captured by the collection device may include not only the participant's cough sounds, but also other sounds such as background noise and equipment noise. Therefore, a cough sound extraction module can be used to extract the participant's cough sounds from the audio data, thereby preventing other noise from affecting subsequent classification results. This embodiment does not specifically limit the method by which the cough sound extraction module extracts cough sounds.

[0027] In an illustrative embodiment, the cough sound extraction module specifically obtains the mel-spectrogram of the audio data, and uses a pre-trained Conformer-BiLSTM network to extract cough sound segments from the mel-spectrogram. Specifically, since the audio data itself has time domain characteristics, it only needs to be converted into the frequency domain to obtain the frequency domain characteristics of the audio data, and then the mel-spectrogram can be obtained by combining the time domain characteristics of the audio data itself. Since the recognition of cough sounds depends not only on the frequency distribution, but also on the dynamic changes of the time dimension, the mel-spectrogram can display the distribution pattern of the sound in the audio data from both the time and frequency aspects, thereby quickly and accurately locating the cough sound.

[0028] The cough sound extraction module in this embodiment may specifically include: a conversion unit, a detection unit, and an extraction unit. The conversion unit is used to perform frequency domain conversion on the audio data. This embodiment does not specifically limit the format of the audio data; for example, it may be audio data with a sampling rate of 16000 Hz and a frame length greater than 32 ms. A mel-spectrogram with time-frequency characteristics is obtained based on the frequency and time domain features of the audio data. The detection unit may use the Conformer-BiLSTM network's CNN layer to extract local time-frequency features from the mel-spectrogram that exhibit spatial dimension compression. Specifically, because the mel-spectrogram has both time (i.e., time domain) and frequency (i.e., frequency domain) dimensions, these two dimensions together construct the spatial structure of the mel-spectrogram. The CNN layer can retain the most critical features of the mel-spectrogram in both dimensions through convolutional layer stride and pooling operations. For example, cough sounds often have sudden and short pulses, which are characteristic of the time domain. In the frequency domain, cough sounds are often more concentrated in frequency. Combining these features for analysis, which is equivalent to comprehensively considering both spatial and temporal dimensions, can improve efficiency in the subsequent cough sound identification stage. For example, the low-frequency energy distribution in some cough sounds is abnormal and the spectrum is discrete. By jointly analyzing the frequency energy distribution and time of the cough sound, we can determine whether its capacity is abnormal and whether its spectrum is discrete.

[0029] The detection unit uses the BiLSTM layer of the Conformer-BiLSTM network to obtain the temporal dependencies of the mel-spectrogram. Specifically, the BiLSTM layer captures the contextual information of the mel-spectrogram in chronological order. This contextual information refers to analyzing the relationships between previous and subsequent frames, using frames as the unit for ease of understanding. It should be noted that the use of frames as the unit is merely an example for ease of understanding and can be adjusted based on actual circumstances. The relationship between the preceding and following seconds of each second can also be analyzed. Based on the temporal relationships that may exist between multiple frames in the audio data (for example, a cough has continuous pulses that occur within a certain timeframe, meaning that frames with pulses close in time can be considered part of a single cough), the Conformer layer performs a residual connection based on local time-frequency features and temporal dependencies to determine the location of the cough frame in the audio data. The Conformer layer primarily performs feature fusion, capturing the relationships between frames while preserving the time-frequency features within the frame, thereby improving cough localization accuracy.

[0030] For easier understanding, please refer to Figure 2 ,like Figure 2As shown in the figure, the mel-spectrogram can be sequentially passed through the CNN layer in the Conformer-BiLSTM network to obtain local time-frequency features. The mel-spectrogram is then sequentially passed through two layers of BiLSTM to obtain the temporal dependency of the mel-spectrogram. Finally, the temporal dependency and local time-frequency features are sequentially passed through three stacked Conformers to output the location of the cough frame.

[0031] The extraction unit can extract the audio in the audio data according to the position of the cough frame extracted by the detection unit to obtain the cough sound segment.

[0032] It should be noted that this embodiment does not specifically limit the specific structure of the Conformer-BiLSTM network. For example, the network structure may include: a three-layer CNN layer, a two-layer BiLSTM, and three stacked Conformer structures. This arrangement allows for full extraction of features from the mel-spectrogram. At the same time, due to the limited spatial complexity of the mel-spectrogram, the three-layer CNN layer can also avoid overfitting during feature extraction. The two-layer BiLSTM can not only learn the temporal relationship between frames, but also capture the temporal relationship between frames in the entire audio data. For example, the first-layer BiLSTM can be used to capture the temporal dependency between each frame and adjacent frames in the mel-spectrogram, while the second-layer BiLSTM can capture the temporal dependency between non-adjacent frames based on the output of the first-layer BiLSTM. This allows for more accurate location of cough frames, improving the accuracy of cough frame determination.

[0033] Optionally, any Conformer structure can include: a multi-head attention branch for capturing global contextual relationships and a convolution branch for local feature enhancement. Figure 3 ,like Figure 3As shown in the figure, Attention Module is the multi-head attention branch, Conv Module is the convolution branch, Feed forward Module is the feedforward module, which is used to perform nonlinear transformation on temporal dependencies and local time-frequency features to enhance feature expression capabilities, LayerNorm is layer normalization, which is used for normalization processing, and Attention Score is the attention score, which is used to learn the association between frames. The module input is the temporal features of the temporal dependency of the Mel-spectrogram extracted by the BiLSTM layer. It first enters the first residual structure, and the features preliminarily processed by the feedforward module are added to the original input. Its output is sent to two branches in parallel. Each branch includes a residual structure. The output of the first branch on the left is the input feature that first passes through the convolution branch and then through the multi-head attention branch to add to the original feature. The input feature of the second branch on the right first passes through the multi-head attention branch and then through the convolution branch to add to the original feature. At the same time, the two branches fuse the key attention scores of each branch to achieve information fusion. After processing by these two branches, the outputs of the left and right branches finally use the adder to complete the third residual connection. The formed comprehensive features are sent to the encoder's feedforward module for further conversion and integration; the output of this module is finally normalized by the layer to obtain the final output of the entire encoder module. This output effectively combines the advantages of convolution and attention. At the same time, the multi-level residual connection retains the original information and promotes feature reuse, enhancing the model's multi-scale feature extraction capability and long-range dependency modeling capability.

[0034] Optionally, the Conformer-BiLSTM network may further include a BatchNorm layer, which is used to normalize the output of the Conformer structure to obtain a frame-level cough probability sequence. Simultaneously, a dynamic threshold detection algorithm is combined to analyze the temporal distribution characteristics to accurately locate the start and end points of the cough sound. For example, after the BatchNorm layer normalizes the output of the Conformer structure, the probability of each frame being a cough sound can be obtained. When this probability is greater than a certain threshold, the frame can be considered a cough frame. This embodiment does not specifically limit the threshold; for example, the threshold is 80%.

[0035] During the training of the Conformer-BiLSTM network, the Adam optimizer can be used to optimize the network parameters. The initial learning rate of the optimizer is 0.001, the batch size is 16, the epoch is set to 100, and the training stops after the maximum number of epochs is reached. The loss function is Focal Loss. In this way, in the training samples used in the training process, the number of cough frames and non-cough frames is often unbalanced, and the number of non-cough frames may be far greater than that of cough frames. Therefore, this embodiment uses Focal Loss to introduce a modulation factor to make the network pay more attention to cough frame samples that are difficult to classify, thereby effectively alleviating the problems caused by category imbalance and improving the network's ability to detect cough frames.

[0036] The specific relationship can be satisfied as follows: ; Among them, FL is the frame-level cough probability sequence; is the true label, is the model prediction probability, is a parameter used to balance the weights of positive and negative samples, It is a focusing parameter for adjusting the weight of easy-to-classify samples. In this embodiment, it is set to 2. The base of the log function is e, that is, the natural logarithm.

[0037] After obtaining the cough sound segment output by the cough sound extraction module, the COPD screening module can use the preprocessing unit, the FBank feature parameter acquisition unit and the classification unit to obtain the classification result.

[0038] In one exemplary embodiment, the preprocessing unit is configured to pre-emphasize cough segments, enhancing the high-frequency components within the segments to better capture audio details. The pre-emphasized signal is then divided into short time frames of a specified duration, and each frame is windowed using a window function to reduce signal discontinuities between frames. Finally, the windowed frames are converted into frequency domain signals using a fast Fourier transform. This results in a clearer, more continuous frequency domain signal for the cough segments.

[0039] Next, the frequency domain signal can be input into the Mel filter, and the frequency domain signal of each frame can be filtered to obtain the FBank feature parameters of the cough sound segment. Among them, the Mel filter bank is a set of triangular filters built based on the Mel scale. When the frequency domain signal passes through the Mel filter, the frequency components can be integrated according to the auditory perception of the human ear, making it easier for subsequent units to obtain the essential characteristics of the cough sound. In addition, the FBank feature parameters output by the Mel filter are obtained by filtering the spectrum and then combining it, that is, the FBank feature parameters have dual characteristics of the time domain and frequency domain.

[0040] Since the FBank feature parameters retain both frequency domain features and time domain features, the FBank feature parameters can be further input into a pre-trained classification model with dynamic convolution and multi-scale attention mechanism to fully explore the association between the two dimensions (such as the pulse change of a certain frequency at a specific time point), thereby outputting the classification results of COPD cough sounds and non-COPD cough sounds.

[0041] It should be noted that this embodiment does not specifically limit the structure of the classification model. Exemplarily, the classification model may include: a full-dimensional dynamic convolution module with a four-branch parallel architecture, a residual network, and a statistical pooling layer.

[0042] In other words, the full-dimensional dynamic convolution module is used to perform cross-dimensional dynamic weighting processing on the number of convolution kernels, input channels, feature information of output channels, and spatial dimensions of the FBank feature parameters, thereby breaking the isolation between dimensions and capturing the time-frequency crossover features in the FBank feature parameters, generating an adaptive weight matrix, and generating preliminary features. In other words, the full-dimensional dynamic convolution module can determine which dimension is more capable of discriminating COPD and non-COPD cough sounds. The input channel and output channel can be regarded as frequency domain dimensions, spatial dimensions, and time domain dimensions. The number of convolution kernels can be understood as the importance of each dimension. For example, COPD cough is often accompanied by low-frequency resonance caused by airway stenosis, which has strong discriminability. Therefore, in the frequency domain dimension, the number of convolution kernels will be given a higher weight.

[0043] The specific relationship can be satisfied as follows: ; in, Represents the convolution kernel The weight of Represents the weight on the convolution kernel space; represents the weight of the input channel, Represents the weight of the output channel, i ranges from 1 to n, x is the initial convolution, y is the final convolution, is the tensor product.

[0044] The residual network can learn more complex pathological features layer by layer from the initial features, providing a basis for subsequent judgment. Among them, this embodiment does not specifically limit the structure of the residual network. Exemplarily, the residual network can include multiple Res2Net residual modules, each of which realizes cross-group feature interaction through group convolution and channel reorganization technology, that is, each group uses an independent convolution kernel to extract different types of features. This embodiment does not specifically limit the number of residual modules. Exemplarily, it can be four. For example, the first group focuses on low-frequency features, the second group focuses on high-frequency features, and so on. And use a multi-scale attention mechanism in the residual path to realize the weighted fusion of multi-scale features in multiple residual modules, strengthen pathology-related features, and output the first feature.

[0045] The statistical pooling layer is used to normalize the first feature, and output the classification results of COPD cough sound and non-COPD cough sound through the sigmoid function. That is to say, the sigmoid function is used to output the probability that the cough sound segment is a COPD cough sound or a non-COPD cough sound. According to the probability, it is possible to determine whether the cough sound segment is a COPD cough sound. Among them, the sigmoid function is a binary classification activation function. The statistical pooling layer in this embodiment is a one-dimensional pooling layer, which can retain the key frequency band timing information while compressing the features along the time axis, and strengthen the channel response of pathology-related features through the frequency gating mechanism, which is more suitable for the characteristics of time series data such as audio. In the analysis of COPD cough sound, the changing trend of the frequency channel over time can be accurately captured, so that the channel attention can be more focused on features such as duration and time-frequency linkage related to the pathology, and ultimately improve the model's accuracy in distinguishing COPD cough.

[0046] Optionally, the system may also include a COPD grading module, which is used to input the FBank feature parameters into a pre-trained screening model to obtain the level of the COPD cough sound when the classification result is the result of COPD cough sound; the level output by the pre-trained screening model includes at least five levels. This level screening model uses a softmax function to obtain the level of the COPD cough sound; wherein, the softmax function is an activation function for five-level classification, and the softmax function can be used to output the probability that the cough sound segment is a certain level. The structure of the screening model except the five-level classification activation function is the same as the structure in the classification model. wherein, the levels include but are not limited to at least one of the following: normal, mild, moderate, severe and extremely severe.

[0047] It should be noted that the COPD grading module may also include other processing structures, which are not specifically limited in this embodiment.

[0048] In an illustrative example, please refer to Figure 4 ,like Figure 4As shown, the COPD classification module takes FBank features of COPD cough sounds as input. It first undergoes a full-dimensional dynamic convolution module for dynamic multi-dimensional feature extraction, followed by a batch normalization layer to normalize the feature distribution. The feature dimensions are then compressed by a max pooling layer before entering a sequence of residual modules in the backbone network: the first two-layer residual module processes basic features, the second three-layer residual module deepens feature abstraction, the third five-layer residual module strengthens feature representation, and finally, a three-layer residual module completes high-level feature compression. The model then selects a pooling type based on requirements: Adaptive Spatial Pooling (ASP), Spatial Adaptive Pooling (SAP), Temporal Adaptive Pooling (TAP), or Spatiotemporal Pooling (TSP), which uniformly scales the input to eliminate input scale differences. The pooled output is fed into a fully connected layer for class space mapping. Finally, a softmax activation function is used to generate a probability distribution for COPD diagnosis, completing the classification result for intelligent decision-making.

[0049] The above is a COPD screening auxiliary system based on cough sounds provided by the present invention. The above embodiments can be combined according to actual conditions, and the present invention does not make specific limitations here.

[0050] To facilitate understanding of the solution provided by the present invention, please refer to Figure 5 ,like Figure 5 As shown in the figure, after the participants complete the data collection, the audio data will be uploaded to the audio database. Next, the cough sound extraction module can be used to detect and extract the cough sound. Furthermore, the cough sound is classified according to the extracted cough sound. If it is classified as COPD, it can be graded.

[0051] In addition, an embodiment of the present invention also provides an electronic device, which may include a memory and a processor, wherein the memory stores a computer program, and the processor executes the program in the memory, such as executing the cough sound-based COPD screening assistance system of any of the above embodiments.

[0052] The advantages of this electronic device correspond to the advantages of any of the above-mentioned screening assistance systems. It can quickly and accurately distinguish whether the audio belongs to COPD. If it belongs to COPD, it can further subdivide its level to ensure the accuracy of the judgment.

[0053] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0054] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; and internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0055] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0056] In the description of this specification, the terms "one embodiment", "some embodiments", "embodiments", "examples", "specific examples" or "some examples" refer to the specific features, structures, materials or characteristics described in conjunction with the embodiment or example and included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and features of different embodiments or examples, unless they are mutually inconsistent.

[0057] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A COPD screening auxiliary system based on cough sound, characterized by: include: A data acquisition module, used to acquire audio data to be processed; a cough sound extraction module, configured to obtain a mel-spectrogram of the audio data and extract cough sound segments from the mel-spectrogram using a pre-trained Conformer-BiLSTM network; The COPD screening module is used to preprocess the cough sound segment and convert it into a frequency domain signal, obtain the FBank feature parameters corresponding to the frequency domain signal through a Mel filter; input the FBank feature parameters into a pre-trained classification model with dynamic convolution and multi-scale attention mechanism, and output the classification results of COPD cough sound and non-COPD cough sound.

2. The system according to claim 1, wherein: The system also includes a COPD grading module, which is used to input the FBank feature parameters into a pre-trained screening model to obtain the level of COPD cough sound when the classification result is COPD cough sound; the levels output by the pre-trained screening model include at least five levels.

3. The system according to claim 2, characterized in that The screening model uses a softmax function to obtain the level of COPD cough sound; wherein the softmax function is a five-category activation function responsible for outputting probability; The structure of the screening model except the cross entropy loss function is the same as the structure of the non-cross entropy loss function in the classification model.

4. The system according to claim 1, wherein: Cough sound extraction module, including: A conversion unit is used to perform frequency domain conversion on audio data with a sampling rate of 16000 Hz and a frame length greater than 32 ms, and obtain a Mel-spectrogram with time-frequency characteristics; The detection unit extracts local time-frequency features with spatial dimension compression performance from the Mel-spectrogram through the CNN layer of the Conformer-BiLSTM network; And, the temporal dependency of the Mel-spectrogram is obtained through the BiLSTM layer of the Conformer-BiLSTM network; The Conformer layer of the Conformer-BiLSTM network performs residual connections based on local time-frequency features and temporal dependencies to obtain the position of the cough frame corresponding to the audio data; The extraction unit is configured to extract cough sound segments from the audio data based on positions of cough frames.

5. The system according to claim 4, characterized in that The Conformer-BiLSTM network includes: three layers of CNN layers, two layers of BiLSTM, and three stacked Conformer structures; Each Conformer structure includes: a multi-head attention branch for capturing global contextual relationships and a convolution branch for local feature enhancement; the Conformer structure is also used to fuse the residual connections output by the multi-head attention branch and the convolution branch respectively; During the training phase, the Adam optimizer was used, with an initial learning rate of 0.001, a batch size of 16, and an epoch of 100. Training stopped after the maximum number of epochs was reached, and the loss function was Focal Loss.

6. The system according to claim 5, characterized in that The Conformer-BiLSTM network further includes a BatchNorm layer, which is used to normalize the output of the Conformer structure to obtain and output a frame-level cough probability sequence.

7. The system according to claim 1, wherein: COPD screening module, including: a pre-processing unit, configured to pre-emphasize the cough sound segment, divide the pre-emphasized signal into short time frames of a specified duration, perform windowing on each frame, and convert the windowed frames into frequency domain signals through a fast Fourier transform; An FBank feature parameter acquisition unit is used to input the frequency domain signal into the Mel filter to obtain the FBank feature parameters; A classification unit is used to input the FBank feature parameters into a pre-trained classification model with dynamic convolution and multi-scale attention mechanism, and output the classification results of COPD cough sounds and non-COPD cough sounds.

8. The system according to claim 7, characterized in that The classification model includes: a full-dimensional dynamic convolution module with a four-branch parallel architecture, a residual network and a statistical pooling layer; The full-dimensional dynamic convolution module is used to perform cross-dimensional dynamic weighting processing on the number of convolution kernels, input channels, feature information of output channels, and spatial dimensions of the FBank feature parameters to generate an adaptive weight matrix to generate preliminary features; The residual network includes multiple Res2Net residual modules, each of which realizes cross-group feature interaction through group convolution and channel reorganization technology, and uses a multi-scale attention mechanism in the residual path to achieve weighted fusion of multi-scale features in multiple residual modules, strengthen pathology-related features, and output the first feature; The statistical pooling layer is used to normalize the first feature and output the classification results of COPD cough sounds and non-COPD cough sounds through a sigmoid function; wherein the sigmoid function is a binary classification activation function.

9. The system according to claim 8, characterized in that The statistical pooling layer is a one-dimensional pooling layer.

10. An electronic device, characterized in that: A cough sound-based COPD screening auxiliary system comprising any one of claims 1 to 9.

Images