Breathing sound data classification system based on multi-view ensemble learning

Abstract

The invention provides a breath sound data classification system based on multi-view ensemble learning, and belongs to the field of audio data analysis in artificial intelligence. The system comprises a data preprocessing module, a multi-view feature extraction module, a multi-view integrated learning module based on meta learning and a parameter optimization module based on a firefly algorithm. The data processing module is used for carrying out preprocessing work such as noise reduction, segmentation and standardization processing on the collected original breath sound data; the multi-view feature extraction module is used for extracting acoustic features, spectrum features and time sequence features of the breath sound data; the multi-view integrated learning module based on meta learning is used for establishing three heterogeneous base classifiers and fusing results output by the three base classifiers by using a stacking technology; according to the method, a parameter optimization module based on a firefly algorithm is established for a plurality of sensitive parameters involved in a multi-view integrated learning module based on meta-learning to carry out multi-parameter optimization.

Description

Technical Field

[0001] The present invention provides a respiratory sound data classification system based on multi-view ensemble learning, which can be used for early detection and large-scale screening of respiratory diseases such as chronic obstructive pulmonary disease, and belongs to the field of audio data analysis in computer science and artificial intelligence. Background Art

[0002] According to the World Health Organization, chronic obstructive pulmonary disease (COPD) is the third leading cause of death worldwide (World Health Organization, 2023). Due to the lack of knowledge or diagnostic experience of COPD by patients and junior clinical workers, they are unable to effectively identify the early symptoms of COPD, resulting in a low diagnosis rate for COPD. Many patients miss the best time for interventional treatment due to misdiagnosis and missed diagnosis (Peng Xianru et al., 2018). Currently, the main diagnostic methods related to COPD include: (1) Clinical symptom assessment: the presence of specific symptoms, such as shortness of breath, persistent cough, sputum or asthma, and a large amount of exposure to hazardous factors, including cigarette smoke or other environmental pollutants; (2) Pulmonary function test: the FEV1 / FVC ratio is less than 0.70 after the use of bronchodilators, which is also the gold standard for the diagnosis of COPD; (3) Imaging examination: X-ray, CT, etc. are used to observe abnormal changes in lung structure.

[0003] At present, researchers have been trying to integrate artificial intelligence technology with COPD risk assessment in order to achieve early detection of COPD. From a data perspective, existing methods mainly use questionnaires (Wang et al. 2023), electronic medical records (Cheng et al. 2017) and medical images (Tang et al. 2020) to establish early screening models for COPD, while there are fewer technical research and developments using audio data to build early screening models for COPD. With the advancement of data collection and storage technology, more and more medical devices are integrated with sensors, especially the emergence of digital stethoscopes, which provides data support for model construction based on audio data. To this end, researchers have conducted some research on respiratory sound data. For example, Windmon et al. (2019) used cough sound data collected from smartphones to establish a diagnostic model for COPD and congestive heart failure; Altan et al. (2019) used the Hilbert-Huang transform method to extract statistical features from multi-channel respiratory sound data and used a deep belief network to build a COPD diagnostic model. Shuvo et al. (2021) used empirical mode decomposition and continuous wavelet transform methods to denoise respiratory sound data, and then used the denoised spectrogram as the input of the convolutional neural network for model training. In the field of audio analysis, different types of audio features play different roles in modeling (Luz et al. 2021), and most existing technologies use a single feature extraction method such as spectral analysis (You et al. 2023) or Mel spectrogram (Roy et al. 2023) as input features to build models.

[0004] Existing COPD detection technologies mainly include screening questionnaires and clinical examinations. Screening questionnaires such as the Clinical COPD Questionnaire (CCQ), the COPD Assessment Test (CAT) and the COPD Population Screening (COPD-PS) are commonly used tools for assessing the risk of COPD, but these methods are affected by the subjective factors of the filler, so the accuracy is low. In clinical practice, the risk assessment of COPD mainly relies on X-ray, CT examination and pulmonary function test, but these examination methods are complicated and costly, and are not suitable for large-scale screening and daily monitoring. At the same time, according to the survey, doctors often use auscultation to assess the patient's lung condition at the initial diagnosis to determine whether the patient has symptoms such as wheezing, rales and stridor that are more typical of COPD. However, since the breathing sounds of different respiratory diseases may be similar, such as asthma and COPD, and the difference in doctors' diagnostic experience, it is difficult to accurately classify the breath sound data by auscultation alone. Therefore, the present invention considers using the powerful modeling ability of artificial intelligence technology to establish a breath sound data classification method that can be used for early screening of COPD based on breath sound data.

[0005] By analyzing the prior art, the present application finds that there are still the following limitations:

[0006] The existing technology has not fully explored and utilized the information contained in the breathing sounds, and has only extracted a certain type of feature of the breathing sound data. For example, although the original audio signal can be converted into a time-frequency representation through a short-time Fourier transform, this processing method may lose detailed information that is important for disease risk assessment. On the one hand, feature extraction methods such as Fourier transform cannot handle nonlinear and non-stationary features, such as special acoustic patterns caused by turbulent airflow or airway stenosis; on the other hand, frequency domain feature extraction methods such as Fourier transform lack the ability to model the global structure of the breathing pattern. Therefore, the present invention considers integrating multiple feature extraction methods to construct a multi-view feature representation of respiratory sound data, thereby effectively capturing the subtle features of airway lesions.

[0007] Existing technologies are still insufficient in learning and integrating multiple types of breathing sound features. On the one hand, different types of information have significant heterogeneity. For example, acoustic features may provide frequency and intensity information about breathing sounds, while time series features may focus on feature dependencies in the time dimension. This means that different learning methods should be appropriately selected when processing different types of features. On the other hand, the importance of different types of features is different, but existing technologies mostly use subjective weighting methods to distinguish the importance of different types of features, which results in suboptimal fusion results and poor stability.

[0008] Existing technologies still have limitations in determining the best combination of parameter values. When building a model, the accuracy of the model will be greatly affected by the selection of sensitive parameters, and these models often have more than one sensitive parameter. However, existing technologies mostly use sequential tuning strategies to adjust sensitive parameters one by one, but this will cause the model to fall into a local optimal solution and fail to obtain the best combination of multiple parameter values, thus affecting the performance of the model. Summary of the invention

[0009] Taking into account the limitations of the prior art mentioned in the background technology, the present invention develops a respiratory sound data classification system based on multi-view ensemble learning, which can accurately assess the risk of chronic obstructive pulmonary disease.

[0010] The technical solution for implementing the present invention is as follows:

[0011] The respiratory sound data classification system based on multi-view ensemble learning proposed in this application mainly includes: a data preprocessing module, a multi-view feature extraction module, a multi-view ensemble learning module based on meta-learning, and a parameter optimization module based on the firefly algorithm.

[0012] First, the data preprocessing module is used to perform preprocessing tasks such as noise reduction, segmentation, and normalization on the collected original respiratory sound data, so as to obtain a standard and standardized respiratory sound data set.

[0013] In the data preprocessing module of the present invention, the noise reduction step is implemented by a spectral gating noise reduction algorithm, and the segmentation is to divide the audio sequence and ensure that the time of each segment is 20 seconds. If the breathing sound is less than 20 seconds, it is extended by filling with zeros.

[0014] Second, a multi-view feature extraction module is used to perform feature extraction of acoustic features, spectral features and time series features of respiratory sound data, so as to obtain multi-view features of respiratory sound data.

[0015] In terms of acoustic features, the acoustic features described in the present invention include: loudness, perceived width, starting detection point, duration, rhythm, and fundamental frequency. In terms of spectral features, the present invention uses Mel-frequency cepstral coefficients to construct a Mel-frequency spectrogram and extracts the spectrogram features of respiratory sounds from image data. In terms of time series features, the present application sorts and splices the sequence features generated by the Mel-frequency cepstral coefficients in time order to obtain the time series features of the respiratory sound signal. in It is the tth Mel-frequency cepstral coefficient component. In this paper, the number of components is set to T = 20, and the length of each component is 626.

[0016] Third, the multi-view ensemble learning module based on meta-learning involves the construction of three base classifiers, which are used to learn the information used for classification in acoustic features, spectral features and time series features respectively. In this module, the stacking technology is used to fuse the results output by the three base classifiers to obtain the final risk assessment result.

[0017] The three base classifiers mentioned in the present invention are: random forest classifier, convolutional neural network classifier and long short-term memory neural network classifier. The random forest classifier is used to learn the acoustic features extracted from the audio signal and output its risk assessment result for the current sample. The convolutional neural network classifier is used to learn the spectral features extracted from the Mel-language spectrogram and output its risk assessment result for the current sample The long short-term memory classifier is used to learn the time series features extracted from the Mel-level spectrogram and output its risk assessment result for the current sample.

[0018] Fourthly, the present invention establishes a parameter optimization module based on the firefly algorithm to perform multi-parameter optimization on multiple sensitive parameters involved in the multi-view ensemble learning module based on meta-learning.

[0019] When performing parameter optimization, the firefly algorithm described in the present invention will use a self-designed adaptive step size mechanism to dynamically adjust the search range, and its calculation method can be expressed as: Among them G max represents the maximum number of iterations, rand(0,0.5] is a random function, and t represents the current number of iterations. The introduction of the adaptive step size mechanism can expand the search range of the firefly algorithm in the early iteration and increase the convergence speed in the later iteration.

[0020] Beneficial effects:

[0021] First, the present invention effectively extracts multi-view features from breathing sound data by fusing various feature extraction methods, thereby achieving the complementarity of different types of information.

[0022] Second, the present invention constructs a multi-view ensemble learning model based on meta-learning for multi-view features, and improves the classification performance and robustness of the model through heterogeneous integration and stacking fusion.

[0023] Thirdly, the present invention improves the step size adjustment mechanism of the firefly algorithm and uses it to optimize multiple sensitive parameters involved in the classification system, thereby determining the best parameter value combination of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 Schematic diagram of a respiratory sound data classification system based on multi-view ensemble learning designed for the present invention. DETAILED DESCRIPTION

[0026] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0027] It should be noted that the following embodiments and features in the embodiments may be combined with each other in the absence of conflict; and, based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making any creative work are within the scope of protection of the present disclosure.

[0028] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0029] The present application embodiment is a respiratory sound data classification system based on multi-view ensemble learning. Figure 1 As shown in the figure, the system aims to use artificial intelligence technology to improve the accuracy of individual COPD risk assessment, thereby improving the efficiency and effectiveness of early screening of COPD. The system includes: data preprocessing module, multi-view feature extraction module, multi-view ensemble learning module based on meta-learning, and parameter optimization module based on the firefly algorithm.

[0030] The data preprocessing module is used to perform noise reduction and segmentation on the collected original respiratory sound data, so as to obtain standardized audio data that can be used for modeling;

[0031] Multi-view feature extraction module, used to extract acoustic features, spectral features and time series features of respiratory sound data;

[0032] The meta-learning-based multi-view ensemble learning module includes the construction of three base classifiers, which are used to learn the input acoustic features, spectral features and time series features respectively, and uses stacking technology to re-learn the outputs of the three base classifiers to obtain the final COPD risk assessment results.

[0033] The parameter optimization module based on the firefly algorithm is used to perform combinatorial optimization on the key sensitive parameters involved in the multi-view feature extraction module in the classification system and the multi-view ensemble learning module based on meta-learning, so as to determine the best parameter value combination suitable for the current task.

[0034] This embodiment extracts the features of breathing sound data from three perspectives: acoustics, spectrogram and time series, and uses the complementary information of different feature types to improve the model accuracy; secondly, a heterogeneous integration framework containing three base classifiers is designed. These base classifiers are used to learn information of different types of features, and stacking technology is introduced to improve the robustness of the model; in addition, the efficiency and results of the parameter optimization process can be improved by improving the step size adjustment mechanism of the firefly algorithm.

[0035] The following is a detailed description of each module of the system:

[0036] The data preprocessing module in the embodiment of the present application is used to perform noise reduction and segmentation on the collected original sound data, so as to obtain standardized audio data that can be used for modeling.

[0037] Noise reduction: The original sound recording contains many irrelevant components that may affect the classification accuracy of the model. Therefore, referring to the research of Sainburg, Thielk, and Gentner, the waveform S of the breathing sound is n , a spectral gated noise reduction algorithm was used to maintain consistency and minimize the loss of key breathing sound signals. The sampling rate of the input signal was set to 14Hz to ensure consistency while reducing the loss of breathing sound signals. Afterwards, the sound signal was normalized to reduce the impact of signal changes on the classification results.

[0038] Segmentation: Typically, each breathing sound recording should consider a sample duration of at least 3 seconds to capture useful breathing sound information. Considering the breathing pattern of COPD, this application chooses to extend the duration of each segment to 20 seconds to allow for a more detailed analysis of the acoustic pattern. If the breathing sound is longer than 20 seconds, the excess will be omitted. Conversely, records less than 20 seconds will be extended by zero padding to reach this predetermined length to ensure consistency in the dataset. Finally, a cleaned breathing sound dataset can be obtained. Where N is the number of samples, Represents the nth breath sound segment in the dataset.

[0039] The multi-view feature extraction module in the embodiment of the present application is used to extract acoustic features, spectral features and time series features of breathing sound data.

[0040] In the field of machine learning, the performance of downstream tasks depends heavily on feature extraction. Inspired by multi-view learning, this application uses different audio data processing methods to extract diverse respiratory sound features. The extracted features mainly include three parts: (1) acoustic features; (2) spectral features; and (3) time series features.

[0041] Acoustic features: Relevant acoustic features can be extracted based on the waveform of the breathing signal.

[0042] (1) Loudness is mainly used to characterize the intensity of sound, which can be calculated using the following formula: Among them, M ERB Indicates the number of Equivalent Rectangular Bandwidth (ERB) scales, is the energy of the signal in the mth bandwidth.

[0043] (2) Perception width S p The timbre width of the breathing sound can be characterized and is calculated as:

[0044]

[0045] (3) Onset detection is used to identify sudden and sharp changes in the breathing sound signal, which usually represents the starting point of the breathing sound event. This application performs onset detection by calling the librosa package in Python.

[0046] (4) The length of the audio recording represents the duration of the breathing sound and is calculated as follows: in Indicates the length of the sound clip, and sample rate indicates the sampling rate.

[0047] (5) Rhythm can be used to describe the rhythmic nature or frequency characteristics of breathing sounds. This application calculates this feature by using the librosa package in Python.

[0048] (6) The fundamental frequency (FF) is the lowest major frequency component in the respiratory sound signal, which usually corresponds to the fundamental vibration frequency of the lungs during breathing and sound production. It can be calculated using the librosa package in Python.

[0049] Spectral features: Mel frequency cepstral coefficients (MFCC) are one of the most widely used tools for spectral feature extraction in audio processing. This application also uses MFCC to represent the spectral features of breathing sounds. Specifically, MFCC is a series of discrete cosine transform (DCT) coefficients obtained from a specific cepstral representation of an audio clip. The calculation process is mainly as follows: ① Obtain the spectrum by Fourier transforming the signal; ② Map the spectrum to the Mel scale: This method is similar to human auditory perception; ③ After mapping to the Mel scale, the Mel frequency scaled spectrum is logarithmically transformed; ④ The Mel logarithmic frequency scaled spectrum is discrete cosine transformed; ⑤ The amplitude spectrum is obtained, that is, the Mel frequency cepstrum coefficients containing important information of the original audio signal are obtained.

[0050] Time series features: Based on the results of MFCC calculation, this application arranges the features generated by MFCC in time order to obtain the time series features of the breathing sound signal in It is the t-th MFCC component. This paper sets the number of components to T = 20, and the length of each component is 626.

[0051] The multi-view ensemble learning module based on meta-learning in the embodiment of the present application includes the construction of three base classifiers, which are respectively used to learn the input acoustic features, spectral features and time series features, and use stacking technology to achieve re-learning of the outputs of the three base classifiers, so as to obtain the final COPD risk assessment result.

[0052] Based on the multi-view features obtained by the feature extraction module, three different base classifiers (i.e., random forest algorithm, convolutional neural network algorithm, and long short-term memory neural network algorithm) are used to learn the mapping function from features to output results from different perspectives.

[0053] Random Forest Algorithm (RF): Random forest is an integrated learning method composed of multiple decision trees. Each decision tree gradually establishes a tree hierarchy by performing binary segmentation on the feature values. In view of the complex nonlinear relationship between the acoustic features of respiratory sounds, and the random forest is just suitable for handling such tasks, this application uses random forest to construct an acoustic feature classifier. Specifically, given the acoustic feature vector A = {a1, a2, ..., a6} composed of six acoustic features extracted from the respiratory sound data, the vector A is input into the random forest model to construct each decision tree, and the final output of the random forest is Determined by majority vote of all decision trees.

[0054] Convolutional Neural Network (CNN): CNN has been widely used in various pattern recognition tasks, including processing data of different modalities such as images and videos. When processing audio data, the Mel-spectrogram can display the energy of each frequency band from the time dimension and convert the audio signal into an image format, and the advantage of CNN is to mine the potential patterns contained in the image. Therefore, this application is based on the image data generated by the Mel-spectrogram, and introduces a CNN architecture consisting of a convolutional layer, a pooling layer, and a fully connected layer to learn the information contained in the Mel-spectrogram, thereby establishing a base classifier for the image view.

[0055] Specifically, the convolution layer uses convolution kernels to extract features within a certain range in the image, and introduces a maximum pooling layer to further compress the dimension of the features. Finally, the results obtained by the maximum pooling layer are input into a fully connected layer to learn the mapping function between features and output results. This process can be expressed as follows:

[0056]

[0057] Where X is the image input, usually a matrix, and K represents the convolution kernel, whose size is a key to be tuned. Ginseng number, b is the bias term of the convolutional layer, φ(·) represents the ReLU activation function, represents the dot product operator, Maxpooling(·) represents the maximum pooling, and FC(·) represents the fully connected layer.

[0058] Long Short-Term Memory Neural Network Algorithm (LSTM): Considering the temporal nature of respiratory sound data, this application uses LSTM to capture the sequence dependency of each sampling data point of respiratory sound. For respiratory sound sequence data with a time step of t, the output h t The formula h t =LSTM(x t ,c t-1 ,h t-1 ) is calculated, where

[0059] c t-1 is the cell state at time t-1. Then, the output of the LSTM layer Will

[0060] It is input into the average pooling layer and the fully connected layer in turn for calculation. The calculation method is:

[0061]

[0062] Where Meanpooling(·) represents the average pooling operation, and FC(·) represents the fully connected layer.

[0063] Using stacking technology to fuse the output results of three base classifiers: Stacking technology is a meta-learning method that aims to improve the overall performance of the system by combining multiple models (also called base classifiers or primary models) in a structured way. Compared with popular ensemble learning methods such as random forest and XGBoost, stacking technology introduces a meta-classifier to integrate the outputs of each base classifier, thereby enhancing the robustness and flexibility of the system. Based on the stacking technology, this application proposes multi-view ensemble learning based on meta-learning, and its calculation process is shown in Algorithm 1. It can be seen from the calculation process that stacking trains the secondary model by using the output of the base classifier as the input of the meta-classifier, thereby enhancing the robustness of the system.

[0064]

[0065]

[0066] The parameter optimization module based on the firefly algorithm in the embodiment of the present application is used to perform combined optimization on the key sensitive parameters involved in the multi-view feature extraction module in the classification system and the multi-view ensemble learning module based on meta-learning, so as to determine the best parameter value combination suitable for the current task.

[0067] In the process of using the firefly algorithm to optimize the model parameters, in order to enhance the efficiency and effectiveness of the firefly algorithm in finding the optimal solution, an innovative adaptive step size mechanism was designed, and the step size adjustment factor in the original algorithm was improved to enhance the algorithm's ability to avoid falling into local optimality. At the same time, the adaptive step size adjustment improves the algorithm's search ability in the early stages and its convergence ability in the later stages.

[0068] The step size adjustment factor can be expressed as: G max represents the maximum number of iterations, rand(0,0.5] is a random function, and t represents the current number of iterations.

[0069] The following is a detailed description of the parameter optimization process based on the firefly algorithm:

[0070] Since many machine learning and deep learning methods are parameter-sensitive methods, the values ​​of key parameters can significantly affect the performance of the system, so it is necessary to select appropriate parameters. At present, most studies rely on manual adjustment or grid search for parameter optimization, but these methods cannot solve the multi-parameter optimization problem. In this application, several key parameters such as the size and number of convolution kernels in CNN, the size and number of hidden layers in LSTM models, etc. will affect the performance of the system. Therefore, the present embodiment introduces the firefly algorithm to solve the combinatorial optimization problem of multiple parameters.

[0071] The firefly algorithm (FA) is an optimization algorithm that simulates the foraging behavior of fireflies in nature. It finds the optimal solution by adjusting the attractiveness and brightness of fireflies. In FA, the brightness of the individual represents the quality of the solution, and the attractiveness represents the degree of attraction between individuals. Each firefly tends to move closer to individuals with higher brightness. When solving the optimization problem, FA iteratively updates the position of the individual to find the optimal solution. The parameter optimization process based on the firefly algorithm designed in this application can be divided into five steps: population initialization, brightness calculation, attractiveness calculation, position update and iterative optimization. The specific calculation process of each step is as follows:

[0072] (1) Population initialization

[0073] Set the basic parameters of the firefly algorithm, such as the number of fireflies ps, initial attraction β0, light absorption coefficient γ, and maximum number of iterations G max Etc., and based on this, the first generation of firefly groups is generated 1 , where each individual Represents a candidate solution.

[0074] (2) Brightness calculation

[0075] According to the fitness function (Formula 3) designed by this application, the brightness of each firefly can be calculated, and a higher brightness value indicates a better candidate solution. In order to achieve the best prediction performance, this application uses the AUC value obtained by K-fold cross validation as the fitness function, which is calculated as follows:

[0076]

[0077] Where K represents the number of cross-validation experiments, AUC(·) is the calculation function of the comprehensive evaluation index AUC, and y pre Represents the prediction result of the model, y act Represents the actual label of the sample.

[0078] (3) Calculation of attractiveness

[0079] In order to determine the relative positions of individual fireflies, we need to calculate their attraction. Each firefly needs to calculate its attraction to each other firefly to provide support for the next update of the firefly position. Suppose there are two fireflies firefly i and firefly j , the attraction between them β(r i,j ) is calculated as:

[0080]

[0081] Where β0 is the attraction when the distance r = 0, γ is the light absorption coefficient, r ij Indicates firefly i and firefly j The distance between them is calculated as:

[0082]

[0083] Where d = 1, 2, ..., D, D represents the number of parameters to be optimized, i,d and fir j,d They respectively represent the d-th parameter candidate value of the firefly individual.

[0084] (4) Location Update

[0085] By comparing the brightness of fireflies in pairs, we can determine the direction of movement of each firefly. Fireflies with lower brightness will be attracted to fireflies with higher brightness. Suppose in the tth iteration, if Brightness ratio Weak, then update The location is updated as follows:

[0086]

[0087] in, is the step size adjustment factor, and rand(0,0.5] is a random function. In order to enhance the efficiency and effectiveness of the firefly algorithm in finding the optimal solution, this application improves the step size adjustment factor in the original algorithm and introduces an adaptive search mechanism with random perturbations to avoid falling into the local optimal solution during the optimization process. At the same time, the adaptive step size improves the search ability of the algorithm in the early stage and the convergence ability in the later stage.

[0088] (5) Iterative Optimization

[0089] Repeat steps (2) to (4) iteratively until the algorithm reaches a predetermined maximum number of iterations G. max ,Finally, the candidate solution corresponding to the firefly with the largest fitness value in the population is the optimal parameter value combination.

[0090] In this embodiment, the establishment of the COPD risk assessment model mainly uses respiratory sound data collected by a digital stethoscope, but other types of data such as questionnaire results and medical images (X-rays, CT) can also be integrated when establishing the model.

[0091] In addition, this application verified the effectiveness of the proposed method by conducting a series of experiments on the ICBHI 2017 dataset. The comparison results with other baseline methods are shown in Table 1. The results show the advantages of this application in the COPD risk assessment task.

[0092] Table 1 Comparison results of this application with other baseline methods

[0093]

[0094] Note: Precision indicates the proportion of samples that are actually high-risk among all samples classified as high-risk by the model; Recall indicates the proportion of samples that are actually high-risk and are classified as high-risk by the model; Accuracy indicates the proportion of correct classification results of the model among all samples; F1-score is the harmonic mean of Precision and Recall; AUC is the area under the receiver operating curve, which indicates the comprehensive performance of the model at different decision thresholds.

[0095] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A respiratory sound data classification system based on multi-view ensemble learning, which is characterized by comprising a data preprocessing module, a multi-view feature extraction module, a multi-view ensemble learning module based on meta-learning, and a parameter optimization module based on the firefly algorithm: The data preprocessing module is used to perform noise reduction and segmentation on the collected original respiratory sound data, so as to obtain standardized audio data that can be used for modeling; Multi-view feature extraction module, used to extract acoustic features, spectral features and time series features of respiratory sound data; The meta-learning-based multi-view ensemble learning module includes the construction of three base classifiers, which are used to learn the input acoustic features, spectral features, and time series features respectively, and use the stacking technology to learn the outputs of the three base classifiers again to obtain the final COPD risk assessment results; The parameter optimization module based on the firefly algorithm is used to perform combinatorial optimization on the key sensitive parameters involved in the multi-view feature extraction module in the classification system and the multi-view ensemble learning module based on meta-learning, so as to determine the best parameter value combination suitable for the current task.

2. According to claim 1, the respiratory sound data classification system based on multi-view ensemble learning is characterized in that the three base classifiers are: a random forest classifier, a convolutional neural network classifier, and a long short-term memory neural network classifier: The random forest classifier is used to learn the acoustic features extracted from the audio signal and output its risk assessment result for the current sample; The convolutional neural network classifier is used to learn the frequency spectrum features extracted from the Mel-language spectrogram and output its risk assessment result for the current sample; The long short-term memory classifier is used to learn the time series features extracted from the Mel-level spectrogram and output its risk assessment result for the current sample.

3. According to the respiratory sound data classification system based on multi-view ensemble learning in claims 1 and 2, its characteristic is that the stacking technology in the multi-view ensemble learning module based on meta-learning uses a logistic regression algorithm to fuse the output results of the three base classifiers.

4. According to the respiratory sound data classification system based on multi-view ensemble learning in claim 2, its characteristic is that the key parameter values ​​in the multi-view ensemble learning module based on meta-learning are obtained by parameter optimization using the firefly algorithm.

5. The respiratory sound data classification system based on multi-view ensemble learning according to claim 4 is characterized in that the firefly algorithm adopts an adaptive step size mechanism, and the step size adjustment factor is: G max represents the maximum number of iterations, rand(0,0.5] is a random function, and t represents the current number of iterations.

6. According to claim 1, the respiratory sound data classification system based on multi-view ensemble learning is characterized in that, in the data preprocessing module, the noise reduction operation is implemented using a spectral gating noise reduction algorithm, and the segmentation operation sets the segmentation time of the respiratory sound segment to 20 seconds. If the respiratory sound is longer than 20 seconds, the excess part will be omitted, and the record less than 20 seconds will be extended by zero padding.

Images