A lung sound digital biomarker signal processing method and system

Abstract

The present application relates to a kind of lung sound digital biomarker signal processing method and system.The method comprises: based on wavelet transform method, lung sound digital biomarker signal is decomposed at least one layer and at least one reconstructed signal is obtained;Select the high-frequency wavelet coefficient after the last layer decomposition and extract energy value;Based on energy value, construct the feature vector related to respiratory cycle, and the feature vector is reduced dimension to obtain lung sound feature.Based on the dimension of the lung sound digital signal in the prior art is less so that the defect that the error of analysis result is larger, the present application is decomposed by wavelet transform method to lung sound digital biomarker signal multiple layer, obtains the lung sound feature of multidimensional number, so as to be able to analyze the rich lung sound feature of lung, and the judgment of breathing condition is more accurate.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and in particular to a method and system for processing digital biomarker signals of lung sounds. Background Technology

[0002] Lung sounds typically include moist rales, dry rales, and wheezing, usually caused by air passing through the moist trachea or bronchi. In addition, wheezing is another lung sound. Currently, stethoscopes are commonly used to collect lung sounds, and computer-aided analysis of these sounds efficiently helps doctors save diagnostic time and reduce diagnostic errors caused by overwork. It also helps patients conduct self-screening and monitoring of their diseases.

[0003] With the development of artificial intelligence technology, advancements in smart healthcare and telemedicine have provided opportunities for the processing of lung sound signals. Many researchers have continuously improved methods for noise reduction and recognition of lung breath sounds based on computer analysis, providing a reliable approach for the objective assessment of lung health. Through research on digital signal recording and lung sound signal characteristics, the storage, analysis, and communication of lung breath sounds have been achieved, overcoming the limitations of traditional auscultation, such as reliance on manual methods, inconvenient recording, and inability to perform repeatable analyses. In many areas, including home healthcare and telemedicine services, lung sound processing methods based on digital auscultation have become a more effective and economical tool for diagnosing lung abnormalities, effectively reducing the harm caused by other examination methods.

[0004] However, current methods for analyzing lung sounds generally involve simple improvements and analyses based on wavelet transform. Due to the limited data dimensionality, the resulting analysis often deviates significantly from the actual lung structure.

[0005] For example, patent application CN116052724A discloses a lung sound enhancement method, system, device, and storage medium. The method includes acquiring lung sound data to be enhanced; converting the lung sound data to be enhanced using a trained lung sound enhancement model to obtain enhanced lung sound data; the lung sound enhancement model is constructed based on a first variational autoencoder, a second variational autoencoder, and a cycle consistency loss model, and is trained using clean lung sound data and noisy lung sound data to obtain the enhanced lung sound data. By using the cycle consistency loss model and noisy lung sound data, the training process of the lung sound enhancement model does not require a large amount of clean lung sound data.

[0006] For example, patent application CN107837091A discloses a single-channel cardiopulmonary sound separation method and system, wherein the method includes: step 1, processing the acquired cardiopulmonary sound mixed signal x(t) to obtain the time spectrum X of the cardiopulmonary sound mixed signal x(t); step 2, taking the column average of the 20Hz to 100Hz part of the time spectrum X to obtain the amplitude-mode time sequence template H0 of the heart sound; step 3, constructing a regularization term of the heart sound time sequence structure based on the amplitude-mode time sequence template H0 of the heart sound, performing non-negative matrix decomposition on the time spectrum X, and extracting the estimated time spectrum Xc of the heart sound signal and the estimated time spectrum Xr of the lung sound signal; step 4, reconstructing the time domain signals of the heart sound and lung sound based on the estimated time spectrum Xc of the heart sound signal and the estimated time spectrum Xr of the lung sound signal.

[0007] The aforementioned existing technologies process lung sound signals, all resulting in high-dimensional vectors. The drawbacks of processing high-dimensional vectors are twofold: First, the curse of dimensionality exists: high-dimensional vectors typically require more computational and storage resources, which can slow down processing speed, increase algorithm complexity, and necessitate more data to train the model to avoid overfitting. Second, there is a model generalization problem: high-dimensional vectors may cause machine learning models to overfit the training data, as noise or irrelevant features in the training data can interfere with the model's generalization performance.

[0008] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0009] Existing methods for analyzing lung sounds are limited to simple algorithmic analysis of digital biomarker signals. The drawback of this approach is that the analysis of digital biomarker signals has a single dimension, making it impossible to accurately assess the respiratory status of the lungs. Therefore, the technical problem this invention aims to solve is how to perform multidimensional analysis on digital biomarker signals of lung sounds collected from the lungs to obtain accurate respiratory characteristics.

[0010] To address the shortcomings of existing technologies, this invention discloses a method for processing digital biomarker signals of lung sounds from a first aspect. The method may include: decomposing the digital biomarker signal of lung sounds into at least one layer based on wavelet transform and obtaining at least one reconstructed signal; selecting the high-frequency wavelet coefficients after the last layer of decomposition and extracting the energy value; constructing a feature vector related to the respiratory cycle based on the energy value; and performing dimensionality reduction processing on the feature vector to obtain lung sound features.

[0011] This invention decomposes the digital biomarker signal of lung sounds multiple times to obtain reconstructed signals across multiple frequency bands. The reconstructed signal comprises high-frequency wavelet coefficients and low-frequency wavelet coefficients. The low-frequency component exhibits a smoother characteristic, while the high-frequency component contains more detailed information. Multi-level decomposition of the reconstructed signal allows for the acquisition of more detailed information, providing rich and detailed data for subsequent respiratory feature analysis.

[0012] According to a preferred embodiment, the step of decomposing the lung sound digital biomarker signal at least one level may include: decomposing the lung sound digital biomarker signal based on a wavelet transform method to obtain high-frequency wavelet coefficients and low-frequency wavelet coefficients, and forming a reconstructed signal based on the sum of the coefficients of the high-frequency and low-frequency wavelet coefficients. The high-frequency features of the reconstructed signal include more respiratory details. Therefore, decomposing the high-frequency and low-frequency features allows for the extraction of more data related to respiratory characteristics.

[0013] According to a preferred embodiment, the step of decomposing the digital biomarker signal of lung sounds at least one level further includes: decomposing the high-frequency wavelet coefficients and low-frequency wavelet coefficients obtained from each level of decomposition again to obtain a plurality of high-frequency wavelet coefficients; and sorting the plurality of high-frequency wavelet coefficients according to frequency range to represent the original signal of lung sounds. Decomposing each reconstructed signal again is equivalent to repeatedly extracting detail-related data from the respiratory features, ultimately obtaining an original signal containing multiple dimensions. The respiratory features extracted from the multi-dimensional original signal are sufficiently rich, which is beneficial for accurate judgment of lung conditions.

[0014] According to a preferred embodiment, the digital biomarker signal of lung sounds is decomposed into at least three levels based on a wavelet transform method. If the number of decomposition levels is less than three, the dimension of the feature vector will be low, or even form a single high-frequency wavelet coefficient, making it impossible to extract multiple details of respiration in data form. Performing a three-level decomposition can obtain high-frequency wavelet coefficients with three dimensions, ultimately forming a feature vector with three dimensions.

[0015] According to a preferred embodiment, during the further decomposition of the high-frequency and low-frequency wavelet coefficients obtained from each level of decomposition, the low-frequency wavelet coefficients obtained from the high-frequency wavelet coefficient decomposition are discarded. The low-frequency wavelet coefficients obtained from the previous level of decomposition are decomposed again, and the resulting new high-frequency and low-frequency wavelet coefficients are retained. Since the low-frequency portion has a relatively smooth characteristic and contains almost no respiratory details, the low-frequency wavelet coefficients are discarded. The high-frequency wavelet coefficients that can provide more respiratory details are further extracted. The more levels of decomposition, the more high-frequency wavelet coefficients related to respiratory details are extracted.

[0016] According to a preferred embodiment, the step of constructing a feature vector related to the respiratory cycle may include: after decomposing the digital biomarker signal of lung sounds during expiration layer by layer, calculating the energy based on the high-frequency wavelet coefficients obtained from the last layer of decomposition and constructing a first feature vector; after decomposing the digital biomarker signal of lung sounds during inhalation layer by layer, calculating the energy based on the high-frequency wavelet coefficients obtained from the last layer of decomposition and constructing a second feature vector; and normalizing the feature vector related to the respiratory cycle formed by combining the first and second feature vectors to obtain the total energy of the feature vector. Only feature vectors extracted from the entire respiratory process of the respiratory cycle can comprehensively reflect the characteristics of lung sounds. Normalizing the feature vectors can eliminate the differences in energy magnitude between signals.

[0017] According to a preferred embodiment, the step of dimensionality reduction of the feature vector may include: reducing the dimensionality of the normalized feature vector based on the PCA algorithm. Dimensionality reduction helps reduce the computational load of subsequent processing without compromising the accuracy of respiratory feature extraction.

[0018] The present invention provides a signal processing system for digital biomarkers of lung sounds from a second aspect, comprising at least one processor. The processor is configured to: decompose the digital biomarker signal of lung sounds into at least one layer based on a wavelet transform method to obtain at least one reconstructed signal; select the high-frequency wavelet coefficients after the last layer of decomposition and extract the energy value; construct a feature vector related to the respiratory cycle based on the energy value; and perform dimensionality reduction processing on the feature vector to obtain lung sound features.

[0019] The system of this invention involves minimal hardware and can be flexibly deployed at the data acquisition site, or even remotely. The lung sound features extracted by the processor of this invention have more dimensional data and represent richer information.

[0020] According to a preferred embodiment, the processor is further configured to: further decompose the high-frequency wavelet coefficients and low-frequency wavelet coefficients obtained from each decomposition layer to obtain a plurality of high-frequency wavelet coefficients; and sort the plurality of high-frequency wavelet coefficients according to their frequency range to represent the original signal of lung sounds. The reconstructed original signal obtained by the system of the present invention can repeatedly extract high-frequency wavelet coefficients that reflect respiratory details, thus providing richer information about respiration.

[0021] According to a preferred embodiment, the system further includes a plurality of lung sound acquisition components connected to the processor via wired or wireless means. The plurality of lung sound acquisition components acquire digital biomarker signals of lung sounds during at least one respiratory cycle. By configuring the lung sound acquisition components, it is beneficial to acquire lung digital biomarker signals in real time. Attached Figure Description

[0022] Figure 1 This is a diagram of a three-layer wavelet decomposition structure according to a preferred embodiment of the present invention; Figure 2 This is a simplified human body placement diagram of a lung sound acquisition component according to a preferred embodiment of the present invention; Figure 3 This is a simplified human placement diagram of a lung sound acquisition component according to another preferred embodiment of the present invention.

[0023] List of reference numerals 1: Lung sound acquisition component; 11: First sound sensor; 12: Second sound sensor; 13: Third sound sensor; 14: Fourth sound sensor; 2A: First point-like sound sensing element; 2B: Second point-like sound sensing element; 2C: Third point-like sound sensing element; 2D: Fourth point-like sound sensing element; 3A: Fifth point-like sound sensing element; 3B: Sixth point-like sound sensing element; 3C: Seventh point-like sound sensing element; 3D: Eighth point-like sound sensing element; 3E: Ninth point-like sound sensing element. Detailed Implementation

[0024] The following is a detailed explanation with reference to the accompanying drawings.

[0025] This invention provides explanations and clarifications for some terms and concepts.

[0026] Low frequency: In wavelet decomposition, the low-frequency component represents the main, lower-frequency components of the signal. This part contains the overall trend and slowly changing features of the signal. High frequency: The high-frequency component contains the high-frequency components and subtle fluctuations in the signal. This part is used to capture local details and rapidly changing features of the signal.

[0027] Low-frequency wavelet coefficients: refer to the low-frequency signal portion after wavelet transform.

[0028] High-frequency wavelet coefficients: refer to the high-frequency portion of the signal after wavelet transform.

[0029] In this invention, the division between high and low frequencies is based on the original wavelet transform. The signal is divided into two parts—a low-pass part and a high-pass part—by a high-pass filter and a low-pass filter. Low-frequency and high-frequency wavelet coefficients do not have fixed frequency ranges. They are two relative concepts obtained after performing one wavelet transform. After one wavelet transform, the frequency range of the low-frequency signal passing through the low-pass filter is called the low-frequency wavelet coefficients; the frequency range of the high-frequency signal passing through the high-pass filter is called the high-frequency wavelet coefficients.

[0030] By employing multi-level decomposition, a signal can be broken down into components at different frequencies and scales, thereby better capturing its diversity and complexity. This enhances the signal's feature extraction capabilities, enabling more accurate identification of key features. Further subdivision of the high-frequency components is intended to capture more local information, allowing for more precise analysis of the signal's fine structure.

[0031] Digital biomarker signals of lung sounds: These are digital signals of lung sounds that can represent a living organism, formed by converting collected lung sounds.

[0032] The lung sound digital biomarker signals in this invention can be collected in real time from animals or humans with lung tissue, or obtained from a lung sound database. Lung sound signals are collected using a digital electronic stethoscope. When collecting lung sound digital biomarker signals, lung sound data for each individual are acquired during six minutes of normal sitting breathing and six minutes of walking breathing tests. From these, 50 typical normal lung sounds, 50 COPD lung sounds, and 50 AECOPD lung sound data are selected to form a lung sound database containing 150 samples.

[0033] Example 1 Existing methods for analyzing lung sounds are limited to simple algorithmic analysis of digital biomarker signals. The drawback of this approach is that the analysis of digital biomarker signals has a single dimension, making it impossible to accurately assess the respiratory status of the lungs. Therefore, the technical problem this invention aims to solve is how to perform multidimensional analysis on digital biomarker signals of lung sounds collected from the lungs to obtain accurate respiratory characteristics.

[0034] To address the shortcomings of existing technologies, this invention discloses a method for processing digital biomarker signals of lung sounds, the method including: S1: Based on the wavelet transform method, the digital biomarker signal of lung sounds is decomposed into at least one layer to obtain at least one reconstructed signal.

[0035] S2: Select the high-frequency wavelet coefficients after the last layer of decomposition and extract the energy value.

[0036] S3: Construct a feature vector related to the respiratory cycle based on energy values.

[0037] S4: Dimensionality reduction of the feature vector is performed to obtain lung sound features.

[0038] This invention decomposes the digital biomarker signal of lung sounds multiple times to obtain reconstructed signals across multiple frequency bands. The reconstructed signal comprises high-frequency wavelet coefficients and low-frequency wavelet coefficients. The low-frequency component exhibits a smoother characteristic, while the high-frequency component contains more detailed information. Multi-level decomposition of the reconstructed signal allows for the acquisition of more detailed information, providing rich and detailed data for subsequent respiratory feature analysis.

[0039] When lung sound digital biomarker signals are used directly for analysis and processing, they contain a significant amount of background noise. Therefore, it is best to preprocess lung sound digital biomarker signals.

[0040] The preprocessing steps for the digital biomarker signal of lung sounds are as follows. An adaptive LMS (Least Mean Square) filter is used to reduce noise in the signal. The acquired lung sound signal inevitably contains some heart sound components, and the two highly overlap in frequency. This paper uses a thresholding method to remove the interference of heart sounds from the wavelet coefficients obtained by decomposition, resulting in relatively pure wavelet coefficients for the lung sound signal.

[0041] The principle of the wavelet transform method in this invention is as follows.

[0042] For the square-integrable signal to be analyzed Its continuous wavelet transform is: The input signal is denoted as 'a', and 'b' are the scale and translation parameters. It is the complex conjugate of the wavelet function. Using wavelet transform, digital biomarker signals of lung sounds can be mapped to a space composed of wavelet basis functions. By increasing parameter 'a', higher frequency resolution can be obtained when analyzing low-frequency signals; by decreasing parameter 'a', higher time resolution can be obtained when analyzing high-frequency signals. Digital biomarker signals of lung sounds can be obtained through wavelet coefficients. Refactor.

[0043] The principle of the discrete wavelet transform method in this invention is as follows.

[0044] Wavelets obtained by raising a to the power of 2 are usually called "binary" wavelets, i.e. At this point, the wavelet basis function can be expressed as: At this time, the signal The discrete wavelet transform is: The digital biomarker signal of lung sounds can also be obtained from Obtained through reconstruction.

[0045] Therefore, the decomposition method mentioned in this invention can refer to either the wavelet transform method or the discrete wavelet transform method. In the process of decomposing lung sound digital biomarker signals layer by layer, the wavelet transform method and the discrete wavelet transform method can also be used in combination at different decomposition levels.

[0046] This invention preferably uses binary wavelet transform to decompose the signal multiple times to obtain reconstructed signals in multiple frequency bands. The low-frequency components have smoother characteristics, while the high-frequency components contain more detailed information. When analyzing high-frequency components, the wavelet transform corresponds to the rapidly changing parts of the time-domain signal. In this case, the time-domain resolution needs to meet the distance requirements between rapidly changing elements, while the frequency resolution requirement is relatively lower and easier to meet. However, when analyzing low-frequency components, the wavelet transform corresponds to the slowly changing parts of the signal. In this case, more emphasis is placed on frequency-domain resolution, while time-domain resolution is relatively easier to meet. Furthermore, the center frequency of the signal analysis should be shifted to a lower frequency position. Therefore, based on the advantages of binary wavelet transform, this invention performs layer-by-layer decomposition of lung sound digital biomarker signals.

[0047] S1: Based on the wavelet transform method, the digital biomarker signal of lung sounds is decomposed into at least one layer to obtain at least one reconstructed signal.

[0048] S11: The digital biomarker signal of lung sounds is decomposed based on the wavelet transform method to obtain high-frequency wavelet coefficients and low-frequency wavelet coefficients. The reconstructed signal is formed by the sum of the coefficients of the high-frequency wavelet coefficients and low-frequency wavelet coefficients.

[0049] This invention uses the Coieflet wavelet decomposition method as an example to illustrate the signal processing method for digital biomarkers of lung sounds.

[0050] initial signal It can be expressed by the following formula: in N represents the number of samples; h and g represent the reconstructed signals of the low-pass and high-pass filters, respectively. Indicates the signal in the i-th layer The low-frequency wavelet coefficients. Indicates the signal in the i-th layer The high-frequency wavelet coefficients.

[0051] The reconstructed signal obtained after decomposition can be represented as: The high-frequency features of the reconstructed signal contain many respiratory details. Therefore, decomposing the high-frequency and low-frequency features allows for the extraction of more data related to respiratory characteristics.

[0052] Low-frequency components typically represent the general outline of a signal, excluding subtle variations and rapid oscillations. They are crucial for capturing the global information and overall structure of a signal. High-frequency components are very useful for describing abrupt changes, noise, or sharp edges in a signal. They provide local details of the signal, enabling more precise analysis of its fine structure. Decomposing a signal into low-frequency and high-frequency components allows for the selective preservation or discarding of high-frequency details. This property is extremely useful for signal compression, as data compression can be achieved by discarding some high-frequency wavelet coefficients while preserving the signal's main structure.

[0053] S12: Decompose the high-frequency wavelet coefficients and low-frequency wavelet coefficients obtained from each layer of decomposition again to obtain several high-frequency wavelet coefficients.

[0054] This invention can decompose the digital biomarker signal of lung sounds into a single layer to obtain a reconstructed signal with a single dimension. This invention can also decompose the digital biomarker signal of lung sounds into a two-layer layer to obtain a reconstructed signal with two dimensions. Furthermore, this invention can decompose the digital biomarker signal of lung sounds into more than two layers to obtain a reconstructed signal with more than two dimensions.

[0055] Preferably, the present invention decomposes the digital biomarker signal of lung sounds into at least three levels. If the number of decomposition levels is less than three, the dimension of the feature vector will be low, or even a single high-frequency wavelet coefficient will be formed, making it impossible to extract multiple details of respiration in the form of data. Performing a three-level decomposition can obtain high-frequency wavelet coefficients with three dimensions, ultimately forming a feature vector with three dimensions.

[0056] This invention illustrates the steps of performing a three-layer decomposition of digital biomarker signals of lung sounds as an example.

[0057] like Figure 1 As shown, the digital biomarker signal of lung sounds is decomposed into its first level using wavelet decomposition to obtain the first low-frequency wavelet coefficients. and the first high-frequency wavelet coefficients .

[0058] The first low-frequency wavelet coefficient Perform a second-level decomposition to obtain the second low-frequency wavelet coefficients. Second high frequency wavelet coefficients .

[0059] The first high-frequency wavelet coefficient Perform a second-level decomposition to obtain the third high-frequency wavelet coefficients. .

[0060] The second low-frequency wavelet coefficient The third-level decomposition yields the third low-frequency wavelet coefficients. and the fourth high-frequency wavelet coefficients .

[0061] The second high-frequency wavelet coefficient Perform a third-level decomposition to obtain the fifth high-frequency wavelet coefficients. .

[0062] The third high-frequency wavelet coefficient Perform a third-level decomposition to obtain the sixth high-frequency wavelet coefficients. .

[0063] like Figure 1 As shown, the digital biomarker signal of lung sounds is decomposed into three levels, and the reconstructed signal includes the fourth high-frequency wavelet coefficient. Fifth high-frequency wavelet coefficients and the sixth high-frequency wavelet coefficient It has three dimensions.

[0064] The frequency of the digital biomarker signal for lung sounds ranges from 100 to 2000 Hz. Based on the Nyquist sampling theorem, the sampling frequency of the experimental signal is set to 4000 Hz for the digital biomarker signal for lung sounds. Observations show that the frequency components of the digital biomarker signal for lung sounds above 1000 Hz are extremely few. Therefore, this invention allows the frequency range of the reconstructed signal to be analyzed to be set to 100–1000 Hz. This frequency range setting also removes the heart sound component.

[0065] Decomposing each reconstructed signal again is equivalent to repeatedly extracting the detailed data from the respiratory features, ultimately obtaining the original signal containing multiple dimensions. The respiratory features extracted from the multi-dimensional original signal are rich enough to facilitate accurate judgment of the lung condition.

[0066] The present invention can also be illustrated by the step of performing a five-level decomposition of the digital biomarker signal of lung sounds as an example.

[0067] The original signal frequency range of the lung sound digital biomarker signal is 0~2000Hz.

[0068] First-level decomposition: The digital biomarker signal of lung sounds was decomposed to obtain the first low-frequency wavelet coefficient. and the first high-frequency wavelet coefficients First low-frequency wavelet coefficients The frequency range is 0~1000Hz. First high-frequency wavelet coefficients. The frequency is 1001~2000Hz.

[0069] Second-level decomposition: The first low-frequency wavelet coefficient Decomposition is performed to obtain the second low-frequency wavelet coefficients. Second high frequency wavelet coefficients Second low-frequency wavelet coefficients The frequency range is 0~500Hz. Second high-frequency wavelet coefficients. The frequency is 500~1000Hz.

[0070] The first high-frequency wavelet coefficient Decomposition is performed to obtain the third high-frequency wavelet coefficients. Third high-frequency wavelet coefficients The frequency is 1500~2000Hz.

[0071] Third-level decomposition: The second low-frequency wavelet coefficient The third low-frequency wavelet coefficients are obtained by decomposition. and the fourth high-frequency wavelet coefficients The third low-frequency wavelet coefficient The frequency range is 0~250Hz. Fourth high-frequency wavelet coefficients. The frequency is 250~500Hz.

[0072] The second high-frequency wavelet coefficient Decomposition is performed to obtain the fifth high-frequency wavelet coefficients. Fifth, high-frequency wavelet coefficients The frequency is 750~1000Hz.

[0073] The third high-frequency wavelet coefficient Decomposition is performed to obtain the sixth high-frequency wavelet coefficients. The sixth high-frequency wavelet coefficient The frequency is 1750~2000Hz.

[0074] Fourth level decomposition: The third low-frequency wavelet coefficient The fourth low-frequency wavelet coefficients are obtained by decomposition. and the seventh high-frequency wavelet coefficients Fourth low-frequency wavelet coefficients The frequency range is 0~125Hz. The seventh high-frequency wavelet coefficients... The frequency is 125~250Hz.

[0075] The fourth high-frequency wavelet coefficient The eighth high-frequency wavelet coefficients were obtained through decomposition. The eighth high-frequency wavelet coefficient The frequency is 375~500Hz.

[0076] The fifth high-frequency wavelet coefficient The ninth high-frequency wavelet coefficients were obtained through decomposition. The ninth high-frequency wavelet coefficient The frequency is 875~1000Hz.

[0077] The sixth high-frequency wavelet coefficient Decomposition yields the tenth high-frequency wavelet coefficients The tenth high-frequency wavelet coefficient The frequency is 1875~2000Hz.

[0078] Fifth level of decomposition: The fourth low-frequency wavelet coefficient The fifth low-frequency wavelet coefficients were obtained through decomposition. And the eleventh high-frequency wavelet coefficients Fifth low-frequency wavelet coefficients The frequency range is 0~63Hz. Eleventh high-frequency wavelet coefficients. The frequency is 63~125Hz.

[0079] The seventh high-frequency wavelet coefficient The twelfth high-frequency wavelet coefficients were obtained through decomposition. The twelfth high-frequency wavelet coefficient The frequency is 188~250Hz.

[0080] The eighth high-frequency wavelet coefficient The thirteenth high-frequency wavelet coefficients were obtained through decomposition. Thirteenth High-Frequency Wavelet Coefficients The frequency is 438~500Hz. The ninth high-frequency wavelet coefficient The fourteenth high-frequency wavelet coefficients were obtained through decomposition. Fourteenth High-Frequency Wavelet Coefficients The frequency is 938~1000Hz.

[0081] The tenth high-frequency wavelet coefficients The fifteenth high-frequency wavelet coefficients were obtained through decomposition. The fifteenth high-frequency wavelet coefficient The frequency is 1938~2000Hz.

[0082] At this point, the frequency range is small enough that further decomposition is not necessary. The frequency ranges corresponding to each decomposition level are shown in Table 1.

[0083] In the above steps, through multi-level decomposition, this invention can decompose the signal into components at different frequencies and scales, thereby better capturing the diversity and complexity of the signal. This helps to enhance the feature extraction capability of the signal, enabling more accurate identification of key features in the signal. Further subdivision of the high-frequency component is to capture more local information, enabling more precise analysis of the signal's fine structure. Although the low-frequency wavelet coefficient range "0~1000Hz" obtained from the first-level decomposition can cover the high-frequency wavelet coefficient range "125~250 Hz" obtained from the fourth-level decomposition, this invention does not select the high-frequency wavelet coefficients from the first level, but only the results obtained from the fifth level. Therefore, there is no contradictory selection between the aforementioned low-frequency and high-frequency wavelet coefficients.

[0084] The advantage of this invention in performing multi-level decomposition of high-frequency wavelet coefficients is that: (1) Achieve more accurate frequency positioning: A smaller frequency range means that the frequency components in the signal can be located more accurately. This is very useful for analyzing local features and rapid changes in the signal, as it can provide high resolution for frequency changes.

[0085] (2) It has better time-frequency locality: its smaller frequency range makes it more suitable for analyzing the instantaneous frequency of a signal, thus better capturing the instantaneous characteristics of the signal.

[0086] (3) Achieve better detail capture: Wavelet basis functions with a small frequency range are more sensitive to the details of the signal. This invention performs well in capturing the fine structure and features in the signal, especially in the case of low noise.

[0087] (4) Achieve better noise separation: A smaller frequency range helps to separate signals and noise because noise is usually more noticeable in the higher frequency range.

[0088] (5) Enable multi-scale analysis: Increasing the number of decomposition layers enables multi-scale analysis, allowing the details of the signal to be viewed at different scales.

[0089] Table 1. Frequency ranges of low-frequency and high-frequency wavelet coefficients for each layer. S13: Sort the obtained high-frequency wavelet coefficients according to their frequency range to represent the original signal of lung sounds.

[0090] Take the high-frequency wavelet coefficients of the 5th level decomposition Arranged from low to high frequency range. At this point, the raw signal of the lung sound digital biomarker can be represented by five high-frequency wavelet coefficients: ].

[0091] Preferably, during the further decomposition of the high-frequency and low-frequency wavelet coefficients obtained from each level of decomposition, the low-frequency wavelet coefficients obtained from the high-frequency wavelet coefficient decomposition are discarded. The new high-frequency and low-frequency wavelet coefficients obtained from the low-frequency wavelet coefficient decomposition are retained. Since the low-frequency portion has a relatively smooth characteristic and contains almost no respiratory details, the low-frequency wavelet coefficients are discarded. The high-frequency wavelet coefficients that provide more respiratory details are further extracted. The more levels of decomposition, the more high-frequency wavelet coefficients related to respiratory details are extracted.

[0092] S2: Select the high-frequency wavelet coefficients after the last layer of decomposition and extract the energy value.

[0093] Different types of lung sounds have different acoustic characteristics, and their energy distributions should differ across the entire frequency range. Wavelet coefficients are precisely the representation of a time-domain signal within a specific frequency band; therefore, the energy of the last layer of wavelet coefficients can be calculated and used as a characteristic of that frequency band. Let... To be related to the wavelet coefficients of the i-th layer The corresponding energy, n is dimensionality for The j-th element. At this point, the original signal of the lung sound digital biomarker can be represented by a feature vector consisting of 5 energy values: ].

[0094] S3: Construct a feature vector related to the respiratory cycle based on energy values.

[0095] One cycle of data for the digital biomarker signal of lung sounds consists of two parts: expiration and inspiration.

[0096] After decomposing the digital biomarker signals of lung sounds during expiration layer by layer, the energy of the high-frequency wavelet coefficients obtained from the last layer of decomposition is calculated, and a first feature vector is constructed. After decomposing the digital biomarker signals of lung sounds during inspiration layer by layer, the energy of the high-frequency wavelet coefficients obtained from the last layer of decomposition is calculated, and a second feature vector is constructed. The feature vectors related to the respiratory cycle, formed by combining the first and second feature vectors, are then normalized to obtain the total energy of the feature vectors. Only feature vectors extracted from the entire respiratory process of the respiratory cycle can comprehensively reflect the characteristics of lung sounds. Normalizing the feature vectors can eliminate differences in energy levels between signals.

[0097] Specifically, the two feature vectors for exhalation and inspiration are combined into one. That is, a feature vector representing the energy distribution of the signal over one cycle is formed. The original signal of the lung sound digital biomarker signal is transformed into a 10-dimensional vector, i.e.: ].

[0098] To eliminate differences in energy levels between signals, normalization is required. Let the total energy of the eigenvectors be... ,but: Normalize each energy level, that is: S4: Dimensionality reduction of the feature vector is performed to obtain lung sound features.

[0099] The steps for dimensionality reduction of feature vectors may include: reducing the dimensionality of the normalized feature vectors using Principal Component Analysis (PCA) algorithm. Dimensionality reduction helps reduce the computational load of subsequent processing without compromising the accuracy of respiratory feature extraction.

[0100] Principal Component Analysis (PCA) is a statistical method that uses orthogonal transformations to convert a set of potentially correlated variables into a set of linearly uncorrelated variables; this transformed set of variables is called the principal components.

[0101] The advantage of the dimensionality reduction process in this invention is that: (1) It can reduce data dimensionality: PCA can reduce high-dimensional data to lower dimensions, thereby reducing redundant information in the data. This helps to simplify the dataset, making it easier to process and understand.

[0102] (2) Ability to remove correlation: The goal of PCA is to find the most important directions (principal components) in the data, which are the directions with the largest variance in the data. By selecting these principal components, PCA helps to reduce the correlation between data features. This is particularly useful when dealing with highly correlated features, as correlated features can lead to overfitting and degrade model performance.

[0103] (3) Noise filtering: PCA helps remove noise from data because it focuses on the principal components with the largest variance, while noise usually has a smaller variance. By retaining the first few principal components, noisy components in the data can be removed.

[0104] (4) Reduce computational complexity: In high-dimensional datasets, computational and storage overhead can be significant. By reducing the data dimensionality, PCA can significantly reduce computational complexity, thereby accelerating the data analysis and modeling process.

[0105] S5: Perform feature selection.

[0106] Data preparation: First, the feature vector obtained by the improved wavelet transform of the lung sound signal is used as input data.

[0107] Building a Random Forest: Construct a random forest classifier using the training dataset. A random forest is an ensemble learning method consisting of multiple decision trees. Each decision tree is trained based on a different subset of features.

[0108] Feature importance assessment: In the constructed random forest model, the built-in feature importance assessment method is used to determine the relative importance of each feature in the model.

[0109] The principle of feature importance assessment is to determine the relative importance of each feature in the model by analyzing how the machine learning model uses different features.

[0110] The steps for feature importance assessment are as follows: S21: Training the Model: First, a random forest model is trained using the fitted data. S22: Record baseline performance: Before beginning feature importance evaluation, record the model's baseline performance metrics, such as accuracy and mean squared error (MSE). This will serve as a reference for evaluating the impact of feature selection on model performance.

[0111] S23: Shuffle Features: Shuffle the value order of features, that is, randomly shuffle the value order of features without changing the values ​​of other features and the target variable.

[0112] S24: Measure Performance Degradation: After shuffling the features, rerun the model and measure the degree of performance degradation. Typically, performance degradation can be measured using changes in model performance metrics such as accuracy and MSE. A significant performance degradation indicates that the feature is important in the model.

[0113] S25: Repeat steps S23 and S24: Perform steps S23 and S24 sequentially for each feature to measure the impact of each feature on model performance. This will generate a performance degradation metric for each feature.

[0114] S26: Feature Importance Ranking: Rank features according to the degree of performance degradation. Generally, features with greater performance degradation are considered more important. You can arrange features in descending order of performance degradation, from the most important features to the least important features.

[0115] Feature selection: Based on the results of the feature importance assessment, select the top 20% of features.

[0116] Model evaluation: The random forest model is retrained using the selected feature vectors, and cross-validation is used to evaluate the model's performance.

[0117] Specifically, the steps of cross-validation include: S31: Dataset splitting: Divide the dataset into k approximately equal subsets, each subset being called a fold.

[0118] S32: Model Training and Testing: For each fold, use it as the test set in sequence, and the other k-1 folds as the training set. This means that k training and testing cycles need to be performed.

[0119] S33: Model Training: In each training iteration, the model is trained using k-1 folds of data. This is the process of model learning.

[0120] S34: Model Testing: In each test, the model's performance is evaluated using the remaining folded data. A mean squared error (MSE) metric is used to measure the model's performance on the test set.

[0121] S35: Performance Recording: Record the performance metrics for each test. Typically, the average performance metric and standard deviation are calculated to comprehensively evaluate the model's performance.

[0122] S36: Repeat steps: Repeat steps S32 to S35 until each fold has been used as a test set.

[0123] S37: Performance Summary: Summarize the performance metrics from each test, calculate the average performance and standard deviation, and obtain the final performance evaluation of the model.

[0124] S38: Model Selection and Tuning: Based on the performance evaluation results, select the best model or adjust the model's hyperparameters to improve the model's performance.

[0125] Iterative optimization: If the model performance is not ideal, the threshold of feature importance can be adjusted to optimize the model and obtain the final feature vector.

[0126] Through feature selection, this invention can establish a standardized model to accurately analyze digital biomarker signals of lung sounds and obtain accurate evaluation results.

[0127] Example 2 This embodiment is a further improvement on embodiment 1, and repeated content will not be described again.

[0128] The present invention provides a signal processing system for digital biomarkers of lung sounds from a second aspect, comprising at least one processor. The processor is configured to: decompose the digital biomarker signal of lung sounds into at least one layer based on a wavelet transform method to obtain at least one reconstructed signal; select the high-frequency wavelet coefficients after the last layer of decomposition and extract the energy value; construct a feature vector related to the respiratory cycle based on the energy value; and perform dimensionality reduction processing on the feature vector to obtain lung sound features.

[0129] The processor is capable of running the coding program for the lung sound digital biomarker signal processing method of the present invention. That is, the processor is a server, dedicated integrated chip and / or server cluster, cloud server, etc., capable of executing the method of the present invention.

[0130] The system of this invention involves minimal hardware and can be flexibly deployed at the data acquisition site, or even remotely. The lung sound features extracted by the processor of this invention have more dimensional data and represent richer information.

[0131] According to a preferred embodiment, the processor is further configured to: further decompose the high-frequency wavelet coefficients and low-frequency wavelet coefficients obtained from each decomposition layer to obtain a plurality of high-frequency wavelet coefficients; and sort the plurality of high-frequency wavelet coefficients according to their frequency range to represent the original signal of lung sounds. The reconstructed original signal obtained by the system of the present invention can repeatedly extract high-frequency wavelet coefficients that reflect respiratory details, thus providing richer information about respiration.

[0132] The specific decomposition steps for each layer will not be elaborated here.

[0133] The system of the present invention may further include a preprocessing module for preprocessing the digital biomarker signals of lung sounds.

[0134] According to a preferred embodiment, the system further includes a plurality of lung sound acquisition components 1 connected to the processor via wired or wireless means. The plurality of lung sound acquisition components 1 acquire lung sound digital biomarker signals during at least one respiratory cycle. By configuring the lung sound acquisition components, it is beneficial to acquire lung digital biomarker signals in real time.

[0135] Specifically, the lung sound acquisition component 1 acquires digital lung sound signals from several individual patients during normal sitting and walking breathing conditions within a predetermined time period.

[0136] According to a preferred embodiment, when the patient wears the lung sound acquisition component 1 during daily activities, the lung sound acquisition component 1 acquires lung sounds generated by airflow friction in the patient's lungs over a continuous time period and sends the generated lung sounds to a preprocessing module for storage as a digital lung sound signal. For example... Figure 2 and Figure 3As shown, the lung sound acquisition component 1 includes a first sound sensor 11, a second sound sensor 12, a third sound sensor 13, and a fourth sound sensor 14 disposed at various locations on the patient's body for acquiring lung sounds. This invention specifically lists four arrangements of the sound sensors, but this does not mean that the invention does not include a fifth or sixth sound sensor disposed at other locations on the patient's body.

[0137] Preferably, such as Figure 2 As shown, a first sound sensor 11 is attached to the cricoid cartilage along the anterior midline of the chest to the sternal angle to collect tracheal and lung sounds. The cricoid cartilage represents the starting point of the patient's cervical trachea, i.e., the suprasternal notch. The sternal angle represents the ending point or bifurcation of the patient's trachea. The first sound sensor 11 includes several sound sensing elements removably attached to the tracheal cartilage and / or sternum of the patient's neck.

[0138] Preferably, such as Figure 2 As shown, a second sound sensor 12 is attached to the left side of the patient, at the beginning of the second rib connected to the sternal angle, to collect lung sounds from the left bronchus. The left side of the second rib connected to the sternal angle represents the beginning of the patient's left main bronchus, i.e., to the left of the patient's manubrium. The second sound sensor 12 includes several sound sensing elements removably attached to the second to sixth ribs along the rib extension direction. The removable attachment method is used because rib positioning cannot provide completely accurate lung sound localization. Therefore, medical personnel can collect sounds from the tracheal branches by changing the position of the second sound sensor 12.

[0139] Preferably, such as Figure 2 As shown, a third sound sensor 13 is attached to the beginning of the second rib on the right side, where the patient's sternal angle connects to the sternal angle, to collect lung sounds from the right bronchus. The second rib on the right side, where the patient's sternal angle connects to the sternal angle, represents the beginning of the patient's right main bronchus, i.e., the right side of the patient's manubrium. The third sound sensor 13 includes several sound sensing elements removably attached to the second to sixth ribs along the rib extension direction.

[0140] Preferably, such as Figure 3 As shown, a fourth sound sensor 14 is disposed in the region between the two scapulae on the patient's back to assist in the acquisition of sounds from the patient's trachea and bronchi. The aforementioned region between the two scapulae on the patient's back refers to the area along the scapular lines from the inferior angles of the two scapulae on both sides. The fourth sound sensor 14 includes several sound sensing elements removably disposed on the thoracic vertebrae and the second to sixth ribs in the region between the scapulae.

[0141] This invention enables the acquisition of lung sounds within the trachea and bronchi from outside the human body by placing sound sensing elements at locations of the trachea and related tracheal cartilages, ribs, thoracic vertebrae, and other human bones.

[0142] Preferably, each of the aforementioned sound sensors can be arranged in a strip shape. The sound sensor includes an adhesive patch that can be attached to the patient's skin and several point-like sound sensing elements capable of collecting and recording localized sound. The point-like sound sensing elements are arranged along the extension direction of the trachea to which the sound sensor is located. Preferably, the adhesive patch is provided with several points for installing the point-like sound sensing elements. Medical personnel can specifically set the position of the sound sensor according to the specific layout of the trachea in the body as determined by X-ray images. Preferably, medical personnel can customize the number of point-like sound sensing elements installed based on the different lengths, diameters, and branching states of the left and right main bronchuses.

[0143] For example, based on the distribution of the patient's trachea as shown on X-ray, such as Figure 2 As shown, the second sound sensor 12 is equipped with four point-like sound sensing elements: a first point-like sound sensing element 2A, a second point-like sound sensing element 2B, a third point-like sound sensing element 2C, and a fourth point-like sound sensing element 2D, to collect sound from the beginning of the left main bronchus, the bifurcation of the left main bronchus, the left upper lobe bronchus, and the left lower lobe bronchus, respectively. The third sound sensor 13 is equipped with five point-like sound sensing elements: a fifth point-like sound sensing element 3A, a sixth point-like sound sensing element 3B, a seventh point-like sound sensing element 3C, an eighth point-like sound sensing element 3D, and a ninth point-like sound sensing element 3E, to collect sound from the beginning of the right main bronchus, the bifurcation of the right main bronchus, the right upper lobe bronchus, the right middle lobe bronchus, and the right lower lobe bronchus, respectively.

[0144] For example, such as Figure 2 As shown, due to the different branches of the trachea on the left and right sides and / or the presence of a heart chamber on the left side of the patient, the second sound sensor 12 and the third sound sensor 13 are asymmetrically arranged.

[0145] Medical personnel arranged several sound sensing elements in the first sound sensor 11 according to the segmentation of the tracheal cartilage rings. They also arranged the second sound sensor 12 and the third sound sensor 13 with reference to the extension length, angle, and branching characteristics of the two main bronchi. Figure 3 As shown, medical personnel primarily deploy the fourth sound sensor 14 on the ribs and thoracic vertebrae between the shoulder blades on the patient's back, based on the placement of the first sound sensor 11, the second sound sensor 12, and the third sound sensor 13. The fourth sound sensor 14 can also supplement the collection of lung sounds in areas of the patient's anterior chest and neck where no lung sounds have been collected.

[0146] Preferably, the lung sound digital signal is a composite sound signal composed of sound signals collected by various point-like sound sensing elements and including the numbering information of each point-like sound sensing element.

[0147] It should be noted that the specific embodiments described above are exemplary. Those skilled in the art can devise various solutions inspired by the disclosure of this invention, and these solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents. This specification contains multiple inventive concepts; terms such as "preferredly," "according to a preferred embodiment," or "optionally" indicate that the corresponding paragraph discloses an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. A method for processing digital biomarker signals of lung sounds, characterized in that, The method includes: The system collects digital biomarker signals of lung sounds from a distributed array consisting of a tracheal sensing unit from the cricoid cartilage to the sternal angle, a left bronchial sensing unit, a right bronchial sensing unit, and a back scapular auxiliary sensing unit. A first acoustic sensor is attached to the cricoid cartilage along the anterior midline of the chest to the sternal angle to collect tracheal lung sounds. A second acoustic sensor is attached to the starting point of the left second rib connected to the sternal angle to collect lung sounds of the left bronchus. A third acoustic sensor is attached to the starting point of the right second rib connected to the sternal angle to collect lung sounds of the right bronchus. A fourth acoustic sensor is placed in the area between the scapulae on both sides of the patient's back to assist in collecting sounds from the trachea and bronchi. The wavelet transform method is used to decompose the digital biomarker signal of lung sounds into at least one layer and obtain at least one reconstructed signal. In the process of further decomposing the high-frequency wavelet coefficients and low-frequency wavelet coefficients obtained from each level of decomposition... Discard the low-frequency wavelet coefficients obtained from the high-frequency wavelet coefficient decomposition. The low-frequency wavelet coefficients obtained from the previous decomposition step are decomposed again, and the new high-frequency wavelet coefficients and new low-frequency wavelet coefficients are retained. Select the high-frequency wavelet coefficients after the last level of decomposition and extract the energy value; Based on the energy value, a feature vector related to the respiratory cycle is constructed. The feature vector is dimensionality reduced to obtain lung sound features.

2. The method for processing digital biomarker signals of lung sounds according to claim 1, characterized in that, The steps of decomposing the digital biomarker signal of lung sounds into at least one layer include: The wavelet transform method was used to decompose the digital biomarker signal of lung sounds to obtain high-frequency and low-frequency wavelet coefficients. The reconstructed signal is formed by the sum of the coefficients of the high-frequency wavelet coefficients and the low-frequency wavelet coefficients.

3. The method for processing digital biomarker signals of lung sounds according to claim 2, characterized in that, The step of decomposing the digital biomarker signal of lung sounds into at least one layer also includes: The high-frequency wavelet coefficients and low-frequency wavelet coefficients obtained from each level of decomposition are further decomposed to obtain several high-frequency wavelet coefficients. The obtained high-frequency wavelet coefficients are sorted according to their frequency range to represent the original signal of lung sounds.

4. The method for processing digital biomarker signals of lung sounds according to claim 3, characterized in that, The wavelet transform method is used to decompose the digital biomarker signal of lung sounds into at least three levels.

5. The method for processing digital biomarker signals of lung sounds according to any one of claims 1 to 4, characterized in that, The steps for constructing feature vectors related to the respiratory cycle include: After decomposing the digital biomarker signal of lung sounds during exhalation layer by layer, the energy is calculated based on the high-frequency wavelet coefficients obtained from the last layer of decomposition, and the first feature vector is constructed. After decomposing the digital biomarker signal of lung sounds during inhalation layer by layer, the energy is calculated based on the high-frequency wavelet coefficients obtained from the last layer of decomposition, and a second feature vector is constructed. The feature vectors related to the respiratory cycle, formed by combining the first and second feature vectors, are normalized to obtain the total energy of the feature vectors.

6. The method for processing digital biomarker signals of lung sounds according to claim 5, characterized in that, The steps for dimensionality reduction of the feature vector include: The normalized feature vectors are then dimensionality-reduced using the PCA algorithm.

7. A lung sound digital biomarker signal processing system based on the lung sound digital biomarker signal processing method according to any one of claims 1 to 6, characterized in that, Includes at least one processor, said processor being configured to: The wavelet transform method is used to decompose the digital biomarker signal of lung sounds into at least one layer and obtain at least one reconstructed signal. Select the high-frequency wavelet coefficients after the last level of decomposition and extract the energy value; Based on the energy value, a feature vector related to the respiratory cycle is constructed. The feature vector is dimensionality reduced to obtain lung sound features.

8. The lung sound digital biomarker signal processing system according to claim 7, characterized in that, The processor is also configured to: The high-frequency wavelet coefficients and low-frequency wavelet coefficients obtained from each level of decomposition are further decomposed to obtain several high-frequency wavelet coefficients. The obtained high-frequency wavelet coefficients are sorted according to their frequency range to represent the original signal of lung sounds.

9. The lung sound digital biomarker signal processing system according to claim 7, characterized in that, It also includes several lung sound acquisition components that are connected to the processor in a wired or wireless manner. Several of the lung sound acquisition components acquire digital biomarker signals of lung sounds during at least one respiratory cycle.

Images