Method and system for detecting pulmonary function abnormalities based on audio signals
Patent Information
- Application Number
- TW113151695
- Authority / Receiving Office
- TW · TW
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2024-12-31
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2044-12-30
AI Technical Summary
Existing lung function testing technologies face challenges in accuracy, data scarcity, model adaptability, and noise filtering, particularly for patient groups that cannot cooperate with traditional tests, and existing machine learning models struggle with insufficient training data and static performance.
A method and system using audio signals to detect abnormal lung function by combining audio features with individual patient parameters, incorporating noise filtering, and employing machine learning models with data augmentation and incremental learning to generate personalized predictions.
Improves accuracy and convenience of lung function testing, especially for non-cooperative patient groups, by using audio signals and machine learning models that adapt to new data, enhancing model performance and generalization.
Smart Images

Figure TWG2TB001920398_001 
Figure TWG2TB001920398_002 
Figure TWG2TB001920398_003
Abstract
Description
Technical Field
[0001] This invention relates to the field of lung function testing technology based on sound signals, and in particular to an application of using sound features combined with a machine learning model to predict abnormal lung function. Prior Technology
[0002] In modern medicine, detecting abnormal lung function is a crucial step in diagnosing and managing respiratory diseases. However, existing testing technologies still face numerous challenges in practical application. Traditional lung function testing methods typically require patients to come to a medical institution for testing and cooperate closely with the testing process, such as performing deep inhalation and forceful exhalation. However, these methods can easily lead to test errors or even fail to produce valid results for certain patient groups (such as the elderly, patients with severely impaired lung function, or patients who cannot accurately follow the testing instructions). Therefore, the convenience and accuracy of these methods still need improvement.
[0003] Furthermore, with the rapid development of machine learning technology, lung function testing models trained based on clinical data have become an important research topic. However, existing models often face the challenge of insufficient data during training. Collecting patients' lung function parameters and their corresponding auditory features requires a significant amount of time and resources, often resulting in a scarcity of labeled samples, especially in certain minority groups (such as patients with moderate to severe illness or rare diseases). This scarcity of samples limits the training effect of the model, thus affecting its generalization ability in practical applications.
[0004] Furthermore, existing machine learning models are typically static, meaning they are not updated once training is complete. However, as new patient data continuously emerges, these static models may fail to adapt to dynamic changes in data distribution, leading to a gradual decline in model performance. Moreover, current technologies lack effective methods for incremental learning on new data, preventing models from dynamically incorporating new data while maintaining existing knowledge to improve long-term applicability.
[0005] In summary, existing technologies still have significant shortcomings in terms of accuracy in detecting abnormal lung function, effectiveness of data augmentation, model adaptability, and noise filtering performance. An improved technology is needed to address these issues and meet the needs of more efficient, accurate, and flexible clinical applications. Summary of the Invention
[0006] The main objective of this invention is to provide a method for detecting abnormal lung function based on audio signals. This method generates personalized predicted lung function parameters by combining audio features with individual patient parameters for regularization. It utilizes real and amplified lung function parameters as training data to address the problem of insufficient training data. Furthermore, it incorporates noise filtering techniques and diverse model training strategies to achieve efficient audio feature extraction and model performance optimization, ultimately providing a non-invasive and convenient solution for lung function testing.
[0007] To achieve the aforementioned objective, this invention provides a method for detecting abnormal lung function based on audio signals, comprising the following steps: receiving raw audio signals from at least one subject, the raw audio signals including the subject's breathing and / or vocalizations; processing the raw audio signals by filtering noise to generate filtered audio signals; extracting at least one audio feature from the filtered audio signals, the at least one audio feature including spectral features and / or vocalization features; inputting the at least one audio feature into a machine learning model to generate a first predicted lung function parameter; normalizing the first predicted lung function parameter and converting it into a second predicted lung function parameter based on a preset reference benchmark, wherein the reference benchmark is provided as a basis for normalization based on the gender, age, and height of at least one subject.
[0008] The method further includes a training process, which includes: obtaining at least one patient's actual lung function parameters; using the actual lung function parameters as a labeled audio sample set for the machine learning model or as a second predicted lung function parameter for correction or verification.
[0009] The method further includes the following steps: adjusting the real lung function parameters and generating at least one amplified lung function parameter; using the amplified lung function parameter as an amplified labeled audio sample set for a machine learning model or for correcting or validating a second predicted lung function parameter.
[0010] The method further includes the following steps: calculating the difference or ratio between the labeled audio sample set and the amplified labeled audio sample set; updating the parameter weights of the machine learning model, and using the difference or ratio as an error measure to correct the accuracy of the machine learning model in predicting abnormal lung function.
[0011] The method further includes the following steps: performing a k-fold cross-validation process, dividing the labeled audio sample set and the amplified labeled audio sample set into multiple subsets, and using each subset in turn as the training set and the validation set to evaluate the prediction accuracy of the machine learning model for lung function abnormalities; and selecting the machine learning model with the optimal model configuration, based on the average accuracy or error performance measured by the cross-validation process.
[0012] The method further includes the following steps: dividing the labeled audio sample set into a training set, a validation set, and a test set using a three-part division method, while dividing the augmented labeled audio sample set into a training set and a validation set; training the machine learning model using the training sets derived from the labeled and augmented labeled audio sample sets; adjusting the hyperparameters or network structure of the machine learning model based on the prediction results derived from the validation sets derived from the labeled and augmented labeled audio sample sets; and verifying the accuracy of the machine learning model in predicting lung function abnormalities using the test set derived from the labeled audio sample set after the hyperparameters or network structure adjustment is completed.
[0013] Furthermore, the steps for adjusting the real lung function parameters also include: correcting at least one of the patient's age, height, and weight to obtain the physiological conditions of the virtual patient; calculating the amplified lung function parameters corresponding to the physiological conditions, which are derived by clinical formulas and / or physiological models; and verifying whether the amplified lung function parameters fall within the acceptable clinical range. If so, the amplified lung function parameters are included in the amplified labeled audio sample set; otherwise, the amplified lung function parameters are discarded.
[0014] Furthermore, the method further includes the following steps: receiving new real lung function parameters from at least one new patient; adding the new real lung function parameters to an incremental labeled dataset, which is used to update the labeled audio sample set; and fine-tuning the machine learning model by using the incremental labeled dataset to perform another training on the machine learning model.
[0015] This invention also provides a system for detecting abnormal lung function based on audio signals, including a receiving module and a processing module. The receiving module receives or captures raw audio signals from at least one subject, including the subject's breathing and / or vocalizations. The processing module is connected to the receiving module and performs noise filtering on the raw audio signals to generate filtered audio signals. It then extracts at least one audio feature from the filtered audio signals, inputs the at least one audio feature into a machine learning model to generate a first predicted lung function parameter, and then converts the first predicted lung function parameter into a second predicted lung function parameter based on a preset reference benchmark.
[0016] The system further includes a training module, which is a signal connection processing module, for obtaining the real lung function parameters of at least one patient, and using the real lung function parameters as a set of labeled audio samples for the machine learning model or as a means to correct or verify a second predicted lung function parameter.
[0017] The system further includes a data augmentation module, a signal connection training module, and a processing module, which are used to adjust real lung function parameters and generate at least one amplified lung function parameter. The amplified lung function parameter is used as an amplified labeled audio sample set for a machine learning model or to correct or verify a second predicted lung function parameter.
[0018] The processing module also calculates the difference or ratio between the labeled audio sample set and the amplified labeled audio sample set, and updates the parameter weights of the machine learning model. The difference or ratio is used as an error measure to correct the accuracy of the machine learning model in predicting abnormal lung function.
[0019] The processing module also performs a cross-validation process, which divides the labeled audio sample set and the amplified labeled audio sample set into multiple subsets, and uses each subset as a training set and a test set in turn to evaluate the accuracy of the machine learning model in predicting lung function abnormalities, and selects the machine learning model with the optimal model configuration. The selection is based on the average accuracy or error performance measured by the cross-validation process.
[0020] The processing module further divides the labeled audio sample set into training, validation, and test sets, while the augmented labeled audio sample set is also divided into training and validation sets. The machine learning model is trained using the training sets derived from both the labeled and augmented labeled audio sample sets. Based on the prediction results from the validation sets derived from both sets, the hyperparameters or network structure of the machine learning model are adjusted. After the hyperparameters or network structure are adjusted, the machine learning model's accuracy in predicting lung function abnormalities is verified using the test set derived from the labeled audio sample set.
[0021] In this process, when adjusting real lung function parameters, the data augmentation module corrects at least one of the patient's age, height, and weight to obtain a physiological condition of a virtual patient and calculates the corresponding amplified lung function parameters. The amplified lung function parameters are derived by a clinical formula and / or a physiological model. The amplified lung function parameters are then checked to see if they fall within an acceptable clinical range. If so, the amplified lung function parameters are included in the amplified labeled audio sample set.
[0022] The training module also receives new real lung function parameters from at least one new patient and adds these parameters to an incremental labeled dataset, which is used to update the labeled audio sample set. The machine learning model is then fine-tuned by using the incremental labeled dataset to perform another training on the machine learning module.
[0023] As mentioned above, noise filtering includes spectral subtraction, which involves selecting speech-free segments from the original audio signal as background noise, calculating the average spectrum of the background noise, and subtracting the average spectrum from the spectrum of the original audio signal using linear subtraction.
[0024] As described above, noise filtering includes Voice Activity Detection (VAD) based on a short-time energy threshold. Short-time energy is the average value of the squared amplitude of the original audio signal within a predetermined time window. The short-time energy threshold is a dynamic reference value calculated from ambient background noise and used to determine voice activity. The voice activity extraction method includes: comparing the short-time energy with the short-time energy threshold window by window; determining the start of voice activity when the short-time energy of multiple consecutive time windows exceeds the short-time energy threshold; determining the end of voice activity when the short-time energy of multiple consecutive time windows is lower than the short-time energy threshold; extracting the time segment of the voice activity as a valid voice segment and excluding background noise.
[0025] As described above, the first predicted lung function parameter, the second predicted lung function parameter, the actual lung function parameter, and the amplified lung function parameter may each include at least one of the following: peak expiratory flow (PEF), predicted peak expiratory flow (PEF BEST), Asthma Control Test (ACT) score, forced expiratory volume in 1 second (FEV1), and forced vital capacity (FVC).
[0026] Compared with prior art, the advantages that this invention can claim include:
[0027] (1) Significantly improves the accuracy of pulmonary function abnormality detection: By combining audio characteristics with individual patient parameters (such as gender, age, and height) for data normalization, the first predicted pulmonary function parameter is converted into a second predicted pulmonary function parameter after individualized (normalized) processing. This method solves the problem that existing audio-based detection technologies cannot take into account individual patient differences, greatly improving the accuracy and clinical applicability of detection, especially showing stronger generalization ability in diverse patient groups.
[0028] (2) Generating amplified pulmonary function parameters to address the scarcity of training data: Amplified pulmonary function parameters reflecting the diverse physiological conditions of patients are generated by adjusting and amplifying real pulmonary function parameters. Specific methods include adjustments based on age, height, and weight, combined with calculations using clinical formulas or physiological models to ensure the clinical rationality of the amplified data when simulating virtual patient conditions. These amplified pulmonary function parameters, as new training data (such as training sets or validation sets), can significantly improve the diversity of training data.
[0029] (3) Enhancing the long-term performance of the model: Fine-tuning techniques are employed on the existing machine learning module, allowing the model to continuously receive real lung function parameters from new patients and integrate them into the incremental labeled dataset for fine-tuning. This dynamic update mechanism ensures that the model can quickly adapt to changes in the distribution of new clinical data, further improving its accuracy in long-term applications.
[0030] (4) Diverse Model Training Strategies: Diverse model training strategies are employed, including cross-validation and a tripartite approach (training set, validation set, and test set), to effectively balance the model's learning effects on real lung function parameters and amplified lung function parameters, further improving the model's accuracy and generalization ability. These strategies ensure stable training of the model under different data conditions, with hyperparameters adjusted through the validation set. By avoiding the use of amplified lung function parameters as the test set, the risk of excessive hyperparameter tuning for the test set can be avoided, ultimately achieving efficient prediction of lung function abnormalities.
[0031] (5) Achieve efficient noise filtering and audio feature extraction: By combining spectral features and vocalization features, the accuracy and comprehensiveness of audio feature extraction are further improved. These vocalization features can effectively capture subtle changes in speech signals, especially the stability and clarity of sound, and can more accurately identify speech features related to abnormal lung function. At the same time, it enhances the model's ability to analyze speech quality and potential health problems, and achieves more accurate detection of abnormal lung function.
[0032] (6) Providing a non-invasive and convenient lung function testing solution: By utilizing sound signals to detect abnormal lung function, this invention provides a non-invasive alternative, particularly suitable for patient groups who cannot cooperate with traditional tests, such as children, the elderly, or critically ill patients. This testing method can significantly improve the convenience of the testing process while reducing the burden on medical resources.
[0033] The following detailed description, using specific embodiments and accompanying drawings, will make it easier to understand the purpose, technical content, features, and effects achieved by this invention. Simple Explanation of the Diagram
[0034] Figure 1 is a first flowchart of the method for detecting abnormal lung function based on sound signals according to an embodiment of the present invention; Figure 2 is a second flowchart of the method for detecting abnormal lung function based on sound signals according to an embodiment of the present invention; Figure 3 is a first block diagram of a system for detecting abnormal lung function based on audio signals according to an embodiment of the present invention; and Figure 4 is a second block diagram of the system for detecting abnormal lung function based on audio signals according to an embodiment of the present invention. Implementation
[0035] Embodiments of the present invention will be further explained below with reference to the accompanying drawings. Wherever possible, the same reference numerals in the drawings and description represent the same or similar components. It is understood that elements not specifically shown in the drawings or described in the description are forms known to those skilled in the art. Those skilled in the art can make various changes and modifications based on the content of this invention.
[0036] According to one embodiment, please refer to Figure 1 first. Figure 1 is a first flowchart of a method for detecting abnormal lung function based on sound signals, including steps S110 to S150.
[0037] As in step S110, a raw audio signal is received from at least one subject.
[0038] In this step, the subject's breathing or vocalizations (e.g., reading aloud) can be recorded using the microphone of a mobile device (such as a smartphone or tablet). The audio format can be WAV or other lossless compression formats to ensure the quality of subsequent processing and feature extraction. Simultaneously, the subject's personal information, such as gender, age, height, weight, and smoking status, can be obtained from any device interface.
[0039] As in step S120, the original audio signal is processed, and noise filtering is performed to generate a filtered audio signal.
[0040] In this step, the original audio is first processed as follows:
[0041] (1) Voice activity detection: Excluding long periods of no voice segments or background noise segments; and / or
[0042] (2) Spectrum subtraction: Using the speechless segments around the subject as background noise samples, the average spectrum is calculated and the corresponding spectrum is subtracted from the whole audio signal to obtain a purer speech signal.
[0043] After the above process, the "filtered audio signal" is obtained, and its signal-to-noise ratio should be significantly better than the original audio signal, which is convenient for subsequent audio feature extraction.
[0044] As in step S130, audio features are extracted from the filtered audio signals.
[0045] In this step, a Short-Time Fourier Transform (STFT) can be performed on the filtered audio signal to obtain spectral information; then, at least one of the following characteristics is calculated:
[0046] (1) Spectral characteristics: such as Mel-Frequency Cepstral Coefficients (MFCC), Delta-MFCC, spectral energy distribution, etc.;
[0047] (2) Vocal characteristics: such as vocal cord vibration changes, fundamental frequency (F0) jitter, amplitude, harmonic to noise ratio (HNR), etc.
[0048] Preliminary feature standardization (e.g., Z-score) can also be performed to ensure that the dimension of a physical quantity of audio characteristics is consistent among different test subjects.
[0049] As in step S140, the audio features are input into the machine learning model to generate the first predicted lung function parameters.
[0050] In this step, a pre-trained machine learning model can be deployed first (e.g., Deep Neural Network (DNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), or a Convolutional Neural Network (CNN) with an attention mechanism). The input layer of the machine learning model is the audio features or their vector values extracted in step S130, and the output layer is the "first predicted lung function parameters," which are floating-point numbers or vector values representing the subject's lung function indicators, such as: peak expiratory flow rate, optimal peak expiratory flow rate, asthma control test score, forced expiratory volume in one second, and maximum vital capacity.
[0051] Furthermore, since the machine learning model at this time mainly relies on the "sound signal" itself and may not necessarily include the subject's "gender, age, and height" factors, the "first predicted lung function parameter" is often a more "general" or "predicted" state.
[0052] As in step S150, the first predicted lung function parameter is normalized to generate the second predicted lung function parameter.
[0053] In this step, the "first predicted lung function parameter" is modified based on a "preset reference baseline" to obtain a "second predicted lung function parameter" that more closely reflects the individual's condition. The specific procedure is as follows:
[0054] 1. Define the reference datum:
[0055] Reference benchmarks can be derived from clinically known formulas for "peak expiratory flow rate (PEF BEST)" or "forced expiratory volume in one second (FEV1)," or can be provided from an internal database. For example:
[0056] Alternatively, it could be a formula validated by various medical research or clinical guidelines (such as the Global Initiative for Asthma (GINA) or respiratory societies around the world).
[0057] 2. Calculate the ideal lung function values corresponding to gender / age / height:
[0058] For example, if the first predictive lung function parameter is "peak expiratory flow (PEF)", then: based on the "Age, Height, Sex" input by the test subject, first calculate... .
[0059] For example: If the test subject is actually 40 years old, 170 cm tall, and male, the corresponding value in the database is "PEF_BEST(40,170,male)= 600 (L / min)".
[0060] 3. Normalization methods may include ratio methods, difference correction methods, or other linear / nonlinear mappings:
[0061] (1) Ratio method: Let the first predictive lung function parameter = P1, and the ideal baseline = B. Second predictive lung function parameter =
[0062] in It may be the average value used as a reference in a machine learning model. This approach allows the "first predicted value" to closely approximate the baseline value for that subject's age / height / gender.
[0063] (2) Difference correction method: Second predictive lung function parameter =
[0064] in Here is the weighting constant. The average baseline is the general population (such as groups of the same sex and age) or the average baseline of machine learning models (such as the average baseline in a set of labeled audio samples); this method can adjust the "first prediction" to make it closer to B.
[0065] 4. Generate "second predictive lung function parameters":
[0066] Ultimately, the predicted parameters obtained are "normalized (individualized)" for this subject, which better reflect the lung function level that should be achievable based on their age, height, and gender. Thus, if the first predicted lung function parameter calculates PEF = 500 (L / min), but the baseline based on their age / height / gender might be 600 (L / min), after normalization (using ratio methods or other correction methods), the expected PEF can be obtained... The "second predictive lung function parameter" allows the results to better reflect the individual's physical condition.
[0067] In this embodiment, the "real lung function parameters (as an annotated audio sample set)" collected in other embodiments described later can be used to continuously optimize the above "normalization formula" or update the correction weights, so that the influence of the subject's age / height / gender is more accurate.
[0068] According to another embodiment, please refer to Figure 2 first. Figure 2 is a second flowchart of the method for detecting abnormal lung function based on sound signals, including steps S210 to S280.
[0069] As in step S210, obtain the actual lung function parameters of the real patient.
[0070] In this step, the patient's actual lung function parameters can be obtained through methods such as peak expiratory flow meter (PEF measurement), spirometry (FEV1, FVC measurement), or ACT score (asthma control test) completed by the physician based on clinical assessment. Therefore, the actual lung function parameters include, but are not explicitly shown, the following: patient ID, asthma control, control level, sex, age, height, weight, smoking, peak expiratory flow (PEF), peak expiratory flow best (PEF BEST), asthma control test score (ACT), expiratory nitric oxide (FeNO), forced expiratory volume in one second (FEV1), and forced vital capacity (FVC).
[0071] Asthma control can be classified into well-controlled, partially controlled, and uncontrolled categories based on global asthma prevention initiatives or physician clinical experience.
[0072] The formula for calculating PEF BEST can be, for example, as follows:
[0073] Adult male PEF BEST = (3.8856 × Height) - (2.9508 × Age) + 43.5846
[0074] Adult female PEF BEST = (4.1028 × Height) - (1.611 × Age) - 173.5476;
[0075] For example, for patient A (male, Age=37, Height=163), the calculated PEF BEST = 568 L / min (example), and by comparing the difference with the actual PEF = 440 L / min in the table, the condition can be determined.
[0076] Furthermore, the actual lung function parameters collected from 30 patients can be shown in Table 1, for example. These actual lung function parameters from real patients serve as an important basis for subsequently generating amplified lung function parameters and training the machine learning module, as detailed in the following examples. Table 1: Real lung function parameters of 30 patients ID asthma control control level Sex* Age Height (cm) Weight (kg) Smoking PEF (L / min) PEF Best (L / min) ACT FeNO (ppb) FEV1 (L) FVC (L) 1 well controlled 1 2 45 158 62 null 440 450 twenty two 10 114% 116% 2 well controlled 1 2 36 163.5 86.2 null 550 550 24 17 N N 3 well controlled 1 2 65 160 67 null 200 385 22 12 37% 40% 4 partial controlled 2 2 65 155 55 null 300 351 18 19 88% 92% 5 well controlled 1 2 59 160.5 47.9 null 350 389 21 30 88% 100% 6 well controlled 1 2 65 156.5 55 null 300 363 22 42 112% 115% 7 well controlled 1 2 63 156 53 null 400 400 23 14 132% 131% 8 well controlled 1 2 66 152 47 null 300 345 22 21 91% 85% 9 well controlled 1 2 61 168 73 null 480 500 19 17 101% 103% 10 well controlled 1 1 70 161.5 60.9 null 560 560 22 51 113% 106% 11 partial controlled 2 2 83 150 89.5 null 250 376 21 65 63% 68% 12 well controlled 1 2 22 157 52 null 250 589 22 10 65% 64% 13 well controlled 1 1 42 165 70 null 550 640 23 N 104% 104% 14 well controlled 1 1 63 176 78 null 580 580 23 18 124% 116% 15 well controlled 1 1 73 163 67 Y 480 480 23 32 N N 16 partial controlled 2 2 36 148.5 49.3 null 330 372 17 N 85% 82% 17 well controlled 1 1 67 169 87 null 450 503 21 N 87% 82% 18 well controlled 1 2 50 168 51.5 null 350 431 23 39 N N 19 well controlled 1 2 51 155 86 null 300 364 21 5 69% 65% 20 well controlled 1 2 73 160 64 null 310 355 22 6 73% 76% 21 partial controlled 1 1 37 163 72 null 440 568 22 N 83% 85% 22 partial controlled 2 1 22 184 80 null 600 694 20 108 107% 105% 23 well controlled 1 2 35 163 54 null 120 439 21 32 89% 91% 24 uncontrolled 3 2 61 158 64 null 300 376 21 N 75% 94% 25 uncontrolled 1 1 57 175 105 null 630 680 22 N N N 26 uncontrolled 1 1 75 147.5 50.4 null 250 314 20 N N N 27 well controlled 1 2 75 170 62 null 390 502 23 33 N N 28 well controlled 1 1 76 160 76 null 270 358 23 16 N N 29 well controlled 1 1 41 155 53 null 450 450 twenty three 19 122% 119% 30 well controlled 1 1 40 159 60 null 450 460 twenty two 18 93% 86% *1 = Male; 2 = Female.
[0077] As in step S220, real lung function parameters (annotated audio sample set) are used as training data for the machine learning model.
[0078] As shown in Table 1, the actual pulmonary function parameters of each patient (ID 1 ~ 30) can be considered as an "annotated audio sample set". This sample set can be partitioned and applied to the training set, validation set, or test set, or used for training processes such as cross-validation. However, in reality, the actual pulmonary function parameters may not be as complete as those in Table 1, and some actual pulmonary function parameters may be missing. But because real clinical data is very valuable, even if there are missing actual pulmonary function parameters in the patient's clinical data, they can still be used as training data for machine learning models.
[0079] Furthermore, even in cases where some real lung function parameters are missing, the missing data can be supplemented by a trained machine learning model. The detailed process can be achieved by applying the data augmentation step in subsequent step S230.
[0080] As in step S230, adjust the actual lung function parameters and generate at least one amplified lung function parameter. Alternatively, as described above, supplement the missing actual lung function parameters in the clinical data of real patients.
[0081] In this step, each patient data point in Table 1 may only represent a certain group (age, height, PEF, etc.); if the sample size of a certain category (e.g., partial control) is too small, data of "virtual patients" can be generated by "minorly adjusting" certain parameters to enrich the training data.
[0082] 1. Examples of minor adjustment methods include:
[0083] (1) Adjustments based on age, height, and weight: For example, for ID 21 (male, Age=37, Height=163, partially controlled, PEF BEST=568, PEF=440), it can be estimated that combinations such as Age=40, Height=165.3 (21a), Age=35, Height=161.5 (21b), etc., can maintain PEF BEST=568 (or a nearby value); if Age=45, Height=169.1 (21c), the calculated PEF BEST is still approximately 568 L / min. In this process, if the calculated PEF BEST, FEV1, or other indicators fall within a reasonable range, it can be considered an acceptable amplified lung function parameter.
[0084] (2) Adjustment based on asthma control (ACT) classification: If a patient is uncontrolled but differs from the clinical boundary of partial controlled by only 1 ACT point or 10% FEV1 value, the corresponding ACT or FEV1 value can be adjusted to expand the data to be classified as partial controlled.
[0085] For example, patient ID 4 is female with partially controlled asthma and an ACT of 18; while ID 4a has an ACT of 16; ID 4b has an ACT of 17; and ID 4c has an ACT of 19. Data from three additional virtual patients (4a, 4b, 4c) were added. Patient ID 16 is also female with partially controlled asthma and an ACT of 17; while ID 4a has an ACT of 16; ID 4b has an ACT of 18; and ID 4c has an ACT of 19. Similarly, ACT was adjusted and data from three additional virtual patients (16a, 16b, 16c) were added. The adjusted expanded pulmonary function parameters are shown in Table 2 below. Table 2: One of the augmented lung function parameters of virtual patients ID asthma control control level Sex Age Height Weight Smoking PEF PEF Best ACT FeNO FEV1 FVC 4 Partially controlled 2 2 65 155 55 null 300 351 18 19 88% 92% 4a Partially controlled 2 2 65 155 55 null 300 351 16 19 88% 92% 4b Partially controlled 2 2 65 155 55 null 300 351 17 19 88% 92% 4c partial controlled 2 2 65 155 55 null 300 351 19 19 88% 92% 16 partial controlled 2 2 36 148.5 49.3 null 330 372 17 N 85% 82% 16a partial controlled 2 2 36 148.5 49.3 null 330 372 16 N 85% 82% 16b partial controlled 2 2 36 148.5 49.3 null 330 372 18 N 85% 82% 16c Partially controlled 2 2 36 148.5 49.3 null 330 372 19 N 85% 82%
[0086] (3) Adjustment based on PEF value: This refers to the actual exhalation data of the patient, while PEF BEST is the optimal estimate calculated using a formula. The optimal PEF BEST estimate can be obtained by converting the gender, age, and height of adult (male and female) patients. Similarly, the optimal PEF BEST estimate can be obtained by converting the gender, age, height, and weight of children (male and female) patients. Adjusting the optimal PEF BEST estimate for adults while maintaining the same age and height, and adjusting the optimal PEF BEST estimate for children while maintaining the same age, height, and weight. For example: an ID 21 patient is male, with partially controlled asthma, aged 37, and with a height of 163 cm, the optimal PEF BEST estimate is 568. Using the following formula, different height combinations can be calculated by randomly assigning different ages (40, 35, 45, 50, 30, etc.). Let PB represent the optimal PEF BEST estimate for adult males, let H represent height, and let A represent age. The formula is: Adult male PB = (3.8856 * H) - (2.9508 * A) + 43.5846 For an adult male, H = [PB + (2.9508 * A) - 43.5846] / 3.8856
[0087] Therefore, data from five virtual patients (ID21a, 21b, 21c, 21d, and 21e) can be amplified: ID21a, Age 40, Height 165.3; ID21b, Age 35, Height 161.5; ID21c, Age 45, Height 169.1; ID21d, Age 50, Height 172.9; and ID21e, Age 30, Height 157.7. These data are shown in Table 3 below, and their PEF BEST estimates are all 568, the same as ID21. Table 3: Augmented Lung Function Parameters of Virtual Patients (Part Two) ID asthma control control level Sex Age Height Weight Smoking PEF PEF Best ACT FeNO FEV1 FVC twenty one Partially controlled 2 1 37 163 72 null 440 568 twenty two N 83% 85% 21a Partially controlled 2 1 40 165.3 72 null 440 568 22 N 83% 85% 21b partial controlled 2 1 35 161.5 72 null 440 568 22 N 83% 85% 21c partial controlled 2 1 45 169.1 72 null 440 568 22 N 83% 85% 21d partial controlled 2 1 50 172.9 72 null 440 568 22 N 83% 85% 21e partial controlled 2 1 30 157.7 72 null 440 568 twenty two N 83% 85%
[0088] 3. The clinical rationale for minor adjustments needs to be verified.
[0089] (1) Use the PEF BEST formula or ACT score range (such as ID16~19) to judge: If an amplified data point has Age=50 and Height=172.9, calculate PEF BEST=568 ± 5 L / min, and correct PEF and FEV1 accordingly. Finally, if the difference is no more than ±10% of the original patient's clinical data, it is acceptable; otherwise, discard the virtual data point.
[0090] For example, ID=24 (female, Age=61, Height=158, uncontrolled, PEF BEST=376, PEF=300). Combinations such as Age=55, 159.5 cm (24b), Age=51, 154 cm (24e) can be amplified. As long as the calculated PEF BEST ≈ 376, these can be considered acceptable amplification parameters for lung function.
[0091] Specifically, for patient ID 24, different heights of 155.5, 159.5, 150.4, 161, and 154 cm can be randomly assigned according to the following formula, and different age combinations of virtual patients can be calculated. Adult female PB = (4.1028 * H) - (1.611 * A) - 173.5476 Adult female A = [(4.1028 * H) - 173.5476 - PB] / 1.611
[0092] Therefore, data for 5 virtual patients (24a, 24b, 24c, 24d, 24e) were generated and are shown in Table 4 below. All of them maintained the original Asthma control (e.g., partial / uncontrolled), but their ages and heights were different, and they still met the reasonable range of the corresponding PEF BEST. Table 4: Augmented Lung Function Parameters of Virtual Patients (Part 3) ID asthma control control level Sex Age Height Weight Smoking PEF PEF BEST ACT FeNO FEV1 FVC 24 uncontrolled 3 2 61 158 64 null 300 376 21 N 75% 94% 24a uncontrolled 3 2 55 155.5 64 null 300 376 21 N 75% 94% 24b uncontrolled 3 2 65 159.5 64 null 300 376 21 N 75% 94% 24c uncontrolled 3 2 42 150.4 64 null 300 376 21 N 75% 94% 24d uncontrolled 3 2 69 161 64 null 300 376 21 N 75% 94% 24e uncontrolled 3 2 51 154 64 null 300 376 21 N 75% 94%
[0093] Furthermore, when selecting data to be augmented, an indexed array can be used. First, an indexed array contains 15 elements, corresponding to fields such as ID, Asthma control, Sex, PEF, PEF BEST, ACT, and FEV1. Fixed fields (such as ID and Sex) always have a value of 1; while adjustable parameters such as PEF, PEF BEST, ACT, FEV1, and FVC can have values of 0 (off) or 1 (on).
[0094] Based on clinical experience, certain parameters (such as FeNO and Smoking) may be set to 0 if clinical data collection is relatively scarce or contains a lot of noise, to avoid affecting enhancement quality; conversely, if PEF or PEF BEST is sufficient to support the amplification logic, it is set to 1. In the process of "amplifying lung function parameters," Age and Height are not simply adjusted randomly, but rather the corresponding formulas or interval constraints are applied according to the parameters whose index array shows 1. For example:
[0095] 1. PEF BEST Formula: If both PEF and PEF BEST are 1 (enabled), after adjusting Age and Height, you can use this formula to verify whether the new combination still yields a reasonable PEF BEST; at the same time, check what percentage of the estimated value the actual PEF represents.
[0096] 2. ACT or FEV1 / FVC: If ACT=1 or FEV1 and FVC=1, then during the transition from "uncontrolled to partially controlled", check whether ACT is only in the range of 16~19 and whether the ratio of FEV1 to FVC falls within the clinical range of the corresponding grade.
[0097] An example of an integrated index array process could be:
[0098] 1. Read the index array combination: Assume Array = [ …, PEF=1, PEF BEST=1, ACT=1, FEV1=0, FVC=0, … ].
[0099] 2. Basis for Amplification: In the above index array combination, when generating amplified pulmonary function parameters for virtual patients, both the "PEF BEST formula" and the "ACT score interval" are checked to ensure they conform to the specified ranges; however, FEV1 or FVC is not checked. If, after adjusting Age and Height, the result after substituting into the PEF BEST formula exceeds the ±10% error range, the amplified pulmonary function parameter is discarded. If ACT is also enabled, it is necessary to check whether the ACT score is clinically defined.
[0100] 3. Generate reasonable amplified lung function parameters: If the examination shows that all parameters meet the clinical criteria, then record the "(Age', Height', PEF', PEF_BEST', ACT')" as a new virtual patient data, and combine it with the original other values (such as fixed fields such as Sex and Smoking) to form a complete amplified labeled audio sample set.
[0101] 4. Incorporate into model training (applied to subsequent step S240): During training, feature / target fields are read only for the parameters of Array=1. If PEF=0 or FEV1=0, no input is made and the error for this parameter is not compared.
[0102] In summary, since each change in the array combination alters the values of amplified lung function parameters, the machine learning module can find the optimal parameter combination through repeated experiments. This ensures that the generated virtual patients are clinically meaningful and effectively increase the amount of data from rare gradations (such as uncontrolled) or specific age groups. However, if subsequent evaluations show that enabling too many parameters leads to excessive constraints and insufficient amplification data, some parameters can be disabled to loosen the constraints. Conversely, if enabling too few parameters causes the amplified lung function parameters to deviate from the actual clinical situation, the parameter activation rate can be increased. Therefore, the role of the index array in data augmentation is not to directly control "how to change Age, Height, and Weight," but rather to determine which clinical formulas or grading rules (such as PEF BEST, ACT) to refer to during the amplification validation phase and which features to select during machine learning model training. In this way, targeted augmentation can be performed for specific needs (such as supplementing the number of samples in a certain gradation) while ensuring that the final generated virtual patient data is clinically reasonable.
[0103] As in step S240, lung function parameters (amplified labeled audio sample set) are used as training data for the machine learning model.
[0104] In this step, real lung function parameters (from the labeled audio sample set) and newly generated amplified lung function parameters (from the amplified labeled audio sample set) are combined and fed into a machine learning model (such as DNN, CNN, LSTM, etc.) to improve the problems of "scarcity of training data" or "insufficient specific classifications". For example, if the total number of uncontrolled cases is too small, the amplified lung function parameters of virtual patients with IDs 21a~21e, IDs 16a~16c, etc., can make up for the lack of specific real lung function parameters.
[0105] As in step S250, train the machine learning model.
[0106] In this step, during the machine learning model training phase, an "annotated audio sample set" (real patient data) and an "amplified annotated audio sample set" (virtual patient data) can be integrated. For each sample, since its corresponding correct lung function parameters (e.g., actual measured lung function parameters) already exist, the machine learning model will predict for each sample and calculate the difference between the predicted output and the target value (i.e., the loss value, such as mean squared error, MSE). For these loss values, the gradient backpropagation algorithm can be applied to dynamically adjust the weights and biases of the machine learning model. This process can further optimize the machine learning model, enabling it to learn various features more accurately (such as audio features, PEF, FEV1, ACT grade, etc.) to improve the accuracy of predicting lung function parameters. If an objective evaluation of model performance is required, cross-validation or a three-part method can be used during the training phase to ensure that the machine learning model has sufficient accuracy and generalization ability when judging the original audio signals of unknown subjects.
[0107] In addition, various validation strategies (such as cross-validation or the three-part method) can be used to objectively evaluate the enhanced machine learning model.
[0108] Furthermore, in the three-part training strategy, real lung function parameters can be used simultaneously in the training, validation, and test sets. This is because real lung function parameters are obtained through clinical measurements or physician diagnoses, and are more representative of the diverse situations of real patients (e.g., different ages, genders, disease grades, etc.), thus serving as a benchmark for evaluating the model's performance in "real clinical scenarios." Therefore, allocating some real lung function parameters to the test set can accurately reflect the model's performance on unseen "actual clinical distributions." However, it is recommended that augmented lung function parameters be used only in the training and validation sets, and not in the test set. Augmented lung function parameters generate "virtual patient" data through clinical formulas (such as PEF BEST) or minor adjustments (Age, Height, etc.). While this can supplement scarce samples during the training phase, it still contains a "synthetic" component. If such synthetic data is allocated to the test set, it may cause the test results to be "overestimated" or "biased," failing to truly represent the model's predictive performance in real clinical scenarios. Testing has shown that amplifying lung function parameters in the training set can significantly compensate for the lack of samples in certain classifications (such as partially controlled / uncontrolled) or specific age groups, allowing the model to learn a wider range of features from diverse samples. Furthermore, in the validation set, the model's behavior on a mixed distribution of "synthetic + real" data can be monitored in real time. Some of the amplified data can be placed in the validation set to observe the model's fit to rare classes, thereby continuously adjusting weights and hyperparameters.
[0109] Therefore, it should be ensured that the test set only represents the real clinical distribution. Thus, it is recommended that the test set consist entirely or mostly of real lung function parameters from real clinical patients to maintain the accuracy of the machine learning module's judgment.
[0110] As in step S260, the trained machine learning model is used to assess lung function based on the subject's original audio signals.
[0111] In this step, steps S110-S150 of Figure 1 are actually executed. Using the initial training of the machine learning model completed in the previous steps (S210-S240), the raw audio signals (after preprocessing and feature extraction) of the "unknown subject" can now be directly fed into the machine learning model. The machine learning model then outputs the corresponding "predicted lung function parameters," such as PEF, FEV1, ACT, or other clinical indicators, based on the learned weights.
[0112] Furthermore, the machine learning model can be referenced from previous normalization benchmarks (such as PEF BEST calculated based on age and height) or asthma control grading information to indicate results such as "green light (80%~100%)", "yellow light (60%~80%)", or "red light (<60%)" to the test subject. Please refer to Table 5 below for the specific light numbers corresponding to PEF BEST. If the results indicate that there may be moderate or severe abnormalities, the test subject can be advised to go to a medical institution for further examination or treatment. Table 5: Correspondence between Peak Expiratory Flow Rate (PEF) Severity Best estimate (PEF BEST) Green light 80%~100% (PEF) Yellow light 60%~80% (PEF) red light < 60% (PEF) aldult (male) 524 419 ~ 524 314 ~ 419 < 314 aldult (female) -174 -139 ~ -174 -104 ~ -139 |PEF| < |-104| child (male) -131 -104 ~ -131 -78 ~ -104 |PEF| < |-78| child (female) -99 -79 ~ -99 -59 ~ -79 |PEF| < |-59|
[0113] As in step S270, check whether new real lung function parameters have been obtained.
[0114] In this step, new true lung function parameters can be set for regular checkups (e.g., weekly or monthly) or for acquisition after a specific event (e.g., a patient's return visit for a formal lung function test). For example, for patient ID 30, new true lung function measurements (PEF, FEV1, ACT, etc.) can be obtained after assessment by a physician using a spirometer, blood oxygen saturation analyzer, or ACT. Alternatively, entirely new patient information can be provided, such as patient ID 31. If no new true lung function parameters are available, the process returns to step S260, maintaining the currently trained machine learning model and continuously providing judgment results for unknown subject audio.
[0115] As in step S280, the machine learning model is fine-tuned using new real lung function parameters.
[0116] In this step, if new real patient data is confirmed in step S270, it can be integrated with the corresponding new real lung function parameters to form an "incremental labeled dataset". Alternatively, based on the logic of previous steps (S230 ~ S240), age or height can be adjusted using clinical formulas (PEF BEST, ACT classification, etc.) to generate new "amplified lung function parameters" to supplement the incremental labeled dataset.
[0117] Generally, when new training data (i.e., new real lung function parameters) becomes available, machine learning models may undergo either full retraining or naive incremental training using only the new data. However, each approach has significant drawbacks. Full retraining drastically increases training time and resource costs; it requires fitting all parameters from scratch; and in medical settings, it may not meet the real-time requirements of outpatient settings. Training with only new data, on the other hand, can lead to catastrophic forgetting: while the model overfits to the new data, it destroys the knowledge learned from the old data; ultimately, the model's predictive ability for the old distribution decreases.
[0118] Therefore, after testing, this embodiment adopts a retraining mechanism of partial layer freezing + fine-tuning. This retains the weights of most layers of the old model, performing open training only on the last few layers or specified subnetworks, while simultaneously allowing the model to see both "partial old data" and "new data" for weight updates. Although in this training method, if the feature distribution of the new and old data differs greatly, fine-tuning only the last few layers may still not be sufficient to learn the features of the new samples, requiring a trade-off between "number of frozen layers" and "learnable parameters," in medical applications, the differences in collected data features are not significant (because the features are concentrated in specific clinical values), and the amount of new data added each time is usually small. If a full retraining is performed with only a small amount of new data each time, the time and resource costs would be excessive, and it would be unfriendly to clinical use. Fine-tuning allows updating only some weights or specified blocks, significantly shortening the loop training time. Even if the feature distribution of new data differs greatly from that of old features in the future, other incremental learning strategies (such as replay-based mechanisms) can be considered; however, overall, fine-tuning training is the most efficient retraining scheme in this embodiment.
[0119] Please continue to refer to Figure 3, which is the first block diagram of a system for detecting abnormal lung function based on sound signals.
[0120] According to another embodiment, as shown in Figure 3, a "System for Detecting Abnormal Lung Function Based on Sound Signals" (hereinafter referred to as System 300) is used. System 300 can be implemented on mobile devices, personal computers, servers, or cloud computing platforms to determine lung function parameters based on the breathing and / or sound signals of the test subject, and can continuously update the model through real or augmented lung function parameters.
[0121] System 300 includes a receiving module 310 and a processing module 320. The receiving module 310 is used to receive or capture raw audio signals from at least one subject, which may include the subject's breathing sounds, coughing sounds, or reading aloud. For example, audio can be collected from a mobile device or an earpiece microphone and transmitted to system 300 via a network or local terminal channel. The receiving module 310 can also simultaneously capture basic information such as the patient's age, gender, and height for use in subsequent formalization or analysis steps.
[0122] The processing module 320 is connected to the receiving module 310 and is responsible for performing noise filtering and feature extraction on the original audio signal, and then inputting it into the machine learning model to generate the first predicted lung function parameters; then, based on a preset reference benchmark (e.g., PEF BEST, ACT score range), the first predicted lung function parameters are converted into the second predicted lung function parameters.
[0123] The noise filtering performed by the processing module 320 can be implemented through spectral subtraction or speech activity detection to reduce background noise or invalid audio segments, producing a filtered audio signal that retains the audio features to the greatest extent possible. Feature extraction, for example, uses acoustic features such as Mel-frequency cepstral coefficients (MFCC) or fundamental frequency (F0) jitter and amplitude jitter as input to the machine learning model. In terms of prediction and normalization, the processing module 320 infers from the extracted feature vectors, first obtaining the first predicted lung function parameters (PEF, FEV1, etc.), and then adjusting the parameters to more personalized second predicted lung function parameters based on preset reference benchmarks (age, height, etc.).
[0124] Please refer to Figure 4, which is a second block diagram of a system for detecting abnormal lung function based on sound signals.
[0125] According to another embodiment, the system 300 further includes a training module 330 and a signal connection processing module 320, used to obtain at least one patient's real lung function parameters (such as clinically measured PEF, ACT, FEV1, or FVC) and use them as a labeled audio sample set for a machine learning model or for verifying and correcting a second predicted lung function parameter. The model training process can employ cross-validation or a three-part method (training / validation / testing) to objectively evaluate model performance, and update weights by calculating error values. Furthermore, if new patient parameters are generated, they can be added to the labeled audio sample set as an incremental labeled dataset, and retraining or fine-tuning can be performed to ensure the model maintains good accuracy in long-term applications.
[0126] According to another embodiment, the system 300 further includes a data augmentation module 340, a signal connection receiving module 310, a processing module 320, and a training module 330, used to adjust the real lung function parameters and generate at least one amplified lung function parameter. This process can adjust factors such as age, height, or asthma control score to obtain the physiological conditions of a virtual patient, and then calculate amplified lung function parameters that meet acceptable clinical ranges, thereby performing the following objectives:
[0127] (1) Sample supplementation: Make minor adjustments to rare disease classifications (such as partially controlled or uncontrolled) to ensure that the machine learning model can obtain enough training examples in this category.
[0128] (2) Confirm clinical rationality: Use the PEF BEST formula or asthma control score intervals to confirm that the amplified sample does not deviate from common clinical sense.
[0129] (3) Retraining / correction: The difference between the prediction and expected values of the comparison amplified labeled audio sample set of the processing module 320 is used to update the parameters of the machine learning model, thereby improving the detection accuracy of abnormal lung function.
[0130] The above description is merely illustrative of preferred embodiments of the present invention and is not intended to limit the scope of implementation. Any simple substitutions and equivalent changes made in accordance with the scope of the patent application and the contents of the patent specification of the present invention shall fall within the scope of the patent application of the present invention.
[0131] S110~S150: Steps S210~S280: Steps 300: System 310: Receiver Module 320: Processing Module 330: Training Module 340: Data Enhancement Module
Claims
1. A computer-based method for detecting abnormal lung function based on audio signals, comprising the following steps: receiving an original audio signal from at least one subject, the original audio signal including the subject's breathing and / or vocalization; processing the original audio signal by performing noise filtering to generate a filtered audio signal; extracting at least one audio feature from the filtered audio signal, the at least one audio feature including a spectral feature and / or a vocalization feature; inputting the at least one audio feature into a machine learning model to generate a first predicted lung function parameter; normalizing the first predicted lung function parameter and converting it into a second predicted lung function parameter based on a preset reference benchmark, wherein the reference benchmark is provided as a basis for normalization according to the gender, age, and height of the at least one subject.
2. The method according to claim 1 further includes performing a training process, the training process including: Obtain at least one true lung function parameter from at least one patient; The real lung function parameters are used as a set of labeled audio samples for the machine learning model or as a means of correcting or validating the second predicted lung function parameters.
3. The method according to claim 2 further includes the following steps: adjusting the real lung function parameter and generating at least one amplified lung function parameter; using the amplified lung function parameter as an amplified labeled audio sample set of the machine learning model or for correcting or validating the second predicted lung function parameter.
4. The method according to claim 3 further includes the following steps: calculating a difference or ratio between the labeled audio sample set and the amplified labeled audio sample set; updating a parameter weight of the machine learning model and using the difference or ratio as an error measure to correct the prediction accuracy of the machine learning model for abnormal lung function.
5. The method according to claim 3 further includes the following steps: performing a cross-validation process to divide the labeled audio sample set and the amplified labeled audio sample set into multiple subsets, and using each subset in turn as a training set and a test set to evaluate the predictive accuracy of the machine learning model for abnormal lung function; and selecting the machine learning model with the optimal model configuration, the selection being based on the average accuracy or error performance measured by the cross-validation process.
6. The method according to claim 3 further includes the following steps: dividing the labeled audio sample set into a training set, a validation set, and a test set, while the augmented labeled audio sample set is divided into the training set and the validation set; training the machine learning model using the training set derived from the labeled audio sample set and the augmented labeled audio sample set; adjusting the hyperparameters or network structure of the machine learning model based on the prediction results derived from the validation set derived from the labeled audio sample set and the augmented labeled audio sample set; and after adjusting the hyperparameters or network structure, validating the machine learning model's prediction accuracy for abnormal lung function using the test set derived from the labeled audio sample set.
7. The method according to claim 3, wherein the first predicted lung function parameter, the second predicted lung function parameter, the actual lung function parameter and the amplified lung function parameter respectively include at least one of peak expiratory flow rate, optimal peak expiratory flow rate, asthma control test score, forced expiratory volume in one second and maximum vital capacity.
8. The method according to any one of claims 3 to 7, wherein the step of adjusting the true lung function parameter further comprises: Modify at least one of the patient's age, height, and weight to obtain a physiological condition of a virtual patient; The amplified pulmonary function parameter corresponding to the physiological condition is calculated, which is derived by a clinical formula and / or a physiological model; and the amplified pulmonary function parameter is checked to see if it falls within an acceptable clinical range. If it does, the amplified pulmonary function parameter is included in the amplified labeled audio sample set; otherwise, the amplified pulmonary function parameter is discarded.
9. The method according to claim 8 further includes the steps of: receiving a new true lung function parameter from at least one new patient; adding the new true lung function parameter to an incremental labeled dataset used to update the labeled audio sample set; and fine-tuning the machine learning model by performing another training on the machine learning model using the incremental labeled dataset.
10. The method of claim 1, wherein the noise filtering includes a spectral subtraction, which involves selecting a speech-free segment from the original audio signal as background noise, calculating an average spectrum of the background noise, and subtracting the average spectrum from the spectrum of the original audio signal by linear subtraction.
11. The method according to claim 1, wherein the noise filtering includes a speech activity detection based on a short-time energy threshold, the short-time energy being the average of the squared amplitude of the original audio signal over a predetermined time window, the short-time energy threshold being a dynamic reference value calculated from ambient background noise, used to determine a speech activity, the speech activity extraction method comprising: The short-time energy is compared with the short-time energy threshold for each time window. When the short-time energy of multiple consecutive time windows exceeds the short-time energy threshold, it is determined as the start of the speech activity. When the short-time energy of multiple consecutive time windows is lower than the short-time energy threshold, it is determined as the end of the speech activity. The time period of the speech activity is extracted as a valid speech segment, and background noise is excluded.
12. A system for detecting abnormal lung function based on sound signals, comprising: A receiving module is configured to receive or capture an original audio signal from at least one subject, the original audio signal including the subject's breathing and / or vocalizations; a processing module is signal-connected to the receiving module, configured to perform noise filtering on the original audio signal to generate a filtered audio signal, extract at least one audio feature from the filtered audio signal, input the at least one audio feature into a machine learning model to generate a first predicted lung function parameter, and then convert the first predicted lung function parameter into a second predicted lung function parameter based on a preset reference benchmark.
13. The system according to claim 12 further includes a training module signal-connected to the processing module for acquiring a real lung function parameter of at least one patient and using the real lung function parameter as a labeled audio sample set of the machine learning model or as a means of correcting or validating the second predicted lung function parameter.
14. The system according to claim 13 further includes a data augmentation module signal-connected to the receiving module, the processing module, and the training module, for adjusting the real lung function parameters and generating at least one amplified lung function parameter, and using the amplified lung function parameter as an amplified labeled audio sample set of the machine learning model or for correcting or validating the second predicted lung function parameter.
15. The system according to claim 14, wherein the processing module further calculates a difference or ratio between the labeled audio sample set and the amplified labeled audio sample set, and updates a parameter weight of the machine learning model, using the difference or ratio as an error measure to correct the accuracy of the machine learning model in predicting lung function abnormalities.
16. The system according to claim 14, wherein the processing module further performs a cross-validation process, dividing the labeled audio sample set and the amplified labeled audio sample set into multiple subsets, and using each subset in turn as a training set and a test set to evaluate the predictive accuracy of the machine learning model for lung function abnormalities, and selecting the machine learning model with the optimal model configuration, the selection being based on the average accuracy or error performance measured by the cross-validation process.
17. The system according to claim 14, wherein the processing module further divides the labeled audio sample set into a training set, a validation set, and a test set, and the augmented labeled audio sample set is divided into the training set and the validation set. The machine learning model is trained by combining the training set segmented from the labeled audio sample set and the augmented labeled audio sample set. The hyperparameters or network structure of the machine learning model are adjusted based on the prediction results combined with the validation set segmented from the labeled audio sample set and the augmented labeled audio sample set. After the hyperparameters or network structure adjustment is completed, the machine learning model is verified for predictive accuracy of lung function abnormalities using the test set from the labeled audio sample set.
18. The system according to claim 14, wherein the first predicted lung function parameter, the second predicted lung function parameter, the actual lung function parameter and the amplified lung function parameter each include at least one of peak expiratory flow rate, optimal peak expiratory flow rate, asthma control test score, forced expiratory volume in one second and maximum vital capacity.
19. The system according to any one of claims 14 to 18, wherein when adjusting the real pulmonary function parameters, the data augmentation module corrects at least one of the patient's age, height, and weight to obtain a physiological condition of a virtual patient, and calculates the amplified pulmonary function parameters corresponding to the physiological conditions, the amplified pulmonary function parameters being derived by a clinical formula and / or a physiological model, and then checks whether the amplified pulmonary function parameters fall within an acceptable clinical range; if so, the amplified pulmonary function parameters are included in the amplified labeled audio sample set.
20. The system of claim 19, wherein the training module further receives a new real lung function parameter from at least one new patient and adds the new real lung function parameter to an incremental labeled dataset used to update the labeled audio sample set; then fine-tunes the machine learning model by using the incremental labeled dataset to perform another training on the machine learning model.
21. The system of claim 12, wherein the noise filtering includes a spectral subtraction method for selecting a speech-free segment from the original audio signal as background noise, calculating an average spectrum of the background noise, and subtracting the average spectrum from the spectrum of the original audio signal by linear subtraction.
22. The system according to claim 12, wherein the noise filtering includes a speech activity detection based on a short-time energy threshold, the short-time energy being the average of the squared amplitude of the original audio signal over a predetermined time window, the short-time energy threshold being a dynamic reference value calculated from ambient background noise, used to determine a speech activity, the speech activity extraction method comprising: The short-time energy is compared with the short-time energy threshold for each time window. When the short-time energy of multiple consecutive time windows exceeds the short-time energy threshold, it is determined as the start of the speech activity. When the short-time energy of multiple consecutive time windows is lower than the short-time energy threshold, it is determined as the end of the speech activity. The time period of the speech activity is extracted as a valid speech segment, and background noise is excluded.
Citation Information
Patent Citations
Method and system for identifying lung physiological sound
CN112690823A
VMD-CNN-based lung sound feature recognition and classification method and system
CN113679413A
Respiratory function detection method and device
CN114680867A
Method and system for establishing lung function examination model
CN115137349A
Lung function parameter measuring method and system based on voice signals
CN115644848A
