AECOPD early warning system based on lung sound digital biomarker

Abstract

The present application relates to a kind of AECOPD early warning system based on lung sound digital biomarker, including the acquisition module for collecting the lung breath sound of subject;For extracting lung breath sound feature data from the lung breath sound of the acquisition module acquisition the feature extraction module;Analysis module comprising prediction model and using the prediction model to the lung breath sound and / or the time series data analysis of the lung breath sound feature data of the subject, and the output module of signal connection with the analysis module, the output module is configured as: according to the analysis result of the time series data of the lung breath sound obtained by the analysis module, the prediction result of the disease development path of the subject in future period of time and the suggestion to the subject are output.This application can overcome the difficulty that prior art cannot exclude interference factors introduced when subject detects by oneself when predicting AECOPD.

Description

Technical Field

[0001] This invention relates to the field of digital medical technology, specifically to an early warning system, and more particularly to an AECOPD early warning system based on digital biomarkers of lung sounds. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a lung disease characterized by airflow limitation that is not completely reversible, progresses over time, and primarily affects the lungs.

[0003] Acute exacerbation of chronic obstructive pulmonary disease (AECOPD) is defined as an acute worsening of respiratory symptoms. It is a critical period for chronic obstructive pulmonary disease (COPD) and a major factor determining the health status and prognosis of COPD patients.

[0004] The course of chronic obstructive pulmonary disease (COPD) is characterized by recurrent exacerbations, acute worsening, gradual deterioration of the condition, continuous decline in respiratory function, and ultimately respiratory failure leading to death. Acute exacerbations are a critical period in the progression of COPD. Therefore, strengthening the prediction and assessment of acute exacerbations (AECOPD) of chronic obstructive pulmonary disease is key to controlling COPD.

[0005] Although COPD exacerbations can be suppressed through continuous treatment when they occur, improper medication use (e.g., patients arbitrarily stopping medication during long-term use; increasing medication frequency without authorization due to weather changes causing shortness of breath or difficulty breathing), environmental factors (air pollution, rapid increase in fine dust, rapid temperature changes), delayed detection of symptoms directly related to COPD, and sudden outbreaks of respiratory viruses (upper respiratory tract viral infections and tracheobronchial bacterial infections) can all lead to rapid deterioration and death in COPD patients within a short period. COPD exacerbations are characterized by recurrent episodes and unpredictable relapse times, and currently there is no accurate method for prediction. Seeking medical attention only after an acute exacerbation can lead to adverse consequences. Therefore, an early warning system capable of predicting COPD exacerbations is needed.

[0006] In recent years, deep learning has been widely used in medicine. Current research utilizes lung sounds for the diagnosis of respiratory diseases. Existing technology, such as Chinese patent application CN115424721A, provides a method for constructing a chronic obstructive pulmonary disease (COPD) identification system based on digitized lung sounds, and its application. The system includes the following modules: a data acquisition module, a preprocessing module, a conversion module, a data module, a data processing module, a data augmentation training module, a deep learning module, an identification module, a grading module, and a display module. The data acquisition module collects lung sounds generated by airflow friction during respiration from candidates, digitizes and stores them for subsequent preprocessing. The preprocessing module removes environmental noise, heart sounds, and power line interference from the collected lung sound records, and performs normalization to obtain valid lung sound data. The conversion module converts the preprocessed audio into Mel spectrograms through Fourier transform and Mel filter banks. This technical solution can analyze lung sound spectrum through a trained deep learning model and classify lung sound spectrum into three categories: COPD, healthy, and other non-COPD lung diseases. However, it cannot effectively predict the occurrence of acute exacerbations of COPD, and it is often only discovered when the candidate has already entered the acute exacerbation phase of COPD.

[0007] However, the worsening of the condition usually occurs after a period of accumulation. Current technology cannot effectively predict the disease from large amounts of health data. It lacks focus on specific markers preceding acute exacerbations of COPD, thus failing to predict the probability of AECOPD and its subsequent disease progression, thereby missing the optimal window for patient intervention. Regarding patient-collected data, factors often influence prediction results. For example, patients may experience shortness of breath after certain activities, and the data collected in such cases may not reflect the patient's true condition. Current technology does not offer a solution to this problem.

[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] To address the shortcomings of existing technologies, this invention provides an AECOPD early warning system based on digital biomarkers of lung sounds, comprising: an acquisition module for collecting lung breath sounds from a subject; a feature extraction module for extracting lung breath sound feature data from the lung breath sounds collected by the acquisition module; an analysis module including a prediction model and using the prediction model to perform time-series data analysis on the subject's lung breath sounds and / or lung breath sound feature data; and an output module signal-connected to the analysis module, the output module being configured to: based on the analysis results of the time-series data of lung breath sounds obtained by the analysis module, output a prediction result of the subject's disease progression path over a future period and suggestions for the subject.

[0010] The beneficial effects of this technical solution are as follows: This technical solution first uses an acquisition module to collect lung breath sounds from the subject. Since directly acquired lung breath sounds contain noise, a feature extraction module is needed to extract characteristic lung breath sound data (lung breath sound feature data). Further, an analysis module performs time-series analysis on the lung breath sound feature data. The analysis module includes a prediction model, which is established by training on digital biomarker data of lung breath sounds. This digital biomarker data consists of lung breath sound data from other related diseases, containing respiratory audio frequencies from various lung diseases. Although these are not specific to AECOPD, they allow the model to understand and abstract the basic patterns of lung breath sounds under various environments and conditions. After obtaining the pre-trained model, it is then fine-tuned using AECOPD-specific lung breath sound data, thereby ensuring that the obtained mature prediction model adapts to the specific sound patterns and prediction tasks of AECOPD. The AECOPD early warning system provided in this application can perform in-depth analysis of time-series data, effectively capturing time-dependent patterns and complex patterns in the data. By associating specific characteristic breath sounds with AECOPD, the early warning system can predict the disease progression path of the subject over a future period, thereby improving the timeliness and accuracy of disease prediction. At the same time, based on the subject's disease progression path, the output module 400 can output targeted suggestions for the subject's reference.

[0011] Current technologies typically determine whether a subject has COPD or AECOPD based on lung breath sounds, outputting a result indicating whether the subject has COPD or has progressed to AECOPD. Since acute exacerbations are the primary pathway for COPD progression and a leading cause of hospitalization and death, early prediction and intervention for AECOPD are crucial. The overall progression time for COPD symptom exacerbation is 0–14 days, while studies show that 90% of AECOPD cases progress from symptom worsening to full acute exacerbation within 0–5 days. When symptoms exceed the usual variability (e.g., increased respiratory rate), some patients may self-adjust their medication, but this self-adjustment does not improve symptoms and requires prompt medical attention for examination and treatment. Without timely treatment during an AECOPD attack, the patient's life may be in danger. This new technology analyzes lung breath sound characteristics data to determine the probability of disease onset and the disease progression path before the subject enters the acute exacerbation phase. This allows for personalized recommendations to be provided in advance, preventing life-threatening situations due to delayed treatment.

[0012] According to a preferred embodiment, the analysis results of the analysis module are characterized by time-frequency characteristic data.

[0013] The beneficial effects of this technical solution are as follows: COPD patients produce additional lung sounds during respiration. Due to the different diameters of the bronchi and bronchioles (and thus different degrees of obstruction), the additional sounds produced by lung respiration vary, exhibiting different time-frequency characteristics within one or several respiratory cycles. When a COPD patient's condition changes, the corresponding time-frequency characteristics of these additional sounds also change. In particular, the worsening of the condition is cumulative, and during this cumulative process, the time-frequency characteristics of lung sounds also change. Therefore, the time-frequency characteristic data of lung sounds can be used to predict the probability of a subject developing AECOPD and the disease progression path in the future.

[0014] According to a preferred embodiment, the disease progression path includes the subject's condition stabilization, improvement, and deterioration, and the output module outputs personalized suggestions for the subject based on the disease progression path.

[0015] The beneficial effects of this technical solution are as follows: Based on the prediction results, subjects can understand the changes in their own condition. According to these changes, subjects can pay more attention to improving their lifestyle habits. For example, when a subject learns through the warning system that their condition is stable, it indicates that their recent lifestyle and medication habits are being maintained well, and they can appropriately engage in some exercise (walking, cycling, abdominal breathing exercises) to improve their quality of life. When a subject learns through the warning system that their condition has improved, it indicates that their physical recovery is good, and they can engage in aerobic exercise (stretching exercises, chest expansion exercises, jogging, rehabilitation exercises, etc.) to improve their athletic ability and muscle strength. When a subject learns through the warning system that their condition has worsened, they need to seek medical treatment at a hospital promptly. In particular, the disease progression path indicating worsening condition serves as a strong reminder for subjects. In addition to seeking medical attention, subjects also need to pay attention to changes in their lifestyle habits and environmental factors, and take extra precautions to prevent further aggravation of their condition.

[0016] Secondly, based on the acquired disease progression path, the output module can provide individualized suggestions for the subject's condition. Some subjects, due to limited cognitive abilities, may adjust their medication dosage themselves instead of seeking medical attention when their symptoms worsen. This improper medication use can further exacerbate the severity of their condition. Alternatively, some subjects, feeling their recovery has improved or stabilized, may gradually neglect prescribed medication procedures, especially discontinuing medication on their own. Furthermore, when subjects feel their recovery has improved or stabilized, they may change previously healthy habits, such as drinking alcohol or smoking, and neglecting personal protective measures.

[0017] In addition, some participants may be in good health, but they may be overly worried about their condition and go to the hospital frequently. This not only increases the participants' treatment costs, but also creates pressure on them to seek medical care.

[0018] This technical solution outputs the subject's disease progression path and personalized suggestions through the output module, making it easier for the subject to understand their own condition. It can remind the subject to seek medical attention in time when they realize that their condition is worsening, and can also eliminate the subject's excessive worry about their condition.

[0019] According to a preferred embodiment, the early warning system further includes a monitoring module for monitoring changes in the subject's posture, and the analysis module is configured to output an analysis result indicating that the subject's condition is deteriorating when the subject's posture collected by the monitoring module has not changed but the time-frequency characteristic data indicates that the subject has a tendency to AECOPD.

[0020] The technical advantages of this solution are as follows: The condition of subjects is highly variable, and many factors can influence monitoring through an early warning system. For example, some subjects may experience wheezing after engaging in activities (changing from lying down to walking, from squatting to standing, from sitting to walking, etc.) during non-acute COPD flare-ups, but this usually resolves after a period of rest. Due to differences in subject cognition, testing may be conducted immediately after walking and sitting down. The change in posture may cause shortness of breath, and the collected time-frequency characteristic data of lung sounds may not accurately reflect the subject's condition. When the monitoring module detects no change in the subject's posture, it eliminates the possibility of abnormal time-frequency characteristic data caused by posture changes. Furthermore, if the time-frequency characteristic data indicates a tendency towards acute exacerbation of COPD (AECOPD), timely and accurate early warnings can be provided, avoiding psychological distress caused by false alarms.

[0021] According to a preferred embodiment, the time-frequency characteristic data includes the duration, dominant frequency, and amplitude of wheezing, which reflect changes in the subject's condition.

[0022] The technical effects of this approach are as follows: When the lung sounds of COPD subjects change, the changes can be reflected in the duration and frequency of wheezing. If the degree of obstruction increases at least in one of the trachea, bronchi, bronchioles, or small bronchioles, the resistance to airflow through these sites increases, and correspondingly, the duration of wheezing will be prolonged, and the frequency of wheezing will also increase (the pitch will become higher).

[0023] According to a preferred embodiment, the output module is configured to: when the subject's posture remains unchanged, and the time-frequency characteristic data of lung breath sounds over at least one respiratory cycle shows that the duration of wheezing exceeds a first duration and / or the dominant frequency exceeds a first frequency response value, output a disease progression path for the subject in the future that may lead to AECOPD, and simultaneously output a first type of suggestion for the subject's condition to worsen. Preferably, the first duration is the duration of a reference wheezing sound indicating a tendency for the subject's condition to worsen. Preferably, the first frequency response value is the dominant frequency of the reference wheezing sound indicating a tendency for the subject's condition to worsen. Preferably, the first type of suggestion is a reminder to the subject to seek medical attention at a hospital promptly.

[0024] The technical benefits of this solution are as follows: The subject's posture remains unchanged, ruling out respiratory abnormalities caused by the subject's own activities. Furthermore, when the duration of wheezing exceeds a first duration and / or the dominant frequency exceeds a first frequency response value within at least one respiratory cycle, it indicates a worsening of the subject's condition. This suggests an increased degree of obstruction in at least one of the trachea, bronchi, bronchioles, or small bronchioles, leading to this phenomenon. When the condition worsens, some subjects (especially those with lower cognitive abilities) may increase their medication dosage or frequency to avoid inconvenience. These actions pose a threat to their safety. Therefore, the first type of recommendation is provided to remind the subject to seek medical attention promptly rather than changing their medication regimen on their own.

[0025] According to a preferred embodiment, the output module is configured to: when the subject's posture remains unchanged, and the time-frequency characteristic data of lung breath sounds over at least one respiratory cycle shows that the duration of wheezing is less than a first duration and / or the dominant frequency is less than a first frequency response value, outputting that the subject's disease progression path for a future period is stable, and simultaneously outputting a second type of suggestion for the subject's stable condition. Preferably, the second type of suggestion is to maintain the lifestyle and medication prescribed by the doctor, and to be able to perform exercise of a first intensity. Preferably, the first intensity of exercise can be walking, abdominal breathing exercises, etc.

[0026] The beneficial effects of this technical solution are as follows: The subject's posture remains unchanged, ruling out respiratory abnormalities caused by the subject's own activities. Furthermore, when the duration of wheezing is less than a first duration and / or the dominant frequency is less than a first frequency response value within at least one respiratory cycle, it indicates that the subject's condition is stable and frequent hospital visits are unnecessary. A second type of suggestion is output to remind the subject to maintain the prescribed lifestyle and medication regimen, and to engage in appropriate intensity exercise to improve quality of life. For subjects who are concerned about their health, this early warning system allows them to monitor their health status in real time, reducing psychological stress. Some subjects, when experiencing abnormalities, immediately think of going to the hospital for examination. However, these abnormalities may be caused by incorrect timing of testing (such as immediately after activity or significant emotional fluctuations). Frequent hospital visits not only increase medical costs but also increase the burden of medical care. This technical solution addresses these issues by using the time-frequency characteristics of lung breath sounds to obtain the subject's disease progression path and providing targeted suggestions, preventing subjects from arbitrarily changing their lifestyle or disobeying medical advice due to insufficient understanding.

[0027] According to a preferred embodiment, the output module is configured to: when the subject's posture changes, and the time-frequency characteristic data of lung breath sounds over at least one respiratory cycle shows that the duration of wheezing is less than a first duration and / or the dominant frequency is less than a first frequency response value, output the subject's disease progression path for a future period as "condition improvement," and simultaneously output a third type of suggestion for the subject's condition improvement. Preferably, the third type of suggestion is to maintain the lifestyle and medication prescribed by the doctor, and to be able to perform exercise of a second intensity. Preferably, the second intensity exercise can be jogging, rehabilitation exercises, stretching exercises, etc.

[0028] Preferably, the second strength is greater than the first strength.

[0029] The beneficial effects of this technical solution are as follows: When the subject's posture changes, it indicates that the subject has engaged in some activity in the period surrounding the collection of lung breath sounds. Activity causes changes in breathing; generally, COPD subjects will experience wheezing, and correspondingly, the duration of wheezing will exceed the first duration and / or the dominant frequency will exceed the first frequency response value. However, if, despite the change in the subject's posture, the duration of wheezing is less than the first duration and / or the dominant frequency is less than the first frequency response value, it indicates that the subject's condition is stable and has improved, thus prompting a third-category recommendation. Even if the subject's condition has improved, it is still necessary to maintain the lifestyle and medication regimen prescribed by the doctor, but increasing the intensity and variety of exercise can further improve the subject's quality of life.

[0030] The targeted suggestions provided by the output module are particularly useful for users with limited cognitive abilities. Even when a subject's condition is stable or improving, some may still frequently visit the hospital due to excessive worry. Based on the warning system provided in this application, subjects can monitor their condition in real time. Furthermore, some subjects, feeling their condition has improved or stabilized, may discontinue medication, reduce dosage, neglect protective measures (due to temperature changes, air quality alterations), or start drinking alcohol—all unhealthy habits. Based on the warning system provided in this application, the output module can provide targeted suggestions based on the individual subject's condition development, reminding them to maintain healthy lifestyle and medication habits.

[0031] According to a preferred embodiment, the output module is configured to extend the duration of the acquisition module acquiring the subject's lung breath sounds when the subject's posture changes and the time-frequency characteristic data of lung breath sounds in at least one respiratory cycle shows that the duration of wheezing exceeds a first duration and / or the dominant frequency exceeds a first frequency response value.

[0032] The beneficial effects of this technical solution are as follows: Changes in the subject's posture may cause changes in the subject's breathing. In this case, the duration of wheezing may exceed the first duration and / or the dominant frequency may exceed the first frequency response value, but this does not reflect the subject's true condition. Therefore, it is necessary to extend the duration of the acquisition module to acquire the subject's lung breath sounds in order to obtain the duration and dominant frequency of the subject's wheezing under normal conditions.

[0033] Preferably, the duration for which the acquisition module acquires the subject's lung breath sounds is extended, i.e., the number of respiratory cycles during detection is increased. Preferably, the number of respiratory cycles during detection can be increased from 2 to 15. Specifically, the number of respiratory cycles during detection is increased by 2. Specifically, the number of respiratory cycles during detection is increased by 5. Specifically, the number of respiratory cycles during detection is increased by 8. Specifically, the number of respiratory cycles during detection is increased by 10. Specifically, the number of respiratory cycles during detection is increased by 15.

[0034] When the subject's posture changes, the duration and frequency of wheezing sounds extracted in the first few respiratory cycles cannot reflect the subject's true condition. However, the subsequent respiratory cycles can reflect the true condition. Therefore, it is necessary to extend the duration of lung breath sounds collected by the acquisition module. Attached Figure Description

[0035] Figure 1 This is a simplified schematic diagram of the module connection relationship of an early warning system according to a preferred embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the workflow of an early warning system according to a preferred embodiment of the present invention.

[0037] List of reference numerals

[0038] 100: Acquisition module; 200: Feature extraction module; 300: Analysis module; 400: Output module; 500: Monitoring module. Detailed Implementation

[0039] The following is a detailed explanation with reference to the accompanying drawings.

[0040] Example 1

[0041] This embodiment provides an AECOPD early warning system based on digital biomarkers of lung sounds, such as Figure 1As shown. The early warning system may include a data acquisition module 100, a feature extraction module 200, an analysis module 300, and an output module 400. The data acquisition module 100 is used to acquire lung breath sounds of the subject. The feature extraction module 200 is used to extract lung breath sound feature data from the lung breath sounds acquired by the data acquisition module 100. The analysis module 300 includes a prediction model. The analysis module 300 uses the prediction model to perform time-series data analysis on the lung breath sounds and / or lung breath sound feature data of the subject. Preferably, the analysis module 300 is signal-connected to the output module 400 to transmit the analysis results to the output module 400. Specifically, the connection between the analysis module 300 and the output module 400 is a WLAN connection or a Bluetooth connection. Preferably, the data acquisition module 100 is signal-connected to the output module 400. Specifically, the connection between the data acquisition module 100 and the output module 400 is a WLAN connection or a Bluetooth connection. Preferably, the output module 400 can save the lung breath sounds acquired by the data acquisition module 100. Preferably, the acquisition module 100 and the feature extraction module 200 are connected by a signal. Specifically, the connection between the acquisition module 100 and the feature extraction module 200 is a WLAN connection or a Bluetooth connection. Preferably, the analysis module 300 is connected by a signal to the feature extraction module 200 to receive the lung breath sound feature data extracted by the feature extraction module 200.

[0042] Preferably, the acquisition module 100 can be a lung sound acquisition device, such as a lung sound sensor.

[0043] According to a preferred embodiment, the output module 400 is configured to: based on the analysis results of the time-series data of lung breath sounds obtained by the analysis module 300, output the predicted results of the subject's disease development path in the future and suggestions for the subject.

[0044] According to a preferred embodiment, the analysis results of the analysis module 300 are characterized by time-frequency characteristic data.

[0045] According to a preferred embodiment, the time-frequency characteristic data includes the duration, dominant frequency, and amplitude of wheezing, which reflect changes in the subject's condition.

[0046] According to a preferred embodiment, the disease progression path includes the subject's condition stabilizing, improving, and worsening. The output module 400 outputs personalized suggestions for the subject based on the disease progression path. This technical solution, through the output module 400, outputs the subject's disease progression path and personalized suggestions, making it easier for the subject to understand their own condition. It can remind the subject to seek medical attention promptly when they realize their condition is worsening, and it can also alleviate the subject's excessive anxiety about their condition.

[0047] According to a preferred embodiment, the early warning system further includes a monitoring module 500 for monitoring changes in the subject's posture. The analysis module 300 is configured to: when the subject's posture collected by the monitoring module 500 has not changed but the time-frequency characteristic data indicates that the subject has a tendency to AECOPD, output an analysis result indicating that the subject's condition is worsening. Preferably, the monitoring module 500 and the analysis module 300 are signal-connected. Specifically, the connection between the monitoring module 500 and the analysis module 300 is a WLAN connection or a Bluetooth connection. Preferably, the monitoring module 500 is configured to: scan the activity area where the subject is located to obtain a detection signal reflecting the subject's current activity posture; determine the subject's current position and posture based on the scanning results; continuously scan the subject's current position to obtain a detection signal reflecting the subject's current position and posture for a specific duration; if the subject's current activity posture is the preset human body's movement pattern at the current time point, it is determined that the subject has not changed posture; if the subject's current activity posture is not the preset human body's movement pattern at the current time point, it is determined that the subject has changed posture. The monitoring module 500 can be, for example, a microwave radar device.

[0048] Figure 2 This is a schematic diagram of the workflow of the early warning system provided in this embodiment.

[0049] According to a preferred embodiment, the output module 400 is configured to: when the subject's posture remains unchanged, and the time-frequency characteristic data of lung breath sounds over at least one respiratory cycle shows that the duration of wheezing exceeds a first duration and / or the dominant frequency exceeds a first frequency response value, output the subject's disease progression path for a future period as a path of potential AECOPD deterioration, and simultaneously output a first type of suggestion for the subject's deterioration. Preferably, the first duration is the duration of a reference wheezing sound indicating a tendency for the subject's condition to worsen. Preferably, the first frequency response value is the dominant frequency of the reference wheezing sound indicating a tendency for the subject's condition to worsen. Preferably, the first type of suggestion is a reminder to the subject to seek medical attention at a hospital promptly. Specifically, the first duration can be within 80–120 ms. Specifically, the first duration can be 80 ms, 90 ms, 100 ms, 110 ms, or 120 ms. Specifically, the first frequency response value can be within 400–2500 Hz. The first frequency response value can be 1000Hz, 1200Hz, 1500Hz, 1800Hz, or 2000Hz.

[0050] According to a preferred embodiment, the output module 400 is configured to: when the subject's posture remains unchanged, and the time-frequency characteristic data of lung breath sounds over at least one respiratory cycle shows that the duration of wheezing is less than a first duration and / or the dominant frequency is less than a first frequency response value, outputting that the subject's disease progression path for a future period is stable, and simultaneously outputting a second type of suggestion for the subject's stable condition. Preferably, the second type of suggestion is to maintain the lifestyle and medication prescribed by the doctor, and to be able to perform exercise of a first intensity. The first intensity of exercise can be walking, abdominal breathing exercises, etc.

[0051] According to a preferred embodiment, the output module 400 is configured to: when the subject's posture changes, and the time-frequency characteristic data of lung breath sounds over at least one respiratory cycle shows that the duration of wheezing is less than a first duration and / or the dominant frequency is less than a first frequency response value, output the subject's disease progression path for a future period as "condition improvement," and simultaneously output a third type of suggestion for the subject's condition improvement. Preferably, the third type of suggestion is to maintain the lifestyle and medication prescribed by the doctor, and to be able to perform exercise of a second intensity. The second intensity exercise can be jogging, rehabilitation exercises, stretching exercises, etc. Preferably, the second intensity is greater than the first intensity.

[0052] According to a preferred embodiment, the output module 400 is configured to: when the subject's posture changes, and the time-frequency characteristic data of lung breath sounds within at least one respiratory cycle shows that the duration of wheezing exceeds a first duration and / or the dominant frequency exceeds a first frequency response value, extend the duration for which the acquisition module 100 acquires the subject's lung breath sounds. Extending the duration for which the acquisition module 100 acquires the subject's lung breath sounds increases the number of respiratory cycles during detection. The number of respiratory cycles increased during detection can be 2 to 15. Specifically, the number of respiratory cycles increased during detection is 2. Specifically, the number of respiratory cycles increased during detection is 3. Specifically, the number of respiratory cycles increased during detection is 4. Specifically, the number of respiratory cycles increased during detection is 5. Specifically, the number of respiratory cycles increased during detection is 6. Specifically, the number of respiratory cycles increased during detection is 10. Specifically, the number of respiratory cycles increased during detection is 15. Preferably, the specific number of respiratory cycles increased during detection can be determined based on the subject's activity type and intensity before detection. Preferably, the specific number of respiratory cycles increased during the test can be determined based on the subject's posture changes during the test. Since the physiological recovery time after physical activity or posture changes varies among subjects, it is necessary to increase the number of respiratory cycles during the test according to individual circumstances.

[0053] Preferably, the output module 400 can be a computer or a mobile device. Specifically, the output module 400 can be a smartphone, tablet, or computer.

[0054] Example 2

[0055] This embodiment is a specific generation scheme for the prediction model provided in Embodiment 1. Preferably, the early warning system includes a time-series database of digital biomarkers of lung breath sounds and a prediction model pre-trained and fine-tuned using artificial intelligence algorithms. The early warning system uses a Transformer model to process this time-series data. First, expert knowledge is introduced to initialize some hyperparameters of the Transformer model (such as the number of layers, hidden units, and heads) to conform to prior knowledge of the relationship between digital biomarkers of lung breath sounds and AECOPD, and then training is performed. The early warning system is first pre-trained on lung breath sound datasets of other related diseases (such as asthma, pneumonia, upper / lower respiratory tract infections, etc.), and then fine-tuned using AECOPD data. Finally, based on the analysis, the early warning system predicts the probability of a subject developing AECOPD and the possible disease progression path in the future, and provides personalized health advice. The results are displayed in the form of charts or reports and can be sent to designated terminals such as doctors' computers and subjects' mobile phones via email, SMS, or an app.

[0056] This embodiment provides a method for training and optimizing a prediction model. The method involves pre-training the prediction model on a large-scale general or respiratory-related sound dataset to learn generalized sound features, thus obtaining a pre-trained model. Then, the pre-trained model is fine-tuned using AECOPD-specific lung breath sound data to adapt it to a specific prediction task, ultimately obtaining the final prediction model. The training process can utilize techniques such as cross-validation and early stopping to prevent overfitting and improve the model's generalization ability.

[0057] According to a preferred embodiment, the specific implementation plan for obtaining the prediction model is as follows: ① Pre-training: The goal of the pre-training stage is to enable the model to learn generalized sound features. To this end, this embodiment selects lung breath sound datasets from other related diseases for pre-training. These datasets contain respiratory audio data from various lung diseases; although they are not specific to AECOPD, they allow the model to understand and abstract the basic patterns of lung breath sounds in various environments and conditions. The pre-training process employs self-supervised learning to train the model, for example, using the masked language model (MLM) task. In the MLM task, a subset of sound data is randomly selected and replaced with a special "mask". The goal of the pre-trained model is to predict these masked sound data. This helps the pre-trained model learn the dependencies between sound data and the inherent structure of sound. The pre-training process can utilize large-scale computing resources (e.g., multiple GPUs or TPUs) and train for multiple epochs until the model's performance on the pre-training task converges, thus obtaining the pre-trained model. ② Fine-tuning: After the pre-trained model learns generalized sound features, it is fine-tuned using AECOPD-specific lung breath sound data to adapt to the specific task of the early warning system, namely, predicting the development of AECOPD. The goal of the fine-tuning phase is to adapt the pre-trained model to the specific sound patterns of AECOPD and the prediction task. AECOPD-specific lung breath sound data is labeled as "normal" or "AECOPD," and then the model is subjected to a binary classification task. During fine-tuning, the model architecture (i.e., the parameters of the Transformer model) remains unchanged; only the model parameters are updated to adapt to the new task. During fine-tuning, some hyperparameters, such as the learning rate and batch size, are adjusted according to the needs of the specific task. In addition, regularization techniques, such as dropout or weight decay, are used during training to prevent overfitting. After fine-tuning, a predictive model capable of predicting the development of AECOPD is obtained. ③ Validation and Early Stopping: During training, the data is divided into training and validation sets. The predictive model is trained on the training set and its performance is validated on the validation set. If performance on the validation set begins to decline while performance on the training set continues to improve, training is stopped to prevent overfitting. This strategy is called "early stopping." ④ Cross-validation: To more reliably evaluate the performance of a predictive model, cross-validation is used. For example, k-fold cross-validation can be used. In this method, the data is divided into k subsets, the predictive model is trained on k-1 subsets, and validated on the remaining subset. This process is repeated k times, with each subset used as a validation set once. Finally, the average of the k validation results is taken as the model's final performance.

[0058] This embodiment also provides a time-series data analysis method, which uses a transformer model to process time-series data (digital biomarkers of lung breath sounds) to reveal the dynamic changes of the disease, thereby improving the accuracy and timeliness of prediction.

[0059] According to a preferred embodiment, the specific implementation plan for time series data analysis is as follows: ① Data preprocessing: Since the original form of data may contain noise, outliers, missing values, etc., preprocessing helps to clean and standardize the data, making its use in predictive model training more accurate and effective. When processing lung breath sound data, it may be necessary to suppress noise in the sound signal and remove interference from non-breath sounds, such as coughing and talking. In addition, since the amplitude of the breath sound signal may be affected by various factors (such as microphone position, body position, etc.), standardization is also required so that different records can be compared under the same standard. Feature extraction: Extracting features of time series data, such as trends, periodicity, peaks, and abrupt changes, to better understand the inherent laws and change patterns of the data, thereby improving the accuracy of prediction. ② Feature extraction: Although the Transformer model can process raw time series data, feature extraction can encode biomedically significant information into the model. For example, this technical solution can calculate various spectral characteristics of breath sounds (such as power spectral density, spectral peak, etc.) or time domain characteristics (such as respiratory cycle, respiratory stability, etc.). These features help the model better understand the inherent patterns and variations in breath sounds, thereby improving prediction accuracy. ③ Sequence Model Training: After preprocessing and feature extraction, the Transformer model is used to process the time-series data. The Transformer model, through its self-attention mechanism, effectively captures long-range dependencies in sequences, and is highly efficient and easily parallelizable. Based on prior expert knowledge, this technical solution can set initial values ​​for the Transformer model's hyperparameters (such as the number of layers, hidden units, and heads) to help the prediction model better capture features in lung breath sound data.

[0060] According to a preferred implementation, the Transformer model assigns weights to the importance of each position in the input sequence using a self-attention mechanism. For lung breath sound data, this means the predictive model can capture the interactions between breath sound features over a period of time. This is a crucial step in processing time-series data, especially data with inherent sequence dependencies such as breath sounds.

[0061] According to a preferred embodiment, the primary function of the multi-head attention mechanism in the Transformer model is to help the model model input sequence data from multiple perspectives simultaneously. Each "head" is essentially an independent self-attention mechanism, and the multi-head attention mechanism can learn and focus on different features in the sequence. Preferably, the multi-head self-attention mechanism allows the model to understand the data from multiple different perspectives, with each "head" focusing on different features and patterns. For example, in a task processing lung breath sounds, some "heads" may focus on identifying changes in respiratory rate, while others may focus more on changes in the texture of the breath sounds. In lung breath sound data, for example, one "head" may focus on changes in pitch, another on the stability of the audio, and yet another on the rhythm of the breath sounds. These "heads" work in parallel and share their "attention" outputs, enabling the model to understand the input lung breath sound data simultaneously from multiple perspectives. Furthermore, local attention mechanisms allow the model to focus more on recent data, as in diseases such as AECOPD, recent breath sound data may contain important information about changes in the condition.

[0062] According to a preferred embodiment, the multi-head attention mechanism first linearly transforms the input data into multiple "heads," each with its own query (Q), key (K), and value (V). Then, each "head" performs a self-attention operation on its query, key, and value, obtaining its own output. Finally, the outputs of all "heads" are concatenated and then subjected to another linear transformation to obtain the final output. Specifically, assuming an input X of d_model dimensions, this scheme first needs to define weight matrices for the query (Q), key (K), and value (V) for each "head." These weight matrices are the parameters that the model needs to learn. The weight matrix for each head is defined as follows:

[0063] WQi∈Rd_model×dk, WKi∈Rd_model×dk, WVi∈Rd_model×dv, where i represents the head number, there are a total of h heads, and dk and dv represent the dimensions of each head.

[0064] Then, calculate Q, K, and V for the i-th head:

[0065] Qi = xWQi

[0066] Ki=xWKi

[0067] Vi = xWVi

[0068] Next, the score for self-attention is calculated, using scaled dot product attention:

[0069] Score=softmax((QiKi)^T / sqrt(dk))Vi

[0070] This score represents the distribution of attention to each input. Softmax ensures that all scores add up to 1, while dividing by sqrt(dk) is to avoid the vanishing gradient problem caused by excessively large scores.

[0071] Then concatenate the scores of all the heads:

[0072] MultiHead(Q,K,V)=Concat(Score1,Score2,...,Scoreh)WO

[0073] Where WO is the output weight matrix, which is also a parameter that needs to be learned. Its dimension is hdv×d_model, which is used to convert the concatenated score back to the model dimension.

[0074] Therefore, when applied to lung sound data, the multi-head attention mechanism allows the model to capture the internal diversity of a lung sound data set, such as pitch, audio stability, rhythm, and other characteristics, thereby improving the accuracy of the model's AECOPD prediction.

[0075] Since the Transformer model itself lacks the ability to capture the sequential information in the input sequence, this scheme introduces positional encoding to enable the model to understand the sequential information. Positional encoding is introduced to provide a unique representation for each different position in the input sequence. Typically, positional encoding can be generated using sine and cosine functions, a method that can provide unique and continuous encodings for sequences of arbitrary length. Its generation method is as follows:

[0076]

[0077]

[0078] Here, pos represents the position, i represents the dimension, and dmodel represents the dimension of the model. Specifically, for a given position pos, the position code is calculated using a sine function in even-numbered dimensions and a cosine function in odd-numbered dimensions.

[0079] This positional encoding method allows the model to distinguish different positions in the input sequence, thus better capturing the temporal characteristics of lung breath sound data. Simultaneously, the periodicity of the sine and cosine functions enables the model to capture the periodic features of the input sequence, which is particularly important when processing lung breath sound data, as this data exhibits obvious periodic characteristics (such as respiratory rate).

[0080] According to a preferred embodiment, the Transformer model is based on an encoder-decoder architecture, where the encoder is used to understand the input sequence, and the decoder is used to generate the prediction sequence. In this scenario, the encoder encodes lung breath sound data, and the decoder generates AECOPD status predictions for a future period based on this encoded information.

[0081] According to a preferred implementation, after the Transformer model is trained, it is applied to sequence prediction. An effective strategy is to use a sliding window method, using data from the most recent N time steps as input for each prediction. For AECOPD, the prediction result may be whether the subject will experience an acute exacerbation of AECOPD within a future period (e.g., the next 24 hours or the next week), thus enabling early warning and necessary intervention.

[0082] According to a preferred embodiment, short-term sequence prediction can be combined with other causally related biomarkers to make longer-term predictions, thereby enabling early intervention. Since these biomarkers may have a causal relationship with disease development, their changes may indicate early signs or future trends of the disease. For example, assuming a specific lung sound characteristic (e.g., breath sound variability) is found to be causally related to the onset of AECOPD, and it is also found that changes in this characteristic typically begin some time before the onset (e.g., one week), this information can be used to guide the predictive model provided in this embodiment: when predicting whether an AECOPD onset will occur within the next week, special attention should be paid to changes in this characteristic.

[0083] To better understand the predictions from a clinical perspective, the attention weights of the Transformer model can be further analyzed to interpret the model's predictions. Attention weights reveal the model's emphasis on various features during the decision-making process; for example, which biomarkers or data points at which times have the greatest impact on the prediction results. This understanding can help physicians further comprehend disease development, identify potential key predictive factors, and thus implement more precise interventions.

[0084] To prevent overfitting, early stopping and regularization techniques (such as Dropout and L2 regularization) can be used. Additionally, to improve the model's generalization ability, techniques such as cross-validation and hyperparameter tuning can be used.

[0085] The AECOPD disease early warning system in this embodiment can perform in-depth analysis of time-series data, effectively capturing time-dependent patterns and complex patterns in the data, thereby improving the accuracy and timeliness of disease prediction.

[0086] According to a preferred embodiment, this embodiment provides a prediction and output method for an early warning system: ① Prediction result generation: Based on the sound features extracted from time-series data processed by the Transformer model, the system predicts the probability of a subject developing AECOPD and the possible disease progression path within a future period. The prediction results may include the disease status (e.g., worsening, stable, improving), influencing factors (e.g., lifestyle habits, environmental factors), and the degree of impact (e.g., mild, moderate, severe). ② Result interpretation and visualization: To make the prediction results easy to understand, the system provides detailed explanations, such as which biomarkers or behavioral factors are most likely to lead to the occurrence or development of the disease, and the degree of influence of these factors. Furthermore, the system visualizes the results using charts, heatmaps, trend lines, etc., to help users understand the prediction results more intuitively. ③ Personalized health recommendations: Based on the prediction results, the system also provides personalized health recommendations, such as lifestyle changes (e.g., quitting smoking, healthy eating), regular checkups (e.g., lung function tests, blood gas analysis), and the use of specific medications (e.g., long-acting bronchodilators, antibiotics). These recommendations will consider the subject's specific circumstances, such as age, gender, and disease history. ④ Results Delivery: Predicted results and health recommendations will be sent to designated terminals, such as doctors' computers and participants' mobile phones, in the form of electronic reports. The system supports multiple communication methods, such as email, SMS, and app notifications. Furthermore, the report format and content can be customized according to the recipient's needs.

[0087] Through the above steps, the AECOPD disease early warning system of the present invention can not only accurately predict the development of the disease, but also provide prediction results and health advice in an easy-to-understand and practical way, thereby helping doctors and subjects to better manage the disease.

[0088] 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 do not constitute a limitation on 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. Throughout the text, features introduced by "preferredly" are merely optional and should not be construed as mandatory. Therefore, the applicant reserves the right to abandon or delete relevant preferred features at any time.

Claims

1. An AECOPD early warning system based on digital biomarkers of lung sounds, comprising: Acquisition module (100) for collecting lung breath sounds of subjects. Feature extraction module (200) for extracting lung breath sound feature data from the lung breath sounds acquired by the acquisition module (100). An analysis module (300) comprising a prediction model and using the prediction model to perform time-series data analysis on the lung breath sounds and / or the lung breath sound feature data of the subject, and an output module (400) signal-connected to the analysis module (300), characterized in that, The output module (400) is configured as follows: Based on the analysis results of the time-series data of the lung breath sounds obtained by the analysis module (300), the prediction results of the disease development path of the subject in the future period and the suggestions for the subject are output. The analysis results of the analysis module (300) are characterized by time-frequency feature data; The disease development path includes the subject's condition stabilization, improvement, and deterioration, and the output module (400) outputs suggestions for the subject based on the disease development path; The system also includes a monitoring module (500) for monitoring changes in the subject's posture, and the analysis module (300) is configured to: When the monitoring module (500) collects data indicating that the subject's posture has not changed and the time-frequency characteristic data indicates that the subject has a tendency to AECOPD, it outputs the analysis results of the subject's condition deterioration.

2. The system according to claim 1, characterized in that, The time-frequency characteristic data includes the duration, dominant frequency, and amplitude of wheezing, which reflect changes in the subject's condition.

3. The system according to claim 1, characterized in that, The output module (400) is configured as follows: When the subject's posture remains unchanged, and the time-frequency characteristic data of lung breath sounds over at least one respiratory cycle shows that the duration of wheezing exceeds a first duration and / or the dominant frequency exceeds a first frequency response value, the disease progression path of the subject in the future is output as a path of possible AECOPD deterioration. Simultaneously, a first type of suggestion for the subject's condition deterioration is output, wherein... The first duration is the duration of a reference wheezing sound that indicates a tendency for the subject's condition to worsen; The first frequency response value is the dominant frequency of the reference wheezing sound that indicates the subject's tendency to worsen.

4. The system according to claim 1, characterized in that, The output module (400) is configured as follows: When the subject's posture does not change, and the time-frequency characteristic data of lung breath sounds within at least one respiratory cycle shows that the duration of wheezing is less than a first duration and / or the dominant frequency is less than a first frequency response value, the disease development path of the subject in the future period is output as stable, and a second type of suggestion for the subject's stable condition is also output.

5. The system according to claim 1, characterized in that, The output module (400) is configured as follows: When the subject's posture changes, and the time-frequency characteristic data of lung breath sounds within at least one respiratory cycle shows that the duration of wheezing is less than a first duration and / or the dominant frequency is less than a first frequency response value, the disease progression path of the subject in the future period is output as improved condition, and a third type of suggestion for the improvement of the subject's condition is also output.

6. The system according to claim 1, characterized in that, The output module (400) is configured as follows: When the subject's posture changes, and the time-frequency characteristic data of the lung breath sounds in at least one respiratory cycle shows that the duration of wheezing exceeds a first duration and / or the dominant frequency exceeds a first frequency response value, the duration of the acquisition module (100) acquiring the subject's lung breath sounds is extended.

7. The system according to claim 6, characterized in that, Extending the duration of the acquisition module (100) acquiring the lung breath sounds of the subject increases the number of respiratory cycles during the detection.

Images