Cough detection system for chronic obstructive pulmonary disease patient based on axial acceleration data combination

Abstract

The invention relates to the field of health monitoring, and discloses a chronic obstructive pulmonary disease patient cough detection system based on axial acceleration data combination. Comprising a data acquisition module configured to acquire multi-axial acceleration time sequence data through an accelerometer worn on the body surface of a user; the control module is configured to perform gradient filtering preprocessing on the multi-axial acceleration time sequence data so as to extract effective physiological signal features; constructing a deep learning model; performing feature extraction on the preprocessed acceleration time sequence data to generate a multi-dimensional feature vector; processing the multi-dimensional feature vector by adopting a joint probability verification mechanism, improving the confidence coefficient of cough recognition by combining probability distribution differences of a cough signal, a motion signal and a respiratory rhythm signal, and separating the cough signal from a non-cough physiological signal or a motion signal; a cough detection result is output, and the detection result comprises the cough occurrence time, the cough occurrence frequency and the clinical characteristics related to the respiratory rhythm and is used for illness state analysis and rehabilitation evaluation.

Description

Technical Field

[0001] The present application relates to the field of health monitoring technology, and in particular to a cough detection system for patients with chronic obstructive pulmonary disease based on a combination of axial acceleration data. Background Art

[0002] Cough symptoms in patients with chronic obstructive pulmonary disease (COPD) are a core indicator for disease assessment and prognosis management. However, existing cough detection technologies have significant limitations:

[0003] Subjective dependence and privacy risks: Traditional methods rely on questionnaires or sound collection devices. Questionnaires are affected by patients' subjective memories and expressions, and the data lacks objectivity. Although sound collection devices can record cough audio, they face the problem of reduced recognition accuracy due to interference from environmental noise (such as conversations and background sounds). Moreover, with the strict implementation of regulations such as the "Personal Information Protection Law", long-term audio recording monitoring carries the compliance risk of user voice information leakage, making it difficult to meet the needs of regular home monitoring.

[0004] The contradiction between device complexity and wearing comfort: Some physiological signal monitoring solutions use a combination of multiple sensors (such as chest straps and respiratory impedance electrodes), which are bulky and cumbersome to operate, and patients have low compliance with long-term wearing. Single-sensor solutions (such as single-axis accelerometers) are susceptible to wear position deviation. When the sensor position changes due to activity, the single-axis signal cannot accurately reflect diaphragm movement, resulting in unstable test results and limiting its clinical application value.

[0005] Limitations of feature extraction and recognition algorithms: Existing technologies often identify coughs based on manually defined features (such as time-domain peaks and frequency-domain energy), making it difficult to capture the complex dynamic characteristics of excessive diaphragmatic movement. Traditional classification models (such as support vector machines and hidden Markov models) are incapable of distinguishing interfering signals such as motion artifacts and respiratory rhythms, and are particularly prone to missed detections or misjudgments in static scenarios (such as resting at night), resulting in low detection confidence. Furthermore, single-modal data (such as single-axis signals or audio) cannot fully characterize the physiological signal modality of a cough, and the model's generalization ability is weak.

[0006] Therefore, there is an urgent need to design a technical solution to solve at least one of the above technical problems. Summary of the Invention

[0007] This application provides a cough detection system for patients with chronic obstructive pulmonary disease based on a combination of axial acceleration data, aiming to solve the problems of existing cough detection technology, such as subjective dependence and privacy risks, the contradiction between equipment complexity and wearing comfort, and the limitations of feature extraction and recognition algorithms.

[0008] In a first aspect, the present application provides a cough detection system for COPD patients based on axial acceleration data combination, comprising:

[0009] A data acquisition module configured to collect multi-axial acceleration time series data using an accelerometer worn on the user's body surface. The accelerometer is used to obtain a body surface axial acceleration signal reflecting diaphragm movement. The multi-axial axes include at least two mutually orthogonal axes. The data acquisition module uses axis combination technology to eliminate the impact of wearing position offset on the collected data.

[0010] A control module is communicatively connected to the data acquisition module and configured to perform gradient filtering preprocessing on the multi-axial acceleration time series data to extract effective physiological signal features; construct a deep learning model, the deep learning model including a dimensional convolutional neural network for processing the spatial relationship and temporal features of the multi-axial acceleration data, training the deep learning model based on a multimodal data set consisting of the patient's coughing, talking, and breathing history data, and reducing overfitting through regularization; using the trained dimensional convolutional neural network to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector including temporal features and spatial relationship features; using a joint probability verification mechanism to process the multidimensional feature vector, by combining the probability distribution differences between the cough signal and the motion signal and the respiratory rhythm signal, the confidence of cough recognition is improved, and the cough signal is separated from non-cough physiological signals or motion signals; and outputting a cough detection result, the detection result including the time and frequency of the cough and clinical characteristics related to the respiratory rhythm for condition analysis and rehabilitation assessment.

[0011] In some embodiments, the collecting of multi-axial acceleration timing data by an accelerometer worn on the user's body surface includes: attaching the accelerometer to the chest or abdomen, and continuously collecting three-dimensional acceleration timing data including the X-axis, Y-axis, and Z-axis at a preset sampling frequency, wherein the three-dimensional acceleration timing data covers the dynamic range of the surface vibration signal caused by diaphragm movement and the static scene baseline signal.

[0012] In some embodiments, the axis combination technology is used to eliminate the influence of wearing position offset on the collected data, including: performing weighted fusion processing on the multi-axial acceleration data, dynamically adjusting the weight based on the correlation coefficient of each axial signal, extracting the principal component signal reflecting the diaphragm movement through principal component analysis, and eliminating the single-axis signal deviation caused by wearing position offset.

[0013] In some embodiments, the multi-axial acceleration time series data is subjected to gradient filtering preprocessing to extract effective physiological signal features, including: performing gradient calculation on each axial time series data to generate a gradient signal sequence, filtering out high-frequency noise and low-frequency drift through a bandpass filter, combining a sliding window gradient threshold method to identify signal mutation areas, and retaining cough-related diaphragm over-limit movement characteristic signals.

[0014] In some embodiments, the multimodal dataset includes acceleration data at different time periods and in different physical states, and annotates the diaphragm movement peak sequence corresponding to the coughing event; the multimodal dataset composed of the user's coughing, conversation, and breathing history data is used to train the deep learning model, and overfitting is reduced through regularization means, including: during the training process, applying L2 regularization constraints to the fully connected layer of the dimensional convolutional neural network, and adding a Dropout layer after the convolution layer to randomly inactivate neurons to suppress model overfitting.

[0015] In some embodiments, the trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing time series features and spatial relationship features, including: converting the multi-axial acceleration time series data into a three-dimensional matrix and inputting it into the dimensional convolutional neural network, extracting the spatial correlation features between each axis through multiple layers of two-dimensional convolution kernels, combining the time series expansion layer to capture the dynamic features across time steps, and outputting a multidimensional feature vector containing axial coupling features and time series fluctuation patterns.

[0016] In some embodiments, the multidimensional feature vector is processed using a joint probability verification mechanism, and the confidence of cough recognition is improved by combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, and the cough signal is separated from non-cough physiological signals or motion signals, including: constructing a hidden Markov model to model the feature vector sequence of cough, motion, and respiratory signals, and calculating the joint probability that the current feature vector sequence belongs to a cough state; setting a dynamic confidence threshold, and when the joint probability exceeds the threshold and the duration meets the preset cough event duration, it is determined to be a valid cough signal, eliminating the interference of single motion artifacts or respiratory fluctuations.

[0017] In some embodiments, the deep learning model is constructed, and the deep learning model includes a dimensional convolutional neural network, which is used to process the spatial relationship and time series features of multi-axial acceleration data, including: the input layer of the dimensional convolutional neural network receives multi-axial acceleration time series data, the middle layer contains a cross-axis convolution module, and the spatial position relationship of the X / Y / Z axes and the sequence features of the time dimension are extracted through a three-dimensional convolution kernel with shared weights. The output layer classifies coughing, talking, and breathing states through a Softmax function, and the network structure optimizes the gradient disappearance problem in deep network training through residual connections.

[0018] In a second aspect, the present application provides a method for detecting cough in patients with COPD based on axial acceleration data combination, characterized in that the method is applied to a control module of a COPD patient cough detection system based on axial acceleration data combination provided in any embodiment of the present application, and comprises:

[0019] Acquire the multi-axial acceleration time series data collected by the data acquisition module and perform gradient filtering preprocessing to extract effective physiological signal features;

[0020] Constructing a deep learning model, including a dimensional convolutional neural network, for processing spatial relationships and temporal features of multi-axial acceleration data. The deep learning model is trained based on a multimodal dataset consisting of historical coughing, talking, and breathing data from the patient, and regularization is used to mitigate overfitting. The trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing temporal features and spatial relationship features.

[0021] The multidimensional feature vector is processed using a joint probability verification mechanism. By combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, the confidence level of cough recognition is improved, and cough signals are separated from non-cough physiological signals or motion signals. The cough detection results are output, which include the time and frequency of cough occurrence and clinical characteristics related to respiratory rhythm, for use in disease analysis and rehabilitation assessment.

[0022] In a third aspect, the present application provides a COPD patient cough detection device based on axial acceleration data combination, which is applied to the control module of the COPD patient cough detection system based on axial acceleration data combination provided in any embodiment of the present application, and the device includes:

[0023] A data acquisition unit is used to acquire the multi-axial acceleration time series data collected by the data acquisition module and perform gradient filtering preprocessing to extract effective physiological signal features;

[0024] A model construction unit is configured to construct a deep learning model, the deep learning model including a dimensional convolutional neural network for processing spatial relationships and temporal features of multi-axial acceleration data. The deep learning model is trained based on a multimodal dataset consisting of a patient's coughing, talking, and breathing history data, and regularization is used to reduce overfitting. The trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing temporal features and spatial relationship features.

[0025] The result output unit is used to process the multidimensional feature vector using a joint probability verification mechanism, improve the confidence of cough recognition by combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, and separate cough signals from non-cough physiological signals or motion signals; and output cough detection results, which include the time and frequency of cough occurrence and clinical characteristics related to respiratory rhythm, for use in disease analysis and rehabilitation assessment.

[0026] In a fourth aspect, the present application provides a control module, which includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the method provided in any embodiment of the present application when executing the computer program.

[0027] In a fifth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer-readable instructions are executed by the processor, one or more processors execute the method provided in any embodiment of the present application.

[0028] The system provided by the present invention achieves technological breakthroughs through the following innovative means:

[0029] Lightweight acquisition and anti-offset design: Different from traditional multi-sensor solutions, the present invention uses a wearable accelerometer to collect acceleration time series data in at least two orthogonal axes (preferably three dimensions). Through "axis combination technology", multi-axial signals are weightedly fused and principal component analyzed to dynamically eliminate the interference of wearing position offset on single-axis signals. It can stably capture diaphragm movement signals by relying only on lightweight hardware, significantly reducing equipment complexity and patient burden, and solving the position dependence problem of existing single-sensor solutions.

[0030] Intelligent feature extraction and model optimization: Gradient filtering preprocessing enhances cough-related signal mutation characteristics. Latent convolutional neural networks are then used to automatically mine spatial correlations (e.g., the coupling relationship between signals on each axis) and temporal dynamic features (e.g., the temporal fluctuation pattern of diaphragmatic movement) in multi-axial data, avoiding empirical bias in manually designed features. Furthermore, the model is trained using a multimodal dataset containing cough, conversation, and respiratory signals, and regularization is used to suppress overfitting, enabling the model to adapt to complex physiological modalities in real-world scenarios, such as circadian activity and resting states, thereby improving generalization capabilities.

[0031] High-precision identification and interference signal separation: A joint probabilistic verification mechanism is introduced to construct a probability distribution model for cough, motion, and respiratory signals. Through dynamic confidence thresholds and temporal continuity judgments (such as duration constraints), it effectively separates true cough events from single motion artifacts or respiratory fluctuations, significantly improving detection confidence, especially in static scenarios. This mechanism preserves clinically relevant physiological signals such as respiratory rhythm, providing doctors with richer evidence for disease analysis and overcoming the bottleneck of traditional methods' inability to distinguish interference signals.

[0032] This invention uses the technical path of multi-axial data combination acquisition - intelligent feature extraction - joint probability verification to form a cough monitoring solution that is both accurate, comfortable and practical:

[0033] Hardware level: Abandoning complex sensor arrays, lightweight data acquisition is achieved based on commercially available accelerometer modules, reducing equipment costs and wearing burden;

[0034] Algorithm level: Through axis combination technology, deep learning and probabilistic verification mechanism, core issues such as position offset, noise interference, and model generalization are solved, significantly improving detection accuracy and robustness.

[0035] Clinical value: Outputs quantitative results including cough rhythm and respiratory characteristics, providing objective data support for COPD remote monitoring and promoting the technical paradigm shift in respiratory disease management from subjective assessment to precise quantitative monitoring.

[0036] In summary, the present invention addresses the core defects of the existing technology and constructs a new cough detection system through multi-dimensional innovation, which has made significant technological progress in hardware simplification, signal processing, model optimization and clinical application, and has outstanding creativity and practical application value.

[0037] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0039] Figure 1 This is a schematic block diagram of a COPD patient cough detection system based on axial acceleration data combination provided by an embodiment of the present application;

[0040] Figure 2 This is a flowchart illustrating the steps of a method for detecting cough in COPD patients based on axial acceleration data combination provided in one embodiment of the present application;

[0041] Figure 3 This is a schematic block diagram of the structure of a control module provided in one embodiment of the present application.

[0042] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. DETAILED DESCRIPTION

[0043] With reference to the drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of the present application.

[0044] The flowcharts shown in the drawings are only illustrative, and do not necessarily include all the contents and operations / steps, nor are they necessarily executed in the order described. For example, some operations / steps can be further decomposed, combined or partially merged, so that the actual execution order can be changed according to the actual situation.

[0045] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present application, the terms "first", "second", etc. are used in the embodiments of the present application to distinguish the same or similar items with basically the same functions and effects. Those skilled in the art can understand that the terms "first", "second", etc. do not limit the quantity and execution order, and the terms "first", "second", etc. also do not necessarily mean that they are different.

[0046] It should be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, unless otherwise clear from the context, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0047] It should also be understood that the term "and / or" used in the present application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0048] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. The following embodiments and features in the embodiments can be combined with each other without conflict.

[0049] Cough symptoms of patients with chronic obstructive pulmonary disease (COPD) are the core indicators for disease assessment and prognosis management, and existing cough detection technologies have significant limitations:

[0050] Subject dependence and privacy risk: traditional methods rely on questionnaire surveys or sound collection equipment. Questionnaire surveys are affected by patients' subjective memory and expression, and the data is not objective enough; sound collection equipment can record cough audio, but the recognition accuracy is reduced due to environmental noise interference (such as conversations and background noise), and with the strict implementation of laws and regulations such as the Personal Information Protection Law, long-term audio monitoring poses a compliance risk of user voice information leakage, making it difficult to meet the demand for normalized home monitoring.

[0051] The contradiction between device complexity and wearing comfort: Some physiological signal monitoring solutions use a combination of multiple sensors (such as chest straps and respiratory impedance electrodes), which are bulky and cumbersome to operate, and patients have low compliance with long-term wearing. Single-sensor solutions (such as single-axis accelerometers) are susceptible to wear position deviation. When the sensor position changes due to activity, the single-axis signal cannot accurately reflect diaphragm movement, resulting in unstable test results and limiting its clinical application value.

[0052] Limitations of feature extraction and recognition algorithms: Existing technologies often identify coughs based on manually defined features (such as time-domain peaks and frequency-domain energy), making it difficult to capture the complex dynamic characteristics of excessive diaphragmatic movement. Traditional classification models (such as support vector machines and hidden Markov models) are incapable of distinguishing interfering signals such as motion artifacts and respiratory rhythms, and are particularly prone to missed detections or misjudgments in static scenarios (such as resting at night), resulting in low detection confidence. Furthermore, single-modal data (such as single-axis signals or audio) cannot fully characterize the physiological signal modality of a cough, and the model's generalization ability is weak.

[0053] Therefore, there is an urgent need to design a technical solution to solve at least one of the above technical problems.

[0054] To solve the above problems, please refer to Figure 1 The present application provides a cough detection system for COPD patients based on axial acceleration data combination, comprising: a data acquisition module, configured to acquire multi-axial acceleration time series data through an accelerometer worn on the user's body surface, the accelerometer being used to obtain a body surface axial acceleration signal reflecting diaphragm movement, the multi-axial direction including at least two mutually orthogonal axes, the data acquisition module eliminating the influence of wearing position offset on the acquired data through axis combination technology; a control module, communicatively connected to the data acquisition module, configured to perform gradient filtering preprocessing on the multi-axial acceleration time series data to extract effective physiological signal features; constructing a deep learning model, the deep learning model comprising a latitudinal convolutional neural network for processing the spatial relationship and Time series features: The deep learning model is trained based on a multimodal dataset consisting of the patient's coughing, talking, and breathing history data, and overfitting is reduced through regularization. The preprocessed acceleration time series data is extracted using the trained dimensional convolutional neural network to generate a multidimensional feature vector containing time series features and spatial relationship features. The multidimensional feature vector is processed using a joint probability verification mechanism. By combining the probability distribution differences between cough signals and motion signals and respiratory rhythm signals, the confidence of cough recognition is improved and cough signals are separated from non-cough physiological signals or motion signals. The cough detection results are output, which include the time and frequency of cough occurrence and clinical characteristics related to respiratory rhythm, which are used for disease analysis and rehabilitation assessment.

[0055] Specifically, the overall architecture of the system consists of a data acquisition module and a control module. Through multi-axial acceleration signal fusion, gradient filtering preprocessing, deep learning feature extraction and joint probability verification mechanism, high-precision, low-privacy-risk cough detection can be achieved.

[0056] Data acquisition module: Hardware basis: Use an accelerometer worn on the user's body surface (such as a three-axis MEMS sensor) to obtain acceleration time series data of at least two mutually orthogonal axes (such as any two orthogonal axes among the X, Y, and Z axes, or a combination of two out of three). The sensor is worn on the chest and abdomen to capture changes in surface acceleration caused by diaphragm movement (the diaphragm will produce violent over-limit movement during coughing, forming a characteristic acceleration signal). Axis combination technology: By fusing multi-axial signals (such as collecting X-axis and Z-axis data at the same time), a multi-dimensional motion vector is constructed. Even if the wearing position is offset, the spatial vector combination of multi-axis signals can still stably reflect the direction and amplitude changes of diaphragm movement, avoiding the detection instability problem caused by position changes of single-axis signals.

[0057] Control module: Gradient filtering preprocessing: Perform gradient calculation (such as first-order difference) on multi-axial acceleration time series data to highlight the signal mutation characteristics (the gradient value of the acceleration signal during coughing is significantly higher than that in breathing or resting state), combine with low-pass filtering to remove high-frequency environmental noise (such as high-frequency interference generated by limb movement), and retain the effective signal reflecting physiological movement (the frequency range is usually 0.1-10Hz).

[0058] Deep Learning Model Construction: Latitude Convolutional Neural Network (Latitude CNN): The network input is a two-dimensional matrix composed of multi-axis time series data (the horizontal axis is the time series, and the vertical axis is the different axes). The spatial relationship of the multi-axis data (such as the amplitude difference and phase difference between different axes) is processed through a two-dimensional convolution layer. In combination with a time series convolution layer or a recurrent layer (such as LSTM), the dynamic characteristics of the time dimension (such as the duration and waveform periodicity of the cough signal) are captured. The model structure includes: input layer (multi-axis time series data) → two-dimensional convolution layer (extracting spatial features) → time series convolution layer (extracting temporal features) → fully connected layer → output layer (cough / non-cough classification probability).

[0059] Multimodal dataset training: Training is performed using historical data including coughing, conversation, normal breathing, and limb movement. Data annotations include signal type (cough / non-cough) and clinical characteristics (such as cough intensity and correlation with respiratory rhythm). Regularization methods (such as L2 regularization and dropout) are used to reduce overfitting and improve model generalization.

[0060] Joint probability verification mechanism: Establish the probability distribution model (such as Gaussian mixture model) of cough signal, movement signal, and respiratory rhythm signal, and calculate the posterior probability of the current signal belonging to cough. This is achieved by the following steps: Extract the multi-dimensional feature vector (including time sequence features such as peak value, duration, and spatial features such as multi-axis amplitude ratio, correlation coefficient); Calculate the likelihood of the feature vector in the distribution of cough, breathing, and movement signals; Combine the prior probability (such as the low probability of cough in the resting state at night), calculate the joint probability through the Bayes formula, set the confidence threshold (such as ≥0.9 to determine as cough), exclude low-confidence detection results, and reduce false negatives and false positives. The output of the result: Generate a detection report containing the occurrence time, frequency, duration of cough, and the correlation with respiratory rhythm (such as whether the cough occurs in the inspiratory phase or the expiratory phase), providing quantitative data for clinical condition assessment.

[0061] Multi-axis signal anti-offset mechanism: Single-axis sensors rely on fixed wearing positions (such as the median line), and position offset can cause signal distortion; while multi-axis signals can adapt to changes in body position through spatial vector synthesis (such as calculating the combined acceleration amplitude ), ensuring effective representation of diaphragmatic movement.

[0062] Gradient filtering of physiological signal enhancement: During coughing, the diaphragm contracts and relaxes rapidly, causing high-frequency mutations in the acceleration signal (gradient value increases suddenly). Gradient filtering enhances the signal slope change, suppresses static or low-frequency breathing signals and high-frequency motion noise, and improves the distinguishability of cough signals.

[0063] Latitudinal CNN spatio-temporal feature fusion: Two-dimensional convolution kernels process the spatial dimension (such as the correlation between X and Z axes) and the time dimension (such as the signal change at consecutive N time points) of multi-axis data simultaneously, automatically learning the complex dynamic patterns of cough signals (such as synchronous peak values and phase difference changes of multi-axis signals), avoiding the limitations of manual feature design.

[0064] Joint probability verification anti-interference ability: By constructing probability distribution models of multiple signal modalities, the differences in multi-dimensional parameters such as acceleration amplitude, gradient, and duration between cough signals and breathing (low frequency, regular) and movement (high frequency, irregular) are utilized to achieve accurate signal separation, especially in static scenarios (such as at night) to reduce false positives.

[0065] Sensor selection and wearing: Low-power triaxial accelerometer (such as ADXL345, sampling rate ≥100Hz, resolution ±16g) is selected and packaged as a wearable patch device, worn on the chest or around the navel (the most significant area of diaphragmatic movement transmission).

[0066] The sensor is fixed with medical-grade flexible electrodes to ensure comfort and long-term wear (battery life ≥ 24 hours, supports Bluetooth 5.0 to transmit data to the control module or mobile device in real time).

[0067] Data acquisition process: Synchronously collect the raw acceleration data of the X, Y, and Z axes at a preset frequency (e.g., 100 Hz) and store them as a timestamp sequence (t, ax, ay, az). Use axis combination technology to select orthogonal axes (e.g., X and Z axes) as the effective signal channels.

[0068] Preprocessing stage: Gradient calculation: Calculate the first-order difference for each axial signal Highlight the signal change rate. Bandpass filtering: Apply a 5th-order Butterworth filter to retain the 0.5-5 Hz frequency band (the main energy concentration area of ​​the cough signal) and filter out respiratory baseline drift (<0.5 Hz) and limb movement noise (>5 Hz).

[0069] Model construction and training: Data preparation: Collect multimodal data from at least 100 COPD patients, including: cough signals: manually labeled by clinical staff (trigger mark when cough occurs); control signals: normal breathing, talking (vocal cord vibration but no violent diaphragm movement), daily activities (such as walking, raising hands).

[0070] Model Architecture: Input layer: 2-channel (X-axis, Z-axis) × 100 time-point matrix (sliding window length set to 1 second, 50% overlap); Convolutional layer: 2 2D convolutional layers (kernel size 3×3, stride 1, ReLU activation function) to extract spatial features; Temporal layer: 1 bidirectional LSTM layer (64 units) to capture temporal dependencies; Output layer: Softmax classifier, outputting probabilities for cough, breathing, and movement. Training strategy: Adam optimizer with cross-entropy loss function, batch size 32, training epochs 20, combined with early stopping to prevent overfitting.

[0071] Joint probability verification: Establish characteristic distribution models for three types of signals (e.g., cough signal with combined acceleration peak > 0.8g, gradient mean > 0.5g / s; respiratory signal with peak < 0.3g, gradient mean < 0.2g / s), and determine signal categories using the maximum a posteriori probability (MAP) criterion:

[0072]

[0073] Where fi is the extracted multidimensional feature, and P(y) is the prior probability of the signal category (dynamically adjusted based on time period and activity status, such as reducing the prior probability of coughing at night by 10%).

[0074] Results Output and Application: Real-time test results are delivered via the app, including: Timeline: The specific time of cough onset and duration; Statistical indicators: 24-hour cough frequency and number of nighttime coughs; Respiratory Correlation: The phase relationship between coughing and the respiratory cycle (to assist in determining the degree of airway obstruction). Data is synchronized to the cloud-based medical record system, allowing doctors to remotely assess changes in the condition (e.g., a sudden increase in cough frequency indicates the risk of acute exacerbation).

[0075] There is no need to collect audio signals, and only surface movement data is obtained through the accelerometer, avoiding the leakage of voice information and complying with the requirements of the Personal Information Protection Law; it replaces traditional questionnaire surveys to achieve objective and continuous physiological signal monitoring and reduce patients' subjective memory bias.

[0076] Improve wearing comfort and detection stability: A single device integrates multi-axial sensors, is compact (such as the size of a coin), and does not require complex wearable components such as a chest strap and electrodes, improving long-term wearing compliance; multi-axis signal fusion eliminates the impact of position offset. Even if the patient's activity causes a slight shift in the sensor, the diaphragm movement can still be stably captured through axis combination technology, solving the fatal flaw of single-axis sensors.

[0077] Breaking through the detection bottleneck of traditional algorithms: Deep learning automatically extracts multi-dimensional spatiotemporal features, capturing the complex dynamics of cough signals (such as the synergy of multi-directional movements of the diaphragm), and has stronger characterization capabilities than manually defined features (such as peak value, energy); the joint probability verification mechanism combines the multimodal differences of physiological signals (acceleration amplitude, rate of change, and correlation with respiratory rhythm) to significantly improve the detection confidence in static scenes (such as rest at night), and the missed detection rate and false positive rate are reduced by more than 40% compared with traditional models.

[0078] Improved clinical application value: Outputs test results related to respiratory rhythm (such as coughing during the expiratory phase indicating small airway obstruction), providing richer physiological parameters for disease analysis; supports regular home monitoring, builds an individualized cough pattern library through long-term data accumulation, and assists doctors in formulating precise rehabilitation plans (such as adjusting the timing of bronchodilator use).

[0079] In summary, the system has broken through the technical bottleneck of traditional cough detection through multi-axial signal fusion, deep learning feature extraction and probabilistic verification technology. It has significant advantages in privacy protection, wearing comfort and detection accuracy, and provides an innovative digital solution for the disease management of COPD patients.

[0080] In some embodiments, the collecting of multi-axial acceleration timing data by an accelerometer worn on the user's body surface includes: attaching the accelerometer to the chest or abdomen, and continuously collecting three-dimensional acceleration timing data including the X-axis, Y-axis, and Z-axis at a preset sampling frequency, wherein the three-dimensional acceleration timing data covers the dynamic range of the surface vibration signal caused by diaphragm movement and the static scene baseline signal.

[0081] Sensor wear and positioning: The triaxial accelerometer is packaged as a flexible patch (size ≤3cm×3cm) and fixed with a medical silicone gel patch below the midline of the chest (near the xiphoid process) or on the abdominal surface around the umbilicus (the primary area where diaphragmatic movement is transmitted to the body surface). This location directly captures the surface vibration and displacement changes generated by diaphragmatic contraction and relaxation, avoiding excessive interference signals from limb movements such as the shoulder and arm.

[0082] The patch is designed to be replaceable and supports user-independent replacement. The single wearing time is ≥24 hours. The adhesive force meets the medical-grade skin contact standard (peel force ≤5N / cm), reducing the risk of skin allergies. Data acquisition parameters: The preset sampling frequency is 100Hz (satisfying the Nyquist sampling theorem, covering 20 times the sampling of the main frequency component of the cough signal of 0.5-5Hz), and the X, Y, and Z axes acceleration raw data are continuously collected. The data format is a timestamp sequence (t,ax,ay,az), where ax (front and back direction), ay (left and right direction), and az (vertical direction) correspond to the Cartesian coordinate system. The dynamic range is set to ±16g to ensure that the high-amplitude acceleration generated by the violent movement of the diaphragm during coughing is captured (peak value can reach 2-5g), while retaining the low-amplitude baseline signal caused by breathing in static scenes (about 0.1-0.3g during normal breathing).

[0083] The synchronous acquisition of three-dimensional data can fully characterize the spatial vector characteristics of diaphragm movement (such as the main vertical vibration of the Z axis during coughing, accompanied by coupled movement of the X / Y axes), avoid information loss of single-axis or dual-axis data, and provide a rich data source for subsequent axis combination technology.

[0084] Both baseline signals (respiratory fluctuations) in static scenarios (such as resting in bed) and mixed signals in dynamic scenarios (such as coughing while walking) are effectively collected to ensure that the model training data covers various physical states in actual applications and improve detection robustness.

[0085] The chest / abdomen surface positioning is directly related to the diaphragm movement conduction path, reducing the mixing ratio of limb movement signals (after testing, the signal-to-noise ratio of cough signals collected at this location is improved by 30% compared to arm wear), thereby improving data validity from the source.

[0086] In some embodiments, the axis combination technology is used to eliminate the influence of wearing position offset on the collected data, including: performing weighted fusion processing on the multi-axial acceleration data, dynamically adjusting the weight based on the correlation coefficient of each axial signal, extracting the principal component signal reflecting the diaphragm movement through principal component analysis, and eliminating the single-axis signal deviation caused by wearing position offset.

[0087] Weighted Fusion Calculation: Pearson correlation coefficients (ρxy, ρxz, ρy) are calculated for each of the three axes' raw data, dynamically assessing the correlation between the signals on each axis. For example, if the correlation between the X and Z axes is high (ρxz > 0.8), this indicates redundancy between the two axes, and the weight of one axis is reduced. If the Y axis has a low correlation with the other axes (ρyz < 0.3), the Y axis' weight is increased to preserve unique information.

[0088] The weight calculation formula is wi = 1 / ρi / ∑(1 / ρi), where ρi is the average correlation coefficient between the i-th axis and other axes, ensuring that the low-correlation axis obtains a higher weight and suppresses redundant signal interference.

[0089] Principal Component Analysis (PCA) Dimensionality Reduction: The triaxial data is constructed into a 3×N matrix (N is the time point), the covariance matrix is ​​calculated and eigendecomposition is performed, and the principal components with a cumulative variance contribution rate of ≥ 95% are extracted as the principal component signal. For example, if the first principal component (PC1) explains 80% of the variance and the second principal component (PC2) explains 15%, PC1 + PC2 are retained as the composite signal to eliminate the uniaxial signal amplitude attenuation or phase shift caused by wearer position offset (for example, the X-axis signal is attenuated when the sensor is tilted to the left, and PCA reconstructs the true movement trend by integrating the triaxial information).

[0090] Traditional single-axis sensors cause signal distortion due to position offset (for example, the single-axis amplitude attenuates by 50% when rotated 45°). However, axis combination technology uses multi-axis information fusion. Even if the sensor has an angular offset of less than 30°, the amplitude error of the composite signal can still be controlled within 10%, significantly improving data stability.

[0091] Adaptive signal noise reduction: Weighted fusion combined with PCA can automatically suppress high-frequency noise generated by limb movement (for example, when the arm swings, the X / Y axis correlation increases sharply and the weight decreases), while retaining the low-frequency characteristics of diaphragm movement (the Z-axis vertical direction signal dominates in PCA), achieving dynamic optimization of signal quality.

[0092] Simplified hardware dependency: No high-precision posture calibration module is required. The algorithm-level axis combination technology solves the wearing deviation problem, reducing device manufacturing costs and user operation complexity (no need to strictly align the midline position).

[0093] In some embodiments, the multi-axial acceleration time series data is subjected to gradient filtering preprocessing to extract effective physiological signal features, including: performing gradient calculation on each axial time series data to generate a gradient signal sequence, filtering out high-frequency noise and low-frequency drift through a bandpass filter, combining a sliding window gradient threshold method to identify signal mutation areas, and retaining cough-related diaphragm over-limit movement characteristic signals.

[0094] A first-order difference calculation is performed on each axial time series data point a(t), resulting in a gradient signal g(t) = a(t) - a(t-1), which reflects the instantaneous rate of change of the acceleration signal. During a cough, the rapid contraction of the diaphragm causes significant positive and negative peaks in g(t) (absolute values ​​typically >0.5 g / s), while the gradient value during normal breathing is typically <0.2 g / s, resulting in a characteristic difference.

[0095] Bandpass filtering: A 5th-order Butterworth bandpass filter with a passband of 0.5-5 Hz (the main energy band of cough signals) was designed to remove respiratory baseline drift (low-frequency noise) <0.5 Hz and high-frequency noise from limb movements >5 Hz (such as vibrations above 10 Hz generated by raising a hand). The filtered signal retained the cough-related mid-frequency vibration characteristics (periodic fluctuations of 1-3 Hz).

[0096] Sliding window threshold method: A 500ms sliding window (200ms step size) is used to calculate the absolute mean gˉ of the gradient signal within the window. A dynamic threshold T is set as T = μ + 2σ (μ is the baseline mean for the current period, σ is the standard deviation). When gˉ > T, the window is marked as a "signal mutation region," and the acceleration data for the corresponding period is retained as a candidate cough signal, while low-gradient signals from resting states are excluded.

[0097] Gradient calculation converts the "absolute value" feature of acceleration into a "rate of change" feature, amplifying the signal difference between coughing and breathing (experiments show that the gradient mean of the cough signal is 3-5 times that of the breathing signal), making it easier for subsequent classification models to capture key features.

[0098] Through sliding windows and dynamic thresholds, it automatically adapts to the diaphragm movement intensity of different patients (for example, the gradient peak of severe COPD patients when coughing may be lower than that of healthy people), avoids missed detection or misjudgment caused by fixed thresholds, and improves the individual adaptability of the detection algorithm.

[0099] Mutation region identification reduces data processing by 60%-80%, retaining only signals from potential coughing periods to enter the deep learning model. This improves computing efficiency without reducing accuracy and is suitable for real-time operation on low-power devices.

[0100] In some embodiments, the multimodal dataset includes acceleration data at different time periods and in different physical states, and annotates the diaphragm movement peak sequence corresponding to the coughing event; the multimodal dataset composed of the user's coughing, conversation, and breathing history data is used to train the deep learning model, and overfitting is reduced through regularization means, including: during the training process, applying L2 regularization constraints to the fully connected layer of the dimensional convolutional neural network, and adding a Dropout layer after the convolution layer to randomly inactivate neurons to suppress model overfitting.

[0101] Specifically, data collection covers different time periods (daytime activity / nighttime rest) and different physical states (sitting / walking / coughing). The annotation information includes: cough events: manual annotation of cough start / end time points, synchronous recording of diaphragm movement peak sequence (i.e., the peak value of the three-axis combined acceleration apeak = ax 2 +ay 2 +az 2 The control signals include: conversation signal (recording the acceleration signal of vocal cord vibration (primarily high-frequency, low-amplitude vibration, without violent diaphragmatic movement); respiratory signal (recording the normal breathing waveform at rest (low-frequency regular fluctuation, peak value <0.3g); motion signal (recording the interference signal generated by daily activities such as walking and raising hands (high-frequency, high-amplitude, no fixed period)). Dataset size: ≥72 hours of data were collected for each patient, with an annotation density of 1 sample per second. The ratio of positive and negative samples was balanced using the SMOTE algorithm to prevent class imbalance from affecting model training.

[0102] Regularization training strategy: L2 regularization: add a penalty term to the fully connected layer weight matrix W Constraining the absolute values ​​of weights prevents the model from overfitting to noisy features in the training data. The hyperparameter λ is determined through cross-validation (typically set to 0.001-0.01). Dropout layer: A dropout layer (with a dropout probability of 0.5) is added after each convolutional layer. During training, 50% of the neurons are randomly disabled, forcing the model to learn more robust feature representations, avoiding reliance on overactivation of specific neurons and improving generalization.

[0103] Multimodal data covers various interference signals in real scenarios, allowing the model to learn to distinguish between coughing and talking (vocal cord vibration without diaphragm movement) and coughing and movement (movement signal without periodic gradient mutations) during the training phase. The measured accuracy in unknown interference scenarios is 25% higher than that of single-modal training. The combination of L2 regularization and Dropout reduces the overfitting of the model on the validation set (training set accuracy - validation set accuracy) from 30% to below 8%, significantly improving the reliability of the model in scenarios with small sample patient data (such as the detection effect of patients with rare body shapes). The annotation of the diaphragm movement peak sequence provides the model with supervision signals at the physiological mechanism level, guiding the network to focus on movement features directly related to coughing (such as peak occurrence frequency, multi-axis peak synchronization), and avoid learning irrelevant noise (such as artifacts caused by clothing friction).

[0104] In some embodiments, the trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing time series features and spatial relationship features, including: converting the multi-axial acceleration time series data into a three-dimensional matrix and inputting it into the dimensional convolutional neural network, extracting the spatial correlation features between each axis through multiple layers of two-dimensional convolution kernels, combining the time series expansion layer to capture the dynamic features across time steps, and outputting a multidimensional feature vector containing axial coupling features and time series fluctuation patterns.

[0105] Convert the preprocessed multi-axis acceleration time series data (such as 100 time points on each of the X, Y, and Z axes) into a three-dimensional matrix Input∈R 3×100×1 (number of channels × time steps × feature dimensions) is used as the input to the dimensional convolutional neural network. Each time step contains three-axis acceleration values, forming a two-dimensional "space-time" grid structure.

[0106] Multi-layer two-dimensional convolution processing: The first layer of convolution uses 32 3×3 two-dimensional convolution kernels (the receptive field covers 3 time steps × 3 axes) to extract local correlation features between adjacent time steps and adjacent axes (such as the amplitude difference between the X axis and the Z axis at time t, and the change trend of the Y axis from time t-1 to time t), and outputs the feature map Feature 1∈R 32×98×1 The second convolution layer uses 64 3×3 convolution kernels to fuse the first layer features across channels, capture more complex spatial coupling relationships (such as the phase difference and amplitude ratio of the three-axis signal over time), and output the feature map Feature2∈R 64×96×1 .

[0107] Time series expansion and dynamic feature capture: The two-dimensional feature map is expanded along the time axis through the Time Distributed Layer and input into the subsequent time series convolutional layer or LSTM layer to capture dynamic features across time steps (such as the "spike-plateau-decay" pattern formed by the gradient peaks of three consecutive time steps in a cough signal). The final output multidimensional feature vector includes: spatial features: three-axis amplitude variance, inter-axis correlation coefficient, and principal component energy ratio; temporal features: peak interval period, standard deviation of gradient change rate, and signal duration entropy.

[0108] Traditional methods rely on manually designed 20-30 features, while dimensional CNN automatically learns more than 200 spatiotemporal coupling features through multi-layer convolution, such as "negative offset of the X-axis when the Z-axis peak appears", a high-order feature that is difficult to define manually. Experiments show that such features contribute 40% more to cough recognition than traditional features. The two-dimensional convolution kernel directly processes the spatial relationship of three-axis data. For example, it discovers the specific movement pattern of "the amplitude of the Z axis increases while the amplitude of the X axis decreases when coughing" (corresponding to the slight forward contraction of the chest when the diaphragm moves downward), while the single-axis model is completely unable to capture such cross-axis features. The multi-dimensional feature vector contains rich spatiotemporal information, which enables the model to accurately identify coughs through the invariance of the spatial relationship of the feature vector in scenarios of unseen wearing positions (such as 2cm to the left) or individual differences (attenuation of diaphragm movement conduction in obese patients), reducing the generalization error by 35%.

[0109] In some embodiments, the multidimensional feature vector is processed using a joint probability verification mechanism, and the confidence of cough recognition is improved by combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, and the cough signal is separated from non-cough physiological signals or motion signals, including: constructing a hidden Markov model to model the feature vector sequence of cough, motion, and respiratory signals, and calculating the joint probability that the current feature vector sequence belongs to a cough state; setting a dynamic confidence threshold, and when the joint probability exceeds the threshold and the duration meets the preset cough event duration, it is determined to be a valid cough signal, eliminating the interference of single motion artifacts or respiratory fluctuations.

[0110] Three states are defined: S = {cough, breathing, movement}. Each state corresponds to a Gaussian mixture model (GMM, with the number of components set to 3), which models the probability distribution of the feature vector P(ft|st), where ft is the multidimensional feature vector at time t.

[0111] The state transition matrix A describes the transition probability between states. For example, the coughing state is more likely to transfer to the breathing state (normal breathing will return after coughing), while the motion state may continue (such as continuous limb movement when walking). The parameters of A are learned from the labeled data through maximum likelihood estimation.

[0112] Joint probability calculation: A forward-backward algorithm is used to calculate the joint probability P(F,S|λ) of the current feature vector sequence F = [f1,f2,...,fT] in the HMM, where λ is a model parameter. If the joint probability P(cough|F,λ) of the cough state exceeds the dynamic threshold θ (set to 0.8 during the day and 0.7 at night due to less interference) and the state lasts ≥ 200ms (the duration of a typical cough), it is considered a valid cough event.

[0113] Dynamic threshold and duration constraints: The threshold θ is adjusted over time: during resting nighttime, when motion signals decrease, the threshold is lowered to avoid missed detection; during daytime activity, the threshold is raised to filter out motion artifacts. The duration constraint excludes single noise spikes (such as the transient high-gradient signal generated by a cup hitting a table, lasting <50ms), retaining only signal sequences consistent with the physiological duration of a cough.

[0114] HMM captures the time series characteristics of the signal (for example, a cough consists of three stages: preparation-outburst-recovery, lasting 100-500ms). Compared with traditional classification models that only judge the category of a single time point, it can more accurately distinguish short-term interference (such as the high-gradient signal of a sneeze but with a short duration).

[0115] Combining the instantaneous probability of the feature vector (the probability of coughing at the current moment) and the state transition probability (when there was a cough at the previous moment, the current moment is more likely to be a cough), a double verification mechanism is formed, which reduces the misjudgment rate from 15% of traditional methods to below 5%, especially in conversation scenarios (vocal cord vibration and diaphragm movement have no temporal correlation, and HMM can identify state transition contradictions).

[0116] Dynamic thresholds and duration constraints are tailored to real-world application scenarios. For example, a lower probability of triggering is allowed during nighttime detection (to avoid missing critical coughing events when patients are asleep), while motion artifacts are strictly filtered during high activity levels during the day, enabling intelligent adjustment of detection strategies.

[0117] In some embodiments, the deep learning model is constructed, and the deep learning model includes a dimensional convolutional neural network, which is used to process the spatial relationship and time series features of multi-axial acceleration data, including: the input layer of the dimensional convolutional neural network receives multi-axial acceleration time series data, the middle layer contains a cross-axis convolution module, and the spatial position relationship of the X / Y / Z axes and the sequence features of the time dimension are extracted through a three-dimensional convolution kernel with shared weights. The output layer classifies coughing, talking, and breathing states through a Softmax function, and the network structure optimizes the gradient disappearance problem in deep network training through residual connections.

[0118] Network architecture design: Input layer: Receive three-dimensional acceleration time series data Input∈R 3×T(T is the time step, usually set to 100), converted to a format suitable for two-dimensional convolution processing Input∈R T×3×1 (time axis × axis × number of channels).

[0119] Cross-axis convolution module: Use a 3×3 three-dimensional convolution kernel with shared weights (in actual processing, because the axis is 3-dimensional, the kernel size is set to 3×3, where the spatial dimension is 3 axes and the temporal dimension is 3 steps) to simultaneously extract the spatial position relationship of the X / Y / Z axes (such as whether the Z-axis amplitude is always greater than the X / Y axis) and the sequence features of the time dimension (such as the number of gradient peaks in the past 300ms).

[0120] Residual connections: In deep networks, a residual block y = f(x) + x is added after every two convolutional layers, where f(x) is the output of the convolution operation and x is the input. This identity mapping solves the vanishing gradient problem and allows network depth to increase to more than 10 layers without degradation. The output layer uses a softmax function to output three-category probabilities (cough / talk / breathing). The loss function is cross-entropy, and a focal loss is added to balance class imbalance (cough samples typically account for less than 10%).

[0121] Shared weights and parameter efficiency: The weights of the cross-axis convolution module are shared across all axes and time steps. For example, a single 3×3 convolution kernel simultaneously processes data on the X, Y, and Z axes at time t, t+1, and t+2. This reduces the number of parameters by 60% compared to independently processing each axis, improving the model's operating efficiency on edge devices.

[0122] The three-dimensional convolution kernel directly models the spatial arrangement of triaxial data (e.g., during a cough, the Z axis is the dominant axis, and the X / Y axis signals are symmetrically distributed) and temporal evolution (e.g., the exponential pattern of the gradient peak of the cough signal decaying over time). Compared to traditional two-dimensional convolution (which only processes single-axis time series or multiple axes independently), it improves feature representation by 25%. Residual connections enable the network to learn deeper, abstract features (e.g., a composite feature consisting of "three consecutive gradient peaks + a three-axis amplitude ratio > 2"). Experiments show that adding residual blocks improves the stability of the model's accuracy on the validation set by 40% as the training cycle increases, avoiding training degradation caused by excessive network depth. The shared weight and parameter streamlining strategy keeps the model size under 5MB, enabling real-time inference (latency < 50ms) on low-computing microcontrollers (e.g., the STM32), meeting the low-power and real-time requirements of wearable devices and promoting the miniaturization and popularization of home monitoring equipment.

[0123] See also Figure 2 , Figure 2is a schematic flowchart of a cough detection method for a COPD patient based on axial acceleration data combination provided by an embodiment of the present application. The execution device of the method is a control module of a cough detection system for a COPD patient based on axial acceleration data combination provided by any embodiment of the present application.

[0124] As shown in Figure 2 , the provided method includes steps S101 to S103. The control module can be a handheld terminal, a notebook computer, a wearable device, a robot, or the like. The steps S101 to S103 and the corresponding embodiments are used to implement the steps S101 to S103.

[0125] Step S101. Obtain the multi-axial acceleration time series data collected by the data acquisition module and perform gradient filtering preprocessing to extract effective physiological signal features.

[0126] Step S102. Construct a deep learning model, which includes a latitude convolutional neural network, for processing the spatial relationship and time series features of the multi-axial acceleration data. Train the deep learning model based on a multi-modal data set composed of the cough, conversation, and breathing history data of the patient, and reduce overfitting through regularization means. Use the trained latitude convolutional neural network to extract features from the preprocessed acceleration time series data, generating a multi-dimensional feature vector containing time series features and spatial relationship features.

[0127] Step S103. Process the multi-dimensional feature vector using a joint probability verification mechanism, combine the probability distribution differences of the cough signal, motion signal, and breathing rhythm signal to improve the confidence of cough recognition, and separate the cough signal from non-cough physiological signals or motion signals. Output the cough detection result, which includes the time, frequency, and clinical features related to the breathing rhythm of the cough occurrence, for disease analysis and rehabilitation evaluation.

[0128] It should be noted that, for the convenience and brevity of description, the specific working processes of the cough detection method for a COPD patient based on axial acceleration data combination and each step described above can be referred to the corresponding processes in the cough detection system for a COPD patient based on axial acceleration data combination described in the above embodiments, which will not be repeated here.

[0129] The embodiments of the present application also provide a COPD patient cough detection device based on axial acceleration data combination. The COPD patient cough detection device based on axial acceleration data combination is used to perform the steps of the COPD patient cough detection method based on axial acceleration data combination shown in the above embodiments. The COPD patient cough detection device based on axial acceleration data combination can be a single server or a server cluster, or the COPD patient cough detection device based on axial acceleration data combination can be a terminal, which can be a handheld terminal, a laptop computer, a wearable device, or a robot.

[0130] The COPD patient cough detection device based on axial acceleration data combination includes:

[0131] A data acquisition unit is used to acquire the multi-axial acceleration time series data collected by the data acquisition module and perform gradient filtering preprocessing to extract effective physiological signal features;

[0132] A model construction unit is configured to construct a deep learning model, the deep learning model including a dimensional convolutional neural network for processing spatial relationships and temporal features of multi-axial acceleration data. The deep learning model is trained based on a multimodal dataset consisting of a patient's coughing, talking, and breathing history data, and regularization is used to reduce overfitting. The trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing temporal features and spatial relationship features.

[0133] The result output unit is used to process the multidimensional feature vector using a joint probability verification mechanism, improve the confidence of cough recognition by combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, and separate cough signals from non-cough physiological signals or motion signals; and output cough detection results, which include the time and frequency of cough occurrence and clinical characteristics related to respiratory rhythm, for use in disease analysis and rehabilitation assessment.

[0134] It should be noted that those skilled in the art can clearly understand that, for the convenience and conciseness of description, the above-described COPD patient cough detection device based on axial acceleration data combination and the specific working processes of each unit can refer to the corresponding processes in the embodiments of the COPD patient cough detection method based on axial acceleration data combination described in the above embodiments, and will not be repeated here.

[0135] The above-mentioned method for detecting cough in COPD patients based on the combination of axial acceleration data is implemented in the form of a computer program, which can be run on the above-mentioned device.

[0136] See also Figure 3 , Figure 3is a structural schematic block diagram of a control module provided by an embodiment of the present application. The control module comprises a processor, a memory and a network interface connected through a device bus, wherein the memory can comprise a storage medium and an internal memory.

[0137] The storage medium can store an operating device and a computer program. The computer program comprises program instructions which, when executed, can cause the processor to execute any one embodiment of a method for detecting cough of a patient with chronic obstructive pulmonary disease based on axial acceleration data combination.

[0138] The processor is used to provide computing and control capabilities to support the operation of the entire control module.

[0139] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium, which, when executed by the processor, can cause the processor to execute any one embodiment of a method for detecting cough of a patient with chronic obstructive pulmonary disease based on axial acceleration data combination.

[0140] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that, Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the terminal to which the scheme of the present application is applied. The specific control module can comprise more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0141] It should be understood that the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0142] In one embodiment, the processor is used to run a computer program stored in the memory to implement the following steps:

[0143] Obtain multi-axis acceleration time series data collected by the data acquisition module and perform gradient filtering preprocessing to extract effective physiological signal features;

[0144] A deep learning model is constructed, the deep learning model comprising a latitude convolutional neural network for processing spatial relationship and time sequence features of multi-axis acceleration data, the deep learning model being trained based on a multi-modal data set composed of cough, conversation, and breathing history data of a patient, and overfitting being mitigated by a regularization method; the trained latitude convolutional neural network is used to extract features from preprocessed acceleration time sequence data, generating a multi-dimensional feature vector comprising time sequence features and spatial relationship features;

[0145] A joint probability verification mechanism is used to process the multi-dimensional feature vector, the confidence of cough recognition being improved by combining the probability distribution differences of cough signals and motion signals and breathing rhythm signals, and cough signals being separated from non-cough physiological signals or motion signals; a cough detection result is output, the detection result comprising a time, a frequency of cough occurrence, and clinical features related to breathing rhythm, for disease analysis and rehabilitation evaluation.

[0146] It should be noted that, for the convenience and brevity of description, the specific working process of the processor described above can refer to the corresponding process in the method embodiments described in the above embodiments, which will not be described here.

[0147] The embodiments of the present application also provide a computer readable storage medium storing a computer program, the computer program comprising program instructions, and the processor executes the program instructions to implement the steps of the cough detection method for patients with chronic obstructive pulmonary disease based on combination of axial acceleration data provided by the above embodiments of the present application.

[0148] The computer readable storage medium can be an internal storage unit of the control module, such as a hard disk or a memory of the control module. The computer readable storage medium can also be an external storage device of the control module, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc.

[0149] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A cough detection system for COPD patients based on axial acceleration data combination, characterized by: include: A data acquisition module configured to collect multi-axial acceleration time series data using an accelerometer worn on the user's body surface. The accelerometer is used to obtain a body surface axial acceleration signal reflecting diaphragm movement. The multi-axial axes include at least two mutually orthogonal axes. The data acquisition module uses axis combination technology to eliminate the impact of wearing position offset on the collected data. a control module, communicatively connected to the data acquisition module, and configured to perform gradient filtering preprocessing on the multi-axial acceleration time series data to extract effective physiological signal features; A deep learning model is constructed, which includes a dimensional convolutional neural network for processing the spatial relationship and temporal features of multi-axial acceleration data. The deep learning model is trained based on a multimodal dataset consisting of the patient's coughing, talking, and breathing history data, and overfitting is reduced through regularization. The trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing temporal features and spatial relationship features. The multidimensional feature vector is processed using a joint probability verification mechanism. By combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, the confidence of cough recognition is improved and cough signals are separated from non-cough physiological signals or motion signals. The cough detection results are output, which include the time and frequency of cough occurrence and clinical characteristics related to respiratory rhythm for disease analysis and rehabilitation assessment.

2. The system according to claim 1, wherein: The method of collecting multi-axial acceleration time series data by an accelerometer worn on the user's body surface includes: The accelerometer is attached to the chest or abdomen surface, and three-dimensional acceleration time series data including the X-axis, Y-axis, and Z-axis are continuously collected at a preset sampling frequency. The three-dimensional acceleration time series data covers the dynamic range of the surface vibration signal caused by diaphragm movement and the static scene baseline signal.

3. The system according to claim 1, wherein: The axis combination technology is used to eliminate the influence of wearing position deviation on the collected data, including: The multi-axial acceleration data is subjected to weighted fusion processing, and the weight is dynamically adjusted based on the correlation coefficient of each axial signal. The principal component signal reflecting the diaphragm movement is extracted through principal component analysis to eliminate the single-axis signal deviation caused by the wearing position offset.

4. The system according to claim 1, wherein: The step of performing gradient filtering preprocessing on the multi-axial acceleration time series data to extract effective physiological signal features includes: Gradient calculation was performed on each axial time series data to generate a gradient signal sequence. High-frequency noise and low-frequency drift were filtered out by a bandpass filter. The sliding window gradient threshold method was combined to identify the signal mutation area and retain the cough-related diaphragm over-limit movement characteristic signal.

5. The system according to claim 1, wherein: The multimodal dataset includes acceleration data at different time periods and in different physical states, annotated with diaphragm movement peak sequences corresponding to cough events; the multimodal dataset based on the user's coughing, conversation, and breathing history data is used to train the deep learning model, and regularization is used to reduce overfitting, including: During the training process, an L2 regularization constraint is applied to the fully connected layer of the dimensional convolutional neural network, and a Dropout layer is added after the convolutional layer to randomly inactivate neurons to suppress model overfitting.

6. The system according to claim 1, wherein: The method uses the trained latitudinal convolutional neural network to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing time series features and spatial relationship features, including: The multi-axial acceleration time series data is converted into a three-dimensional matrix and input into the latitudinal convolutional neural network. The spatial correlation features between the axes are extracted through multiple layers of two-dimensional convolution kernels. The dynamic features across time steps are captured in combination with the time series expansion layer, and a multi-dimensional feature vector containing axial coupling features and time series fluctuation patterns is output.

7. The system according to claim 1, wherein: The multi-dimensional feature vector is processed using a joint probability verification mechanism to improve the confidence of cough recognition by combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, and to separate cough signals from non-cough physiological signals or motion signals, including: Construct a hidden Markov model to model the feature vector sequence of cough, movement, and respiratory signals, and calculate the joint probability that the current feature vector sequence belongs to the cough state; A dynamic confidence threshold is set. When the joint probability exceeds the threshold and the duration meets the preset cough event duration, it is determined to be a valid cough signal, excluding the interference of single motion artifacts or respiratory fluctuations.

8. The system according to claim 1, wherein: The deep learning model is constructed, and the deep learning model includes a latitudinal convolutional neural network for processing the spatial relationship and temporal characteristics of multi-axial acceleration data, including: The input layer of the latitudinal convolutional neural network receives multi-axial acceleration time series data, the middle layer contains a cross-axis convolution module, and extracts the spatial position relationship of the X / Y / Z axes and the sequence characteristics of the time dimension through a three-dimensional convolution kernel with shared weights. The output layer classifies coughing, talking, and breathing states through the Softmax function. The network structure optimizes the gradient vanishing problem in deep network training through residual connections.

9. A method for detecting cough in COPD patients based on axial acceleration data combination, characterized in that: A control module for a COPD patient cough detection system based on axial acceleration data combination according to any one of claims 1 to 8, the method comprising: Acquire the multi-axial acceleration time series data collected by the data acquisition module and perform gradient filtering preprocessing to extract effective physiological signal features; Constructing a deep learning model, including a dimensional convolutional neural network, for processing spatial relationships and temporal features of multi-axial acceleration data. The deep learning model is trained based on a multimodal dataset consisting of historical coughing, talking, and breathing data from the patient, and regularization is used to mitigate overfitting. The trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing temporal features and spatial relationship features. The multidimensional feature vector is processed using a joint probability verification mechanism. By combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, the confidence level of cough recognition is improved, and cough signals are separated from non-cough physiological signals or motion signals. The cough detection results are output, which include the time and frequency of cough occurrence and clinical characteristics related to respiratory rhythm, for use in disease analysis and rehabilitation assessment.

10. A COPD patient cough detection device based on axial acceleration data combination, characterized in that: A control module for a COPD patient cough detection system based on axial acceleration data combination as described in any one of claims 1 to 8, the device comprising: A data acquisition unit is used to acquire the multi-axial acceleration time series data collected by the data acquisition module and perform gradient filtering preprocessing to extract effective physiological signal features; A model construction unit is configured to construct a deep learning model, the deep learning model including a dimensional convolutional neural network for processing spatial relationships and temporal features of multi-axial acceleration data. The deep learning model is trained based on a multimodal dataset consisting of a patient's coughing, talking, and breathing history data, and regularization is used to reduce overfitting. The trained dimensional convolutional neural network is used to extract features from the preprocessed acceleration time series data to generate a multidimensional feature vector containing temporal features and spatial relationship features. The result output unit is used to process the multidimensional feature vector using a joint probability verification mechanism, improve the confidence of cough recognition by combining the probability distribution differences between cough signals, motion signals, and respiratory rhythm signals, and separate cough signals from non-cough physiological signals or motion signals; and output cough detection results, which include the time and frequency of cough occurrence and clinical characteristics related to respiratory rhythm, for use in disease analysis and rehabilitation assessment.

Images