Auxiliary expectoration control method and device

Abstract

The invention provides an expectoration auxiliary control method and device. The method comprises the steps that audio signals of breathing sounds and cough sounds at multiple positions of the chest of a patient are collected through an audio sensor array, and a standardized audio feature data set is obtained through noise reduction, segmentation and feature extraction; a dynamic weighted graph model is constructed, and a random edge sampling algorithm and a parallel batch processing dynamic algorithm are combined to analyze and obtain a sputum viscosity index and a sputum distribution position map; based on the result, mapping the vibration parameter space into an unweighted disk diagram, and obtaining a personalized vibration treatment scheme by using a shortest path algorithm; in combination with the real-time breathing cycle of the patient, working parameters of the sound wave vibrator and the negative pressure suction device are synchronously controlled, and a coordinated and consistent multi-mode treatment execution instruction is obtained; and collecting real-time feedback data in treatment, and dynamically adjusting working parameters through reinforcement learning to obtain an optimized expectoration adjuvant therapy scheme. Intelligent adjustment can be achieved based on the real-time breathing state and sputum characteristics of the patient, the expectoration efficiency is improved, and discomfort of the patient is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical equipment, and in particular to a method and device for assisting in controlling sputum coughing, which is suitable for sputum clearing treatment of patients with respiratory diseases. Background Art

[0002] Patients with respiratory diseases such as chronic obstructive pulmonary disease (COPD), bronchiectasis, and cystic fibrosis often face difficulty clearing sputum, which can lead to airway obstruction, increased risk of infection, and decreased lung function. Clearing sputum from the respiratory tract is a crucial step in clinical treatment and plays a crucial role in patient recovery and prognosis.

[0003] Traditional sputum removal techniques primarily include manual back patting and simple mechanical vibration devices. Manual back patting requires a healthcare professional to tap the patient's chest and back at a specific frequency and force to loosen and guide sputum. Simple mechanical vibration devices, on the other hand, utilize fixed-frequency vibrations to aid in sputum clearance. These methods are widely used in clinical practice.

[0004] Existing advanced cough-assistance systems use pre-set programs to control the vibration frequency and negative pressure suction device, and can adjust the operating mode based on simple preset parameters. These systems generally include a vibration generator, a negative pressure suction device, and a basic control module. They can provide more stable vibration output than manual back patting and assist in sputum discharge through negative pressure.

[0005] However, existing technologies have obvious shortcomings: first, they are unable to make intelligent adjustments based on the patient's real-time respiratory status and sputum characteristics, and the setting of treatment parameters mainly relies on experience and judgment; second, they lack the ability to accurately analyze sputum characteristics (such as viscosity and distribution location), making it difficult to achieve precise treatment; third, there is insufficient coordination between the various components of the system, and it is impossible to form an optimal treatment strategy, resulting in limited coughing efficiency and possible discomfort to patients. Summary of the Invention

[0006] The purpose of the present invention is to overcome the above-mentioned deficiencies in the prior art and provide a method and device for assisting in the control of expectoration, which can make intelligent adjustments based on the patient's real-time respiratory status and sputum characteristics, accurately analyze the sputum characteristics, and form an optimal treatment strategy through the synergy between the various components of the system, thereby improving the efficiency of expectoration and reducing patient discomfort.

[0007] To achieve the above object, the present invention provides a method for assisting in controlling sputum cough, comprising:

[0008] The audio sensor array collects audio signals of breathing and coughing sounds from multiple locations in the patient's chest cavity, performs noise reduction, segmentation, and feature extraction on the audio signals, and obtains a standardized audio feature dataset.

[0009] A dynamic weighted graph model was constructed based on a standardized audio feature dataset. A random edge sampling algorithm was applied to estimate the average degree of the dynamic weighted graph model. A parallel batch dynamic algorithm was used to analyze the spectral feature changes of cough sounds in the standardized audio feature dataset to obtain the sputum viscosity index and sputum distribution location map.

[0010] Based on the sputum viscosity index and sputum distribution location map, the vibration parameter space is mapped into an unweighted disk map. The shortest path algorithm is used to calculate the optimal vibration frequency, amplitude and time parameter combination to obtain a personalized vibration treatment plan.

[0011] According to the personalized vibration treatment plan, combined with the patient's real-time breathing cycle, the working parameters of the sonic vibrator and the negative pressure suction device are synchronously controlled to obtain coordinated and consistent multimodal treatment execution instructions;

[0012] Real-time feedback data including audio change data and patient physiological indicators are collected during the treatment process, compared and analyzed with multimodal treatment execution instructions, and working parameters are dynamically adjusted based on reinforcement learning algorithms to obtain an optimized cough-assisted treatment plan.

[0013] Preferably, the audio signal is subjected to noise reduction, segmentation, and feature extraction processing to obtain a standardized audio feature dataset, including:

[0014] Based on the audio signals of breathing and coughing sounds, adaptive filtering algorithms and wavelet transform are applied to filter out environmental noise and medical equipment interference to obtain audio signals with high signal-to-noise ratio;

[0015] Based on high signal-to-noise ratio audio signals, time-domain-frequency domain feature analysis and energy threshold detection algorithm are used to identify respiratory cycles, cough events, and quiet periods, and obtain time-stamped audio event sequences.

[0016] Based on the time-stamped audio event sequence, the spectral energy distribution, MFCC coefficients, harmonic structure and time domain features are extracted. The most discriminative feature set is screened through the feature selection algorithm to obtain a standardized audio feature dataset.

[0017] Preferably, a random edge sampling algorithm is applied to estimate the average degree of the dynamic weighted graph model, including:

[0018] Based on the dynamic weighted graph model, the importance sampling technique is used to randomly select edge samples from the edge set, so that the probability of edge selection is proportional to the weight of the edge sample, and a random edge sample set is obtained;

[0019] Based on a random edge sample set, the local degree of the nodes at both ends of each edge is calculated, and the global average degree of the dynamic weighted graph model is estimated by summation operation to obtain the sputum distribution density index;

[0020] Based on the sputum distribution density index, the lungs are divided into multiple anatomical regions and the average degree in each region is calculated to obtain the sputum density distribution maps of different lung regions.

[0021] Preferably, the spectral feature changes of the cough sound of the standardized audio feature dataset are analyzed by a parallel batch dynamic algorithm, including:

[0022] Based on the sputum density distribution map, the cough sounds in the standardized audio feature dataset are grouped by lung region and corresponding sub-graphs are constructed to obtain the acoustic characteristic representation of each region;

[0023] To represent the acoustic characteristics of each region, a parallel computing architecture is used to simultaneously extract the spectral energy distribution, harmonic components, and time-frequency characteristics to obtain a multi-dimensional acoustic feature vector.

[0024] Based on the multi-dimensional acoustic feature vector, a mapping relationship between the spectrum characteristics and the sputum viscosity is established, and the viscosity index of each region is calculated to obtain the sputum viscosity index;

[0025] The sputum density distribution map and sputum viscosity index are integrated with the patient's lung anatomical structure parameters to generate an accurate sputum distribution location map through a three-dimensional spatial interpolation algorithm.

[0026] Preferably, the shortest path algorithm is applied to calculate the optimal vibration frequency, amplitude and time parameter combination, including:

[0027] Based on the sputum viscosity index and sputum distribution location map, the frequency parameter and amplitude parameter are mapped to plane coordinates, and the time parameter is mapped to the size of the point to obtain an unweighted disk map;

[0028] Based on the unweighted disk graph, the distance between two points is defined as the weighted combination of the inverse of the effect gain and the patient's tolerance. The Dijkstra algorithm is applied to search for the shortest path from the initial state to the target state to obtain the optimal parameter path.

[0029] Based on the optimal parameter path, the dynamic programming algorithm is applied to design the curve of vibration parameter changes over time. Combined with the patient tolerance model, the time allocation is optimized to obtain a personalized vibration treatment plan.

[0030] Preferably, according to the personalized vibration therapy plan, combined with the patient's real-time breathing cycle, the working parameters of the sonic vibrator and the negative pressure suction device are synchronously controlled to obtain a coordinated multimodal treatment execution instruction, including:

[0031] Based on the personalized vibration therapy plan, the vibration parameters are converted into driving signals, the working frequency, amplitude and phase of the acoustic vibrator are controlled, and the vibrator driving instructions are obtained;

[0032] Based on the personalized vibration therapy plan and the patient's real-time breathing cycle, the negative pressure intensity, suction time window and pressure change curve coordinated with the vibration parameters are calculated to obtain the negative pressure control instructions;

[0033] By monitoring the patient's respiratory phase in real time, using a prediction algorithm to estimate the optimal intervention time, the start and stop sequence of the sonic vibrator and the negative pressure suction device is controlled to obtain a respiratory synchronization signal;

[0034] Based on the vibrator drive instructions, negative pressure control instructions and breathing synchronization signals, an execution time table for collaboration between devices is generated to obtain a device collaborative control strategy;

[0035] Integrate vibrator drive instructions, negative pressure control instructions, respiratory synchronization signals and equipment collaborative control strategies to ensure the coordinated cooperation of acoustic vibration therapy and negative pressure suction device therapy, and obtain coordinated and consistent multimodal treatment execution instructions.

[0036] Preferably, by real-time monitoring of the patient's respiratory phase, using a prediction algorithm to estimate the optimal intervention time, and controlling the start and stop timing of the acoustic vibrator and the negative pressure suction device, a respiratory synchronization signal is obtained, including:

[0037] By fusing the chest and abdominal motion sensors, airflow sensors, and sound sensors to collect respiratory signals, the respiratory cycle is divided into four phases: early inspiration, late inspiration, early expiration, and late expiration, and the respiratory phase recognition results are obtained.

[0038] Based on the respiratory phase recognition results, wavelet transform and peak detection algorithm are applied to capture the phase transition point. The phase time point of the next cycle is predicted by combining the patterns of the last 3-5 respiratory cycles to obtain the respiratory phase prediction result.

[0039] Based on the respiratory phase prediction results, the optimal intervention time is determined, with vibration therapy starting at the early stage of exhalation and negative pressure suction implemented in the middle and late stages of exhalation. The time difference between vibration and negative pressure is designed to obtain a respiratory synchronization signal.

[0040] Preferably, the working parameters are dynamically adjusted based on the reinforcement learning algorithm to obtain an optimized cough auxiliary treatment plan, including:

[0041] The patient's physiological state, sputum characteristics, and treatment progress are used as the state space, and the working parameters and treatment mode selection are used as the action space. A reward function is constructed based on the short-term sputum cough effect index to obtain a formal representation of the reinforcement learning problem.

[0042] Based on the formal representation of reinforcement learning problems, the policy gradient method and neural network parameterization strategy are used to optimize the treatment strategy through gradient estimation and parameter update to obtain the initial policy network;

[0043] Maintain an experience replay buffer to store treatment process data in real-time feedback data, prioritize sampling important experiences based on the size of the temporal difference error, perform batch learning and parameter updates, continuously update the initial strategy network parameters, and obtain an optimized strategy network;

[0044] Design an exploration strategy with parameter noise and entropy regularization, conduct planned exploration within a safe parameter range, adjust the exploration degree according to the uncertainty estimated by the value function, and obtain an exploration-exploitation balance strategy;

[0045] Based on the optimal strategy network and exploration-exploitation balance strategy, the recommended parameters are calculated according to the current patient status in the real-time feedback data. Safety constraint checking and smoothing processing are applied to dynamically adjust the working parameters to obtain the optimized cough-assisted treatment plan.

[0046] Preferably, a reward function is constructed based on the short-term sputum cough effect indicator, including:

[0047] Overlapping sliding windows were used to continuously analyze the treatment process data in the real-time feedback data. The degree of sputum loosening was assessed by the energy ratio of breath sounds, changes in sputum sounds, and improvement in sound conduction to obtain the sputum loosening index.

[0048] Based on acoustic event statistics, negative pressure system feedback, and cough effectiveness evaluation, the sputum output estimation index is obtained by weighted summation of acoustic event counts, negative pressure recovery estimation, and cough energy.

[0049] The patient comfort index is obtained by combining the fluctuation of physiological indicators, discomfort reaction detection and subjective discomfort score, and calculating the basic score minus the weighted influence of various discomfort factors;

[0050] Compare the gap between the current status and treatment goals in real-time feedback data, track changes in treatment status and target achievement in each lung region, evaluate the treatment progress rate per unit time, and obtain treatment progress indicators;

[0051] Based on the sputum looseness index, sputum output estimation index, patient comfort index and treatment progress index, a comprehensive effect score was constructed through weighted summation and multi-objective balance to obtain the short-term sputum cough effect index.

[0052] The present invention also provides a sputum coughing auxiliary control device, comprising:

[0053] An audio signal acquisition module is used to collect audio signals of breathing sounds and coughing sounds at multiple locations in the patient's chest cavity based on an audio sensor array, and to perform noise reduction, segmentation, and feature extraction on the audio signals to obtain a standardized audio feature dataset;

[0054] The sputum characteristic analysis module is used to build a dynamic weighted graph model based on a standardized audio feature dataset. The random edge sampling algorithm is used to estimate the average degree of the dynamic weighted graph model. The parallel batch dynamic algorithm is used to analyze the spectral characteristic changes of cough sounds in the standardized audio feature dataset to obtain the sputum viscosity index and sputum distribution location map.

[0055] The vibration scheme optimization module is used to map the vibration parameter space into an unweighted disk map based on the sputum viscosity index and sputum distribution location map, and apply the shortest path algorithm to calculate the optimal vibration frequency, amplitude and time parameter combination to obtain a personalized vibration treatment plan;

[0056] A multimodal control module is used to synchronously control the operating parameters of the acoustic vibrator and the negative pressure suction device according to the personalized vibration treatment plan and the patient's real-time breathing cycle, thereby obtaining coordinated and consistent multimodal treatment execution instructions;

[0057] The parameter optimization module is used to collect real-time feedback data during the treatment process, including audio change data and patient physiological indicators, compare and analyze them with multimodal treatment execution instructions, and dynamically adjust working parameters based on the reinforcement learning algorithm to obtain an optimized cough-assisted treatment plan.

[0058] The beneficial effects of the present invention are:

[0059] 1. This invention uses an efficient sputum distribution density estimation method based on random edge sampling. This optimizes the traditional sputum distribution analysis method, which requires a complete traversal of the graph structure, into a random edge sampling technique. By estimating the average degree of the graph, the sputum distribution density is quickly obtained, significantly reducing computational complexity and enabling the system to respond to changes in patient status in real time.

[0060] 2. This invention uses a parallel batch dynamic algorithm for sputum characteristic analysis technology, simultaneously analyzing multi-channel audio features and establishing a dynamic mapping relationship between sputum viscosity and audio features. This effectively improves the speed and accuracy of sputum characteristic analysis and provides a data foundation for precision treatment.

[0061] 3. This invention uses an unweighted disc graph-based vibration parameter optimization method to abstract the vibration parameter space into an unweighted disc graph model. Using a shortest path algorithm, the optimal parameter combination from the initial state to the target sputum loosening effect is solved, enabling precise calculation and dynamic adjustment of treatment parameters.

[0062] 4. The present invention establishes a closed-loop optimization system for collaborative control of multimodal treatment, establishes a collaborative control mechanism for vibration therapy and negative pressure suction, and continuously optimizes treatment parameters through a reinforcement learning algorithm, forming a complete closed-loop system of "perception-analysis-execution-evaluation-optimization", thereby realizing intelligent and personalized auxiliary treatment of cough and sputum. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the conventional technology, the following briefly introduces the drawings required for use in the embodiments or the conventional technology descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0064] Figure 1 A flowchart of a sputum cough assisting control method provided by an embodiment of the present invention;

[0065] Figure 2 This is a structural block diagram of the cough assisting control device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] To make the above-mentioned objects, features, and advantages of the present application more clearly understood, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the scope of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0067] It should be noted that when a component is referred to as being "fixed to" or "disposed on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "connected to" another component, it may be directly connected to the other component or there may be a central component at the same time. The terms "vertical", "horizontal", "upper", "lower", "left", "right" and similar expressions used in the specification of this application are for illustrative purposes only and do not represent the only implementation method.

[0068] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of such features. Throughout the description of this application, "plurality" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0069] In this application, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it can mean that the first feature is directly in contact with the second feature, or the first feature and the second feature are indirectly in contact through an intermediary. Furthermore, when a first feature is "above," "above," or "above" a second feature, it can mean that the first feature is directly above or diagonally above the second feature, or simply means that the first feature is higher in level than the second feature. When a first feature is "below," "below," or "below" a second feature, it can mean that the first feature is directly below or diagonally below the second feature, or simply means that the first feature is lower in level than the second feature.

[0070] Unless otherwise defined, all technical and scientific terms used in the specification of this application have the same meaning as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The term "and / or" used in the specification of this application includes any and all combinations of one or more of the relevant listed items.

[0071] See also Figure 1 , Figure 1 This is a flow chart of a method for assisting in controlling sputum coughing provided by an embodiment of the present invention. The method includes:

[0072] Step S101: collecting audio signals of breathing sounds and coughing sounds at multiple locations of the patient's chest cavity based on an audio sensor array, performing noise reduction, segmentation, and feature extraction on the audio signals to obtain a standardized audio feature dataset;

[0073] Step S102: constructing a dynamic weighted graph model based on the standardized audio feature dataset, applying a random edge sampling algorithm to estimate the average degree of the dynamic weighted graph model, and analyzing the spectral feature changes of the cough sounds in the standardized audio feature dataset using a parallel batch dynamic algorithm to obtain a sputum viscosity index and a sputum distribution location map;

[0074] Step S103: Based on the sputum viscosity index and the sputum distribution location map, the vibration parameter space is mapped into an unweighted disk map, and the shortest path algorithm is applied to calculate the optimal vibration frequency, amplitude and time parameter combination to obtain a personalized vibration treatment plan;

[0075] Step S104: According to the personalized vibration therapy plan, combined with the patient's real-time respiratory cycle, synchronously controlling the working parameters of the acoustic vibrator and the negative pressure suction device to obtain a coordinated and consistent multimodal therapy execution instruction;

[0076] Step S105: Collecting real-time feedback data including audio change data and patient physiological indicators during the treatment process, comparing and analyzing them with the multimodal treatment execution instructions, and dynamically adjusting the working parameters based on the reinforcement learning algorithm to obtain an optimized cough-assisted treatment plan.

[0077] In step S101, the audio signal is subjected to noise reduction, segmentation, and feature extraction processing to obtain a standardized audio feature dataset, including:

[0078] Step S201: Based on the audio signals of the breathing sounds and coughing sounds, an adaptive filtering algorithm and wavelet transform are applied to filter out environmental noise and medical equipment interference to obtain an audio signal with a high signal-to-noise ratio.

[0079] In this step, raw audio signal streams collected from multiple locations in the chest cavity are first received. These signals often contain various noise artifacts, such as ambient noise (e.g., air conditioning, conversations), medical equipment interference (e.g., monitor alarms, infusion pump operation), and noise generated by patient movement. To obtain high-quality breath and cough sounds, a two-stage noise reduction approach is employed. In the first stage, an adaptive filtering algorithm is applied, which automatically adjusts filter parameters based on the dynamic characteristics of the ambient noise. Specifically, a minimum mean square error (LMS) adaptive filter is used to dynamically estimate the noise model using ambient noise samples collected by a reference microphone. The estimated noise component is then subtracted from the main acquired signal. In the second stage, wavelet transform technology is applied for further noise reduction. The wavelet transform decomposes the signal into frequency components, and a soft thresholding method is used to process the wavelet coefficients, retaining large coefficients representing the useful signal and suppressing small coefficients representing noise. When reconstructing the signals, particular attention is paid to the key frequency range of breath and cough sounds (typically 100-2000Hz) to ensure signal integrity in these frequency bands. Through these two stages of processing, various types of noise interference can be effectively removed, while retaining the key acoustic features of breathing sounds and coughing sounds, and outputting audio signals with a high signal-to-noise ratio, laying the foundation for subsequent analysis.

[0080] Step S202: Based on the high signal-to-noise ratio audio signal, respiratory cycles, cough events and quiet periods are identified using time-domain-frequency domain feature analysis and energy threshold detection algorithm to obtain a time-stamped audio event sequence.

[0081] In this step, the continuous audio signal needs to be segmented into meaningful event segments, including respiratory cycles, cough events, and quiet periods. First, a short-term energy function is calculated in the time domain, and the signal energy within each window is calculated using a sliding window (typically 20-50ms). To identify respiratory cycles, the energy curve's periodic variation pattern is analyzed. Inhalation and exhalation typically manifest as energy peaks and troughs. An adaptive energy threshold algorithm dynamically adjusts the threshold based on the current signal's average energy, effectively addressing differences in respiratory intensity among patients. Simultaneously, the signal's spectral characteristics are analyzed in the frequency domain. Breathing sounds are typically concentrated in the 100-600Hz frequency band, while coughs have a wider spectral distribution (100-3000Hz) and higher energy. To identify cough events, combining time-domain energy mutation characteristics with frequency-domain feature analysis reveals that coughs typically manifest as short bursts of high energy with a specific spectral pattern. An event detection algorithm based on a hidden Markov model (HMM) is employed to identify different types of sound events using a pre-trained acoustic model. Quiet periods are also identified: signal segments where energy falls below a specific threshold and persists for a certain period of time. Finally, an accurate timestamp and type tag are added to each identified event to form a time-tagged audio event sequence, providing structured data input for the next step of feature extraction.

[0082] Step S203: Based on the time-tagged audio event sequence, extract the spectrum energy distribution, MFCC coefficients, harmonic structure and time domain features, and screen the most discriminative feature set through a feature selection algorithm to obtain the standardized audio feature data set.

[0083] In this step, in-depth feature extraction is performed on each labeled audio event to capture the multidimensional characteristics of the sound signal. First, the spectral energy distribution is calculated. The signal is converted to the frequency domain using a short-time Fourier transform (STFT), and the energy distribution of different frequency bands is calculated. Particular attention is paid to the energy ratios in the low-frequency band (100-300Hz), mid-frequency band (300-800Hz), and high-frequency band (800-2000Hz). These ratios are important for distinguishing different types of breath sounds and sputum conditions. Second, Mel-frequency cepstral coefficients (MFCCs) are extracted. These are acoustic features that mimic the characteristics of human auditory perception. Typically, 13-20 dimensional MFCC coefficients and their first- and second-order differences are extracted to capture the timbre characteristics of the sound. Third, the harmonic structure of the signal is analyzed, and the energy distribution of the fundamental frequency (F0) and its harmonic components is calculated. This harmonic structure is particularly valuable for distinguishing normal breath sounds from sputum-containing breath sounds. Fourth, time domain features are extracted, including zero-crossing rate, short-time energy, and energy entropy. These features can reflect the time-varying characteristics of the signal. After feature extraction, the challenge is that excessive feature dimensionality can lead to the "curse of dimensionality." To address this issue, feature selection algorithms, including the mutual information-based maximum relevance minimum redundancy (mRMR) algorithm and recursive feature elimination (RFE), are applied to select the most discriminative feature subset. Finally, the selected features are normalized, typically using Z-score or Min-Max normalization, to ensure comparability of features of different dimensions. This series of processes ultimately yields a standardized audio feature dataset, providing high-quality input data for subsequent sputum characterization analysis.

[0084] In step S102, a random edge sampling algorithm is applied to estimate the average degree of the dynamic weighted graph model, including:

[0085] Step S301: Based on the dynamic weighted graph model, edge samples are randomly selected from the edge set using importance sampling technology, so that the selection probability of the edge is proportional to the weight of the edge sample, and a random edge sample set is obtained.

[0086] In this step, a dynamic weighted graph model is first constructed based on the standardized audio feature dataset. In this graph model, the node set V represents the acoustic features of different acquisition locations, the edge set E represents the connection relationship between nodes, and the weight set W represents the connection strength. The traditional method needs to traverse all edges to analyze the graph structure, and the computational complexity is , which is inefficient when the number of nodes is large. To improve computational efficiency, this paper innovatively applies importance sampling technology to perform random edge sampling. Specifically, instead of selecting edges uniformly and randomly, the sampling probability distribution is designed based on the weight of the edge, so that edges with larger weights (indicating stronger acoustic correlation) have a higher probability of being sampled. The sampling probability calculation formula is ,in It's the edge This importance-based sampling strategy ensures that key edges are prioritized, preserving the graph's key structural features even with a low sampling ratio. Typically, 10%-20% of the total number of edges is sampled, significantly reducing computational effort while maintaining estimation accuracy. In practical implementations, efficient algorithms such as acceptance-rejection sampling or the alias method are used for sampling to minimize the computational overhead of the sampling process itself. This yields a significantly reduced random edge sample set that retains key structural information, laying the foundation for subsequent average degree estimation.

[0087] Step S302: Based on the random edge sample set, the local degrees of the nodes at both ends of each edge are calculated, and the global average degree of the dynamic weighted graph model is estimated by summation operation to obtain a sputum distribution density index.

[0088] In this step, the random edge sample set obtained in the previous step is used to efficiently estimate the average degree of the graph and then infer the sputum distribution density. , calculate the local degree of its two end nodes u and v and In a weighted graph, the degree of a node is defined as the sum of the weights of all edges connected to the node. Based on the local degree information of the sampled edges, the unbiased estimation formula is applied to calculate the global average degree of the graph , where m is the number of sampled edges, are the two end nodes of the i-th sampled edge. The theoretical basis of this estimation method is that the average degree of the nodes connected by the randomly selected edges is proportional to the average degree weighted by the node degree. In order to further improve the estimation accuracy, a stratified sampling strategy is adopted to divide the edges into multiple levels according to the edge weights, using a different sampling rate for each level, and then the final estimate is obtained through weighted combination. An adaptive sampling mechanism is also implemented to dynamically adjust the number of samples by monitoring the convergence of the estimated value, minimizing the computational overhead while ensuring accuracy. After obtaining the average degree estimate, a mapping function is applied to convert it into a sputum distribution density index: , where α and β are parameters calibrated by clinical data. This mapping is based on acoustic theory: the higher the sputum density, the more obvious the change in sound conduction characteristics, and the higher the corresponding graph node connectivity. In this way, The complexity of m is much smaller than the total number of edges |E|, which can efficiently estimate the sputum distribution density and provide key input for subsequent analysis.

[0089] Step S303: Based on the sputum distribution density index, the lungs are divided into multiple anatomical regions and the average degree in each region is calculated to obtain sputum density distribution maps of different lung regions.

[0090] In this step, the overall sputum distribution density index is refined into a distribution map of different anatomical regions of the lungs to provide more accurate spatial positioning information. First, the lungs are divided into multiple regions according to standard lung anatomy, usually including the upper lobe, middle lobe (right lung) or lingula (left lung) and lower lobe of the left and right lungs. Each lobe can be further subdivided into front, back, inside and outside regions, forming a detailed division of 12-18 regions. Each sensor acquisition position is mapped to the corresponding lung anatomical region, and a corresponding relationship between the sensor position and the lung region is established. For each lung region , extract the subgraph consisting of all nodes in the area , and apply the aforementioned random edge sampling algorithm to estimate the average degree of the subgraph To handle the region boundary problem, a soft boundary strategy is adopted, which allows nodes near the boundary to belong to multiple adjacent regions at the same time and sets weights based on distance. The baseline differences between different regions are also taken into account. By comparing with the healthy baseline data, the relative density index is calculated: , which helps to eliminate the basic sound conduction differences caused by anatomical differences. Finally, the sputum density indicators of each region are visualized as a heat map, using a gradient color scale from blue (low density) to red (high density) to represent different density values. This is superimposed on the standard lung anatomy map to form an intuitive sputum density distribution map in different lung regions. This distribution map intuitively shows the hot spots where sputum accumulates, providing spatial precision for subsequent treatment strategies, allowing treatment to act precisely on the areas that need intervention most.

[0091] In step S102, analyzing the spectral feature changes of the cough sound in the standardized audio feature dataset using a parallel batch dynamic algorithm includes:

[0092] Step S304: Based on the sputum density distribution map, the cough sounds in the standardized audio feature dataset are grouped according to the lung regions and corresponding sub-graphs are constructed to obtain acoustic characteristic representations of each region.

[0093] In this step, the sputum density distribution map obtained in the previous step is combined with the standardized audio feature dataset to perform region-specific acoustic analysis. First, based on the mapping relationship between the sensor position and the lung anatomical region, the audio feature data of the cough sound is grouped by lung region. For each lung region , extract the cough sound features collected by all sensors in the area and form a region-specific feature set We pay special attention to cough sounds because the interaction between airflow and sputum during coughing produces characteristic acoustic patterns that contain rich information about sputum characteristics. To capture the internal relationship of acoustic features within a region, we construct a local acoustic feature subgraph for each region. In the subgraph, the node Represents the acoustic feature vector in the area, edge Represents the association between features, weight Quantify the strength of association. An adaptive similarity calculation method is used to determine the edge weights, taking into account the Euclidean distance, cosine similarity, and temporal correlation of the feature vectors. To enhance the expressive power of the subgraph, temporal information is also introduced to encode the characteristic change trend of continuous coughing events into the graph structure to form a dynamic temporal graph. Graph embedding techniques, such as graph convolutional networks (GCNs) or Node2Vec, are further applied to compress high-dimensional graph structure information into low-dimensional embedding vectors, resulting in a more compact representation of the acoustic characteristics of each region. This region-specific acoustic characteristic representation retains the detailed information of the local acoustic pattern, while capturing the complex relationship between features through the graph structure, providing rich input data for subsequent sputum viscosity analysis.

[0094] Step S305: for the acoustic characteristic representation of each region, a parallel computing architecture is used to simultaneously extract the spectrum energy distribution, harmonic components and time-frequency characteristics to obtain a multi-dimensional acoustic feature vector.

[0095] In this step, the acoustic characteristic representation of each lung region is deeply analyzed to extract the multi-dimensional acoustic feature vector. Traditional methods usually adopt serial processing to analyze the data of each region in turn, which is inefficient. The present invention innovatively adopts a parallel computing architecture to process data from multiple regions at the same time, significantly improving the processing speed. In specific implementation, a computing thread or processing unit is assigned to each region, and the feature extraction algorithm is executed in parallel on a multi-core CPU or GPU. For each region, the spectral energy distribution is first analyzed to calculate the energy distribution of the cough sound in different frequency bands. , specifically focusing on the energy ratios of the low-frequency band (100-300Hz, primarily reflecting large airway vibrations), the mid-frequency band (300-800Hz, reflecting medium airway and sputum vibrations), and the high-frequency band (800-2000Hz, reflecting small airway and sputum friction). The energy ratios of these frequency bands are important indicators of sputum viscosity. Secondly, harmonic structure features are extracted, identifying the fundamental frequency (F0) and its harmonic components in the cough sound. The harmonic-to-noise ratio (HNR) and harmonic energy decay rate are calculated. Different sputum viscosities exhibit distinct differences in harmonic structure: high-viscosity sputum typically results in an irregular harmonic structure and rapid decay. Thirdly, time-frequency characteristics are analyzed, using short-time Fourier transform (STFT) or wavelet transform to generate a time-frequency spectrum, from which dynamic features such as energy profile, frequency modulation, and energy decay curve are extracted. Particular attention is paid to the initial phase and tail characteristics of the cough sound, as these characteristics are closely related to the degree of sputum looseness. Through parallel batch processing, the acoustic characteristics of multiple lung regions can be analyzed simultaneously, increasing processing speed by approximately k times (k is the number of regions). Ultimately, a multi-dimensional acoustic feature vector containing 30-50 dimensions is generated for each region, fully characterizing the acoustic characteristics of the cough sound in that region.

[0096] Step S306: Based on the multi-dimensional acoustic feature vector, a mapping relationship between the spectrum feature and the sputum viscosity is established, and the viscosity index of each region is calculated to obtain the sputum viscosity index.

[0097] In this step, the extracted multidimensional acoustic feature vector is converted into a clinically meaningful sputum viscosity index. Sputum viscosity is a key factor in the formulation of treatment plans. Traditional methods usually rely on the experience and judgment of medical staff and lack objective quantitative standards. This paper establishes a scientific mapping model based on the biophysical relationship between acoustic characteristics and sputum viscosity. First, based on preclinical research data, a mapping function between spectral characteristics and sputum viscosity is constructed. This mapping is based on the following biophysical principle: sputum viscosity affects the vibration mode generated when airflow passes through the airway, thereby changing the acoustic characteristics of the cough sound. A multivariate regression model is used to calculate the viscosity index V:

[0098]

[0099] in Represents the energy of different frequency bands, H represents the harmonic characteristics, T represents the time domain characteristics, These weight coefficients are learned from labeled data using a machine learning algorithm, employing algorithms such as gradient boosted decision trees (GBDT) or support vector regression (SVR) to minimize prediction error. To address individual patient differences, an adaptive normalization mechanism is introduced to adjust model parameters based on the patient's baseline acoustic characteristics. A dynamic update mechanism is also implemented to dynamically update the viscosity estimate for each region as new cough sound data is input. The update rule is: , where α is a smoothing factor that controls the weighting of new and old data. Finally, the calculated viscosity index V is normalized to the range [0, 10] and categorized into five levels (very low, low, medium, high, and very high), providing an intuitive reference for clinical decision-making. This approach provides an objective and quantitative assessment of sputum viscosity for each lung region, providing a precise basis for optimizing subsequent treatment parameters.

[0100] Step S307: Integrate the sputum density distribution map and the sputum viscosity index, combine with the patient's lung anatomical structure parameters, and generate an accurate sputum distribution position map through a three-dimensional spatial interpolation algorithm.

[0101] In this step, the previously obtained sputum density distribution map and viscosity index are integrated into a visual three-dimensional distribution map to provide spatial positioning guidance for precise treatment. First, the density data ρ and viscosity data V are normalized to obtain standardized data and , to ensure that the two types of data are comparable. Then, for each lung region Calculation of the composite severity index :

[0102]

[0103] in and is an adjustable weight parameter, the default value is 0.5, and it can be adjusted according to clinical needs. To generate an accurate three-dimensional distribution map, the most matching lung three-dimensional model M is selected or adjusted from the anatomical model library based on the patient's height, weight, age, and gender information. Since the sensor acquisition points are discrete, spatial interpolation is required to obtain a continuous distribution map. The improved radial basis function is used. Interpolation algorithm: , where p is the coordinate of any point in the lung, are the coordinates of the known data points, φ is Kernel function, The coefficients are obtained by solving the linear equations. The kernel function selection is optimized, and multi-scale is used according to the characteristics of the lung anatomical structure. kernel, which can capture both local details and global trends. After interpolation, the continuous severity index Mapped onto a three-dimensional lung model M, a heat map color scheme is used: red indicates areas with severe sputum (high density and high viscosity), yellow indicates areas with moderate severity, green indicates areas with mild sputum, and blue indicates areas with normal or almost no sputum. Treatment priorities are also automatically marked based on the severity index S, dividing lung regions into 3-5 priority levels, marked with numbers on the map (1 indicates the highest priority). To assist in evaluating treatment progress, historical map data is stored and a time-evolution sequence of sputum distribution is generated to help medical staff evaluate treatment effectiveness. The final output sputum distribution location map is a precise three-dimensional visualization model that intuitively shows the distribution of sputum in the lungs, its severity, and treatment priority, providing precise spatial positioning for subsequent vibration solution design.

[0104] In step S103, the application of the shortest path algorithm to calculate the optimal vibration frequency, amplitude and time parameter combination includes:

[0105] Step S401: Based on the sputum viscosity index and the sputum distribution position map, the frequency parameter and the amplitude parameter are mapped into plane coordinates, and the time parameter is mapped into the size of a point to obtain an unweighted disk map.

[0106] In this step, the abstract vibration parameter optimization problem is innovatively transformed into an intuitive geometric space search problem. First, a multidimensional space of vibration parameters is defined, including three key dimensions: frequency dimension (usually ranging from 5-200Hz), amplitude dimension (indicating vibration intensity, usually 0-10mm displacement or 0-2g acceleration), and action time dimension (single vibration duration, usually 1-30 seconds). To facilitate visualization and algorithm processing, this three-dimensional parameter space is mapped to an unweighted disk diagram on a two-dimensional plane. In this mapping, the frequency parameter f is mapped to the x-coordinate of the plane, and the amplitude parameter a is mapped to the y-coordinate of the plane, forming a point on the plane The time parameter t is represented by the size or color of the dot. Usually, the radius or color depth of the dot is proportional to the time. A nonlinear mapping function is used to optimize the parameter distribution to ensure that the key areas in the parameter space have enough space to be represented on the disk diagram. For example, for the frequency parameter, a logarithmic mapping is used. , so that the low-frequency area (which is more critical to sputum loosening) can be represented more precisely. The physiological constraints between parameters are also considered. For example, high-frequency vibrations usually require smaller amplitudes to ensure safety. These constraints are manifested as "forbidden zones" on the disc diagram. In addition, based on the sputum viscosity index and distribution position map, the expected effect areas of different parameter combinations are marked on the disc diagram. High-effect areas are usually represented by brighter colors or special marks. Through this mapping transformation, the complex parameter optimization problem is transformed into the problem of finding the optimal path on a two-dimensional disc diagram, which greatly simplifies the computational complexity and provides an intuitive visual representation.

[0107] Step S402: Based on the unweighted disc diagram, define the distance between two points as a weighted combination of the inverse of the effect gain and the patient tolerance, apply the Dijkstra algorithm to search for the shortest path from the initial state to the target state, and obtain the optimal parameter path.

[0108] In this step, the shortest path algorithm is applied to the constructed unweighted disk graph to find the optimal vibration parameter sequence. The key innovation lies in the definition of the distance function. Instead of using the traditional Euclidean distance, a functional distance metric is defined to take into account the treatment effect and patient comfort. Specifically, for the disk graph Two points in and (Represents parameter combination and ), defining the distance between them It is a weighted combination of the inverse of the effect gain and the patient's tolerance. Indicates that the parameter Switch to Expected improvement in sputum loosening, calculated based on a sputum properties model; patient tolerance Indicates the comfort level of the parameter switching process, with a value range of [0,1], where 1 indicates complete comfort. The distance function design ensures that parameter transitions with large effect gains and high tolerance have shorter "distances" and are preferred. After determining the starting and end points on the disk graph, the improved Dijkstra algorithm is applied to find the shortest path. Starting point Represents the current no-vibration state, the end point Represents the ideal state of sputum loosening effect. The algorithm starts from the starting point, gradually explores the reachable nodes, updates the shortest distance to each node, until the shortest path to the end point is found. To improve the efficiency of the algorithm, a priority queue is used and a heuristic search strategy is introduced, such as the A algorithm, which uses the estimated distance to the target to guide the search direction. During the search process, multiple constraints are considered: frequency change smoothness (the frequency change of adjacent points does not exceed the threshold ), amplitude change smoothness (the amplitude change of adjacent points does not exceed the threshold ) and patient safety margins (excluding parameter areas that may cause discomfort or risk). After obtaining the initial path, a path smoothing algorithm is applied to reduce unnecessary parameter oscillations and make the vibration parameter changes smoother and more continuous. Finally, an optimal path in the parameter space is output. ,This path represents the optimal parameter sequence from the initial state to the target sputum loosening effect, taking into account both treatment effect and patient comfort.

[0109] Step S403: Based on the optimal parameter path, a dynamic programming algorithm is applied to design a curve showing the change of vibration parameters over time, and time allocation is optimized in combination with a patient tolerance model to obtain the personalized vibration therapy plan.

[0110] In this step, the optimal parameter path obtained in the previous step is Convert it into a specific time execution plan to form a complete personalized vibration treatment plan. First, the total treatment time Divide into n time periods, each of which can apply different vibration parameters. Define the state space of dynamic programming, each state Indicates that the jth parameter in the parameter path is used in the i-th time period For each state, define the value function Indicates the maximum cumulative effect that can be obtained from the beginning of this state to the end of treatment. Establish the core recursive equation of dynamic programming, considering the effect of the current time period and the optimal decision of the future time period. In the calculation process, introduce the patient tolerance model , describing the patient's tolerance for a parameter p for a duration t. When tolerance falls below a threshold, parameters are automatically adjusted or rest periods are inserted to ensure a comfortable treatment experience. Tighter parameter change constraints are set for specific patient groups (such as the elderly or children), while retaining flexibility for fine-tuning parameters. If patient discomfort (such as increased heart rate or decreased blood oxygen) is detected, vibration intensity can be dynamically reduced. After time planning is completed, lung regions are prioritized for treatment based on regional priorities in the sputum distribution map. Basic vibration parameters are fine-tuned based on regional characteristics (such as sputum viscosity and anatomical location): low-frequency components and vibration intensity are increased in highly viscous areas, vibration duration and energy are increased in deep areas, and vibration amplitude is reduced and frequency is increased in sensitive areas. Treatment time allocation is also optimized, with more treatment time allocated to areas of high severity. Furthermore, optimal vibrator placement and activation sequence are designed. The primary vibrator is placed at the chest wall location corresponding to the target treatment area, while auxiliary vibrators are placed at auxiliary locations to form a coordinated vibration field. Phase differences between the vibrators are designed to create a directional vibration wave. Ultimately, a visual treatment flowchart is generated, including a timeline, parameter curves, area indications, and operation prompts. Final adjustments are made based on individual factors such as the patient's age, body shape, and comorbidities to output a highly personalized vibration treatment plan.

[0111] In step S104, according to the personalized vibration therapy plan, combined with the patient's real-time breathing cycle, the operating parameters of the sonic vibrator and the negative pressure suction device are synchronously controlled to obtain a coordinated multimodal therapy execution instruction, including:

[0112] Step S501: Based on the personalized vibration therapy plan, the vibration parameters are converted into driving signals, the operating frequency, amplitude and phase of the acoustic vibrator are controlled, and a vibrator driving instruction is obtained.

[0113] In this step, the optimized vibration parameters are converted into precise hardware drive signals, achieving precise vibration effects on the target lung area. First, the parameter set in the vibration therapy plan, including frequency, amplitude, and duration, is analyzed and preprocessed to ensure they are within the hardware's supported range. Next, the most appropriate waveform type is designed based on the specific vibration requirements. For general vibration requirements, a sine wave is typically used as the base waveform. For highly viscous sputum, a triangle wave with a faster energy ramp may be chosen. For extremely difficult-to-loosen sputum areas, a square wave with a stronger impact may be used. In some cases, multiple base waveforms are superimposed in specific proportions to form a composite waveform with specific spectral characteristics to achieve optimal sputum loosening. For systems equipped with multiple vibrators, the phase relationship between each vibrator is precisely controlled to achieve various vibration modes: in-phase vibration synchronizes all vibrators to enhance the vibration effect; alternating vibration alternates the vibrators to reduce patient fatigue; wave vibration activates the vibrators in a specific sequence, forming a directionally propagating vibration wave; and focused vibration constructively adds the vibration waves of multiple vibrators to the target area, enhancing the localized effect. To avoid sudden changes at the start and end of vibration, a smooth amplitude envelope curve is designed, including an exponentially rising starting segment, a stable segment with constant amplitude, and an exponentially decreasing ending segment. The generated theoretical waveform also requires digital signal processing to adapt to the hardware characteristics, including sampling rate optimization, quantization processing, filtering processing, and distortion compensation. Finally, based on the electrical characteristics of the vibrator, a matching drive current / voltage signal is generated: a PWM modulated drive current signal is generated for the electromagnetic vibrator, a high-voltage and low-current drive voltage signal is generated for the piezoelectric vibrator, and a speed control signal is generated for the eccentric wheel vibrator. Through this series of processing, accurate vibrator drive instructions are output to ensure the accurate implementation of vibration therapy.

[0114] Step S502: Based on the personalized vibration therapy program and the patient's real-time breathing cycle, the negative pressure intensity, suction time window and pressure change curve coordinated with the vibration parameters are calculated to obtain a negative pressure control instruction.

[0115] In this step, an optimal vacuum suction control strategy is designed to work synergistically with vibration therapy to improve sputum clearance efficiency. First, the patient's respiratory cycle characteristics are analyzed, including inspiratory and expiratory time, respiratory rate, and tidal volume. Based on these parameters, the optimal window for vacuum suction is determined, typically during mid-to-late expiration (60% to 90% of expiratory time), when the direction of natural airway airflow aligns with the direction of vacuum suction, maximizing suction effectiveness. Vacuum intensity, a key parameter determining suction effectiveness, is dynamically calculated based on multiple factors: sputum viscosity index (higher viscosity requires greater vacuum), airway anatomical location (deeper airways require greater vacuum delivery), and patient tolerance (the upper limit of vacuum is adjusted based on the patient's age and physical condition). To enhance comfort and effectiveness, a smooth vacuum curve is designed, including a startup phase that slowly increases vacuum, a treatment phase that maintains target vacuum, and a release phase that gradually decreases vacuum to avoid sudden airway collapse and mucosal damage. The optimal synergy between vibration and negative pressure has also been designed: the sequential mode first uses vibration to loosen sputum and then uses negative pressure to aspirate; the alternating mode alternates vibration and negative pressure to reduce patient discomfort; the superimposed mode applies vibration and low-intensity negative pressure simultaneously, followed by increased negative pressure; and the pulsed mode pulses the negative pressure, forming a specific proportional relationship with the vibration frequency. Multiple safety limits are set to prevent adverse reactions caused by negative pressure aspiration, including maximum negative pressure limits, duration limits, oxygen saturation monitoring, and heart rate monitoring. Ultimately, a precise sequence of negative pressure control instructions is generated, including the operation time point accurate to the millisecond, the target negative pressure value at each time point, the rate control parameter for pressure change, and the maintenance time of each pressure stage. These instructions directly drive the electronically controlled valve of the negative pressure generating device, achieving precise negative pressure control and forming an optimal synergistic effect with vibration therapy.

[0116] Step S503: By monitoring the patient's respiratory phase in real time, using a prediction algorithm to estimate the optimal intervention time, controlling the start and stop timing of the acoustic vibrator and the negative pressure suction device, and obtaining a respiratory synchronization signal.

[0117] In this step, the vibrator and the vacuum suction device are precisely synchronized with the patient's respiratory rhythm, ensuring optimal timing for therapeutic intervention. First, respiratory signals are collected through the fusion of multiple sensors, including a thoracoabdominal motion sensor (collecting displacement signals of chest and abdominal expansion and contraction), an airflow sensor (measuring the velocity and direction of airflow through the mouth and nose), and an acoustic sensor (collecting respiratory audio features). A signal fusion algorithm dynamically adjusts the weights of each signal source based on different circumstances to obtain a stable and reliable integrated respiratory signal. The respiratory cycle is divided into four key phases: early inspiration (when the chest begins to expand), late inspiration (when the chest reaches maximum expansion), early expiration (when the chest begins to contract), and late expiration (when chest contraction is nearly complete). Phase identification utilizes wavelet transforms and peak detection algorithms to accurately capture phase transition points. To overcome response delays, a prediction algorithm is employed to proactively identify phase changes. These include short-term predictions based on the first half of the current respiratory cycle to predict the second half, medium-term predictions based on the patterns of the last three to five respiratory cycles to predict the next cycle, and long-term predictions that integrate patient historical data to establish a personalized respiratory model. The optimal timing of intervention is determined based on treatment goals: vibration therapy is typically initiated during early exhalation, when airflow within the airway facilitates the outward movement of loosened sputum. Vacuum suction is typically applied during mid-to-late exhalation, leveraging natural expiratory flow to enhance suction. A precise time difference between vibration and vacuum is also designed, typically 0.5-1.5 seconds. The system dynamically adjusts intervention timing based on changes in the patient's respiratory rhythm, adapting to changes in respiratory rate and depth, and addressing irregular breathing. A high-precision clock synchronizes multiple treatment modules, including low-latency hardware trigger signals, pre-trigger compensation to account for device startup delays, and phase-locking the device operating cycle to the respiratory cycle. The system also recognizes and responds to emergencies such as coughing, immediately adjusting or pausing treatment interventions when coughing occurs and automatically adjusting treatment strategies after coughing to leverage the loosening effect of coughing. These technologies output precise timing control signals, achieving millisecond-level respiratory synchronization, significantly improving treatment efficacy and minimizing patient discomfort.

[0118] Step S504: Based on the vibrator drive instruction, the negative pressure control instruction and the respiratory synchronization signal, an execution timing table for collaboration between devices is generated to obtain a device collaboration control strategy.

[0119] In this step, a precise multi-device coordinated execution strategy is designed to ensure that each treatment module works together to achieve optimal therapeutic results. First, a complete treatment schedule is constructed, containing detailed information about all treatment events: event type (vibration start / stop, negative pressure start / stop, parameter adjustment, etc.), trigger conditions (time trigger, respiratory phase trigger, or physiological indicator trigger), execution parameters (detailed parameter settings for each event), and priority (the priority level for handling event conflicts). The schedule adopts a hierarchical structure, comprehensively planning the entire treatment process, from macro-treatment stages to the respiratory cycle level and then to precise time points. A low-latency inter-device communication protocol is designed, including a main controller responsible for global coordination and decision-making, standardized control interfaces for each functional module, a low-latency communication protocol implemented using real-time bus technologies (such as EtherCAT or CAN FD), and microsecond-level device synchronization using the IEEE 1588 Precision Time Protocol (PTP). Clear event prioritization rules are defined, with safety-related events receiving the highest priority, patient-initiated actions receiving the next highest priority, core treatment events receiving the lowest priority, and monitoring and recording events receiving the lowest priority. When multiple events conflict, the event with the highest priority is executed. Multiple treatment modes and their switching conditions are defined, including a preparation mode to help patients adapt, a loosening mode focused on loosening sputum, a discharge mode that combines vibration and negative pressure, and a rest mode for patients to rest. Mode switching uses a smooth transition algorithm to avoid patient discomfort caused by sudden parameter changes. Closed-loop control is implemented, with strategies adjusted in real time based on treatment effectiveness. This includes real-time effect evaluation based on analysis of indicators such as audio changes and sputum discharge volume, parameter adaptation that automatically fine-tunes treatment parameters based on effect evaluation, and progress tracking that records treatment progress in each area and dynamically adjusts focus. Execution instructions are also optimized, including merging similar instructions with similar execution times, preloading upcoming instructions into the device cache, allocating communication bandwidth based on task urgency, and designing instruction confirmation and failure retry mechanisms. These strategies generate a precise and comprehensive device collaborative control strategy to ensure seamless coordination among treatment modules and maximize treatment effectiveness.

[0120] Step S505: Integrate the vibrator drive instruction, the negative pressure control instruction, the respiratory synchronization signal and the device collaborative control strategy to ensure the coordinated cooperation of the acoustic vibration treatment and the negative pressure suction device treatment to obtain the coordinated multimodal treatment execution instruction.

[0121] In this step, the various control instructions and strategies generated in the previous steps are finally integrated to form a complete multimodal treatment execution instruction set. First, the vibrator drive instructions, negative pressure control instructions, and respiratory synchronization signals are time-aligned and synchronized to ensure that all instructions are based on a unified time base. Each instruction is accurately timestamped, and a time compensation algorithm is applied to eliminate response delay differences between different devices. Next, based on the device collaborative control strategy, the various instructions are prioritized and conflicts resolved. When multiple instructions execute within the same time window, the execution order is determined according to predefined priority rules to ensure that critical instructions are not delayed or lost. The instruction sequence is also verified for integrity and safety, checking for potentially dangerous instruction combinations or parameter violations. If any issues are identified, adjustments are automatically made or safety warnings are triggered. To improve execution efficiency, instructions are batched and optimized, combining logically related instructions into instruction blocks to reduce communication overhead. An instruction preloading mechanism is also implemented to pre-send upcoming instructions to the cache of each execution unit, ensuring precise execution without being affected by communication delays. In order to deal with abnormal situations that may occur during the execution process, a complete exception handling process has been designed, including device response timeout handling, retry strategies for failed instruction execution, and safety degradation plans in emergency situations. A detailed instruction execution log recording mechanism has also been established to record the sending time, receipt confirmation, and execution completion status of each instruction, providing data support for subsequent analysis and optimization. Ultimately, a multimodal treatment execution instruction set with a complete structure, precise timing, and internal coordination is output. These instructions directly drive the operation of each treatment device, ensuring that vibration therapy and negative pressure suction therapy can accurately coordinate and form a combined force to achieve the best cough-assisting effect. Compared with traditional independent control, this collaborative execution strategy can significantly improve the overall effect and efficiency of treatment and reduce equipment conflicts and idling time.

[0122] In step S503, the patient's respiratory phase is monitored in real time, the optimal intervention timing is estimated using a prediction algorithm, and the start and stop timing of the acoustic vibrator and the negative pressure suction device is controlled to obtain a respiratory synchronization signal, including:

[0123] Step S5031: The respiratory signal is collected by fusing the chest and abdominal motion sensor, the airflow sensor, and the sound sensor, and the respiratory cycle is divided into four phases: the early inspiration phase, the late inspiration phase, the early expiration phase, and the late expiration phase, to obtain the respiratory phase recognition result.

[0124] In this step, a variety of sensors work together to fully capture the patient's respiratory activity information. Chest and abdominal motion sensors usually use elastic belt or capacitive sensors, which are worn on the patient's chest and abdomen respectively to monitor the expansion and contraction movements of the chest and abdomen in real time. These sensors can provide intuitive information on the amplitude and rhythm of respiratory movements, but may be interfered with by the patient's body movements. The airflow sensor is installed on the mask or nasal cannula to directly measure the airflow speed and direction at the patient's mouth and nose, providing more direct respiratory airflow data, but wearing it may increase the patient's discomfort. The sound sensor (usually a highly sensitive MEMS microphone) is attached to a specific position on the patient's chest wall to collect sound signals generated during the breathing process. These signals contain rich respiratory phase information and do not interfere with the patient's normal activities. A weighted fusion algorithm is used to integrate the data of these three types of sensors. The fusion formula is: , where the weight coefficient 、 and Dynamic adjustments are made based on signal quality to ensure stable respiratory monitoring even when the signal quality of one sensor type degrades (e.g., when patient movement interferes with the chest strap signal). Based on the fused respiratory signal, peak-valley detection and slope analysis algorithms are applied to accurately segment the continuous respiratory waveform into four key phases: early inspiration (when the chest cavity begins to expand and the signal begins to rise), late inspiration (when the chest cavity reaches maximum expansion and the signal peaks), early expiration (when the chest cavity begins to contract and the signal begins to decline), and late expiration (when the chest cavity nears completion and the signal reaches a valley). Machine learning algorithms (such as support vector machines or random forests) are also used to calibrate the phase segmentation results. These algorithms use pre-trained data to learn to identify phase characteristics of different respiratory patterns, improving recognition accuracy. For irregular breathing patterns, adaptive thresholding technology is used to dynamically adjust the phase determination criteria based on the characteristics of the most recent respiratory cycles. Ultimately, accurate respiratory phase identification results are output, including the onset time, duration, and relative intensity of each phase, providing foundational data for predicting the timing of subsequent interventions.

[0125] Step S5032: Based on the respiratory phase identification result, wavelet transform and peak detection algorithm are applied to capture the phase transition point, and the phase time point of the next cycle is predicted in combination with the patterns of the latest 3-5 respiratory cycles to obtain the respiratory phase prediction result.

[0126] This step requires not only identifying the current respiratory phase but also predicting future phase changes to overcome response delays and achieve precise intervention timing control. First, the respiratory signal is subjected to time-frequency analysis using a wavelet transform. Compared to the traditional Fourier transform, the wavelet transform better captures the local characteristics of non-stationary signals. Wavelet basis functions (typically Daubechies or Morlet wavelets) are selected to suit the characteristics of the respiratory signal. The signal is then decomposed at multiple scales to extract the time-varying characteristics of different frequency components. Within the wavelet coefficients, an improved peak detection algorithm is applied to precisely locate phase transition points. This algorithm, incorporating information from the first-order derivative (slope) and second-order derivative (curvature) of the signal, accurately identifies inflection points and extreme values, maintaining stability even in the presence of noise or small fluctuations. To predict future respiratory phases, a multi-level prediction strategy is employed. Short-term prediction uses the first half of the current respiratory cycle to predict the second half, leveraging the symmetry and continuity characteristics within the respiratory cycle. Medium-term prediction analyzes patterns from the last three to five respiratory cycles, including cycle length, the proportion of each phase, and amplitude trends, to construct a local time series model to predict the characteristics of the next cycle. An autoregressive moving average (ARMA) model or Kalman filter is used for time series prediction, and model parameters are dynamically updated based on historical data. For longer-term predictions, a personalized breathing pattern model is established by combining the patient's historical respiratory data and current physiological state. This model takes into account factors such as the patient's age, weight, disease type, and treatment stage. A prediction confidence assessment mechanism is also implemented, calculating the confidence of the current prediction based on breathing regularity and historical prediction accuracy. When the confidence falls below a threshold, a more conservative intervention strategy is adopted. Through these technologies, phase transition points can be accurately predicted 50-500 milliseconds in advance, providing ample preparation time for treatment intervention and effectively overcoming the impact of device response delays.

[0127] Step S5033: Based on the respiratory phase prediction result, determine the optimal intervention time for vibration therapy to start at the early stage of exhalation and negative pressure suction to be implemented in the middle and late stages of exhalation, design the time difference between vibration and negative pressure, and obtain the respiratory synchronization signal.

[0128] In this step, the optimal timing for therapeutic intervention is precisely calculated based on the predicted respiratory phase information, generating synchronized signals to control the vibrator and vacuum device. First, the optimal intervention window is determined based on respiratory physiology. Vibration therapy is typically initiated during early expiration. This timing is based on two key considerations: first, airflow in the airway is directed outward at this time, helping to transport loosened sputum; and second, high airway pressure and moderate tissue tension during early expiration optimize vibration transmission efficiency. Specifically, the vibration initiation point is set 10%-20% after the onset of the expiratory phase. This precise timing is dynamically adjusted based on the patient's respiratory rate and sputum characteristics. For vacuum suction, the optimal time is mid-to-late expiration (typically 60%-80% of the expiratory phase). At this point, natural expiratory flow has already transported some of the loosened sputum into the larger airways, allowing vacuum suction to most effectively expel it. Carefully designing the time difference between vibration and vacuum is crucial, as it determines the synergistic effect of the two treatment modalities. The time difference is usually set to 0.5-1.5 seconds, and the specific value is calculated based on the patient's respiratory rate, airway length, and sputum characteristics: ,in is the expiratory time, L is the estimated airway length, V is the sputum viscosity, 、 and is the weighting coefficient. The variability of the respiratory cycle is also taken into account. When changes in respiratory rate are detected, the relative position of the intervention timing is adjusted accordingly to ensure that intervention always occurs at the optimal physiological time. For irregular breathing patterns, an adaptive strategy is adopted, which may temporarily switch to an intervention mode based on absolute time, or wait until breathing becomes regular before implementing synchronized intervention. The generated respiratory synchronization signal is a digital signal stream containing precise timestamps, marking the precise time points of vibration start / stop and negative pressure start / stop. These signals directly control the working state of the treatment device, ensuring that the treatment intervention is precisely synchronized with the patient's breathing at the millisecond level, significantly improving the treatment effect and reducing patient discomfort.

[0129] In step S105, the working parameters are dynamically adjusted based on the reinforcement learning algorithm to obtain an optimized cough and sputum auxiliary treatment plan, including:

[0130] Step S601: The patient's physiological state, sputum characteristics, and treatment progress are used as the state space, and the working parameters and treatment mode selection are used as the action space. A reward function is constructed based on the short-term sputum cough effect index to obtain a formal representation of the reinforcement learning problem.

[0131] In this step, the treatment parameter optimization problem is innovatively formalized into a reinforcement learning framework, which enables the continuous improvement of treatment strategies through interactive learning. First, a comprehensive state space is defined, which includes multiple key dimensions: the patient's physiological state vector P includes blood oxygen saturation ( ), heart rate (HR), respiratory rate (RR), and heart rate variability (HRV), which reflect the patient's overall condition and tolerance to treatment. The sputum feature vector M, including parameters such as density distribution, viscosity, and cleared percentage, describes the current physical properties and distribution of sputum. The treatment history vector H records information such as the current treatment stage, cumulative treatment time, and recent treatment effects, providing a temporal context for treatment progress. The patient feature vector F contains static features such as age, disease severity, and tolerance, which are used to personalize treatment strategies. Due to the high dimensionality of the original state space, dimensionality reduction techniques such as principal component analysis or autoencoders are applied to extract the most informative feature combinations and construct a compact and information-rich state representation. The action space A includes all controllable parameters: vibration frequency, vibration amplitude, negative pressure intensity, treatment mode selection, etc. A multi-objective balanced reward function is designed, with the primary reward based on improvements in sputum looseness and discharge volume. A comfort penalty term (uncomfortable treatment reduces the total reward), a time penalty term (to encourage efficient treatment), and a safety constraint term (to strongly penalize behaviors that may pose safety risks) are also introduced. The design of the reward function balances the three key goals of treatment effect, patient comfort and treatment efficiency, and adjusts the relative importance of each goal through the weight coefficient. The ultimate goal is to find an optimal strategy , maximizing the expected cumulative reward from following this policy starting from any state s. This reinforcement learning-based formalization enables autonomous learning and adaptation in complex medical environments, gradually discovering the optimal combination of treatment parameters without pre-programming all possible scenarios.

[0132] Step S602: Based on the formalized representation of the reinforcement learning problem, a policy gradient method and a neural network parameterization strategy are used to optimize the treatment strategy through gradient estimation and parameter updating to obtain an initial policy network.

[0133] In this step, the core algorithm of reinforcement learning is implemented to optimize the treatment strategy through continuous trial and learning. The policy gradient method is used, which is a type of reinforcement learning algorithm that directly optimizes the policy function and is particularly suitable for problems in continuous action space. First, the policy function is parameterized using a neural network. The network takes a state s as input and outputs the probability distribution of each action a chosen in that state (for discrete actions) or the mean and variance of the actions (for continuous actions). The network typically uses a multi-layer perceptron (MLP) with 2-3 hidden layers, each with 64-128 neurons, and uses ReLU or tanh activation functions. To capture the temporal nature of the state, an LSTM or GRU layer may be added at the front end to process the historical state sequence. The policy network's parameters θ are initialized to small random values ​​and then continuously updated and optimized through interaction with the environment. The policy gradient is calculated using the REINFORCE algorithm or its variants. The basic idea is to increase the probability of actions that lead to high rewards and decrease the probability of actions that lead to low rewards. To reduce the variance of the gradient estimate, a baseline function (usually a value function estimate) is introduced, only considering the advantage relative to the baseline. Importance sampling techniques are also used to improve sample efficiency by utilizing off-policy data. Parameter updates are performed using the Adam optimizer, combined with gradient clipping to prevent gradient explosion. To improve training stability, various techniques were implemented: entropy regularization prevents the policy from prematurely converging to a suboptimal solution; experience replay breaks sample correlations; and the target network slows down the rate of change in value estimates. Furthermore, a conservative policy iteration mechanism, tailored to the specificities of medical applications, ensures that new policies do not significantly deviate from the currently effective policy, thus avoiding drastic fluctuations in treatment quality. These techniques enable the gradual learning and improvement of treatment strategies based on limited interaction data, outputting a preliminarily trained policy network capable of recommending appropriate treatment parameters based on the patient's condition and sputum characteristics. This initial policy network provides a good starting point for subsequent online optimization.

[0134] In this embodiment, the policy network adopts a 4-layer structure: the input layer receives a 24-dimensional state vector (including 8-dimensional physiological state, 6-dimensional sputum characteristics, 6-dimensional treatment history, and 4-dimensional patient characteristics); the first hidden layer contains 128 neurons and uses the ReLU activation function; the second hidden layer contains 64 neurons and also uses the ReLU activation function; the output layer corresponds to the action space dimension (8 dimensions in this embodiment, including vibration frequency, vibration amplitude, negative pressure intensity, etc.), uses the tanh activation function for continuous actions and scales them to an appropriate range, and uses the softmax activation function for discrete actions. The key hyperparameter settings are as follows: the initial value of the learning rate is set to 0.001, and the Adam optimizer is used. The experience replay buffer has a capacity of 10,000; the discount factor γ = 0.95; the batch size is 64; and the target network update frequency is every 100 steps. Training data primarily comes from two sources: clinical data collected from 300 cases of cough-assisted sputum treatment (including patient status, treatment parameters, and effect evaluation); and interaction data accumulated in simulation environments and real-world applications. Training employs a two-stage strategy: in the first stage, supervised learning is used to initialize network parameters by imitating expert behavior from clinical data; in the second stage, reinforcement learning is used to continuously optimize the strategy through interaction with the environment.

[0135] Step S603: Maintaining the experience replay buffer to store the treatment process data in the real-time feedback data, sampling important experiences first according to the size of the time difference error, performing batch learning and parameter updating, continuously updating the initial strategy network parameters, and obtaining an optimized strategy network.

[0136] In this step, the strategy network is continuously optimized through the continuous learning mechanism to adapt to the individual characteristics and status changes of the patient. The core technology is the experience replay mechanism, which maintains a fixed-capacity experience replay buffer D to store the experience tuples collected during the treatment process. , including the current state, the action performed, the reward obtained and the next state. This mechanism breaks the temporal correlation between samples and improves learning efficiency and stability. Using the priority experience replay strategy, according to the temporal difference error The size of the time series difference (TDD) determines sample priority. Samples with large temporal difference errors (indicating a large discrepancy between prediction and reality) have a higher sampling probability, allowing for focused learning of unexpected or informative experiences. To prevent oversampling of high-priority samples, importance sampling weights are introduced for bias correction. The sampling strategy combines uniform random sampling with priority sampling to ensure sample diversity while focusing on important samples. Mini-batches (typically 32-128 samples) are periodically drawn from the buffer for batch learning. The current policy output is calculated using forward propagation, and network parameters are then updated via backpropagation. To improve learning efficiency, a multi-step bootstrapping method is implemented, using n-step returns instead of single-step returns to reduce estimation bias. A dual-network architecture is also employed, maintaining a target network and an action network. The target network parameters are updated slowly to improve training stability. To address the unique characteristics of medical scenarios, a conservative update mechanism is implemented, limiting the policy update amplitude through trust region constraints or KL divergence penalties to ensure smooth evolution of the treatment strategy. An adaptive learning rate adjustment mechanism is also designed to dynamically adjust the learning rate based on learning progress and data quality. Through these continuous learning mechanisms, the initial strategy network continuously absorbs new experiences and optimizes parameters, gradually evolving into a more precise and personalized optimization strategy network that can provide optimal treatment parameter recommendations based on the characteristics of specific patients.

[0137] Step S604: Design an exploration strategy with parameter noise and entropy regularization, conduct planned exploration within the safety parameter range, adjust the exploration degree according to the uncertainty estimated by the value function, and obtain an exploration-exploitation balance strategy.

[0138] In this step, the key challenge in reinforcement learning is solved: the exploration-exploitation balance problem. This problem is particularly important in medical scenarios, as it is necessary to explore possible better treatment parameters while ensuring patient safety and treatment effectiveness. A multi-level exploration strategy is designed. First, a parameter noise method is used to add Gaussian noise to the parameters output by the policy network. The noise level is dynamically adjusted as the treatment progresses. Compared with traditional action space noise, parameter noise can produce more coherent exploration behavior and is more suitable for medical applications. The noise level follows the annealing mechanism and gradually decreases as treatment experience accumulates. The formula is: , where t is the treatment duration and τ is the time constant. A policy entropy term is also incorporated into the objective function to encourage the policy to maintain moderate stochasticity and prevent premature convergence to a suboptimal solution. The weight of the entropy regularization term is also dynamically adjusted over time, initially high to encourage extensive exploration and later reduced to focus on already discovered effective strategies. To ensure exploration safety, a planned exploration mechanism is designed that divides the parameter space into a safe zone, a transition zone, and a risk zone. Exploration is performed only in the safe zone and part of the transition zone, with a strictly limited exploration step size. Uncertainty-driven exploration is also implemented. By integrating multiple policy networks (e.g., using dropout or bootstrap methods), the uncertainty of the value function is estimated, prioritizing exploration in regions with high uncertainty in the value estimate, which typically contain more unknown information. Particular attention is paid to the safe margin of exploration, dynamically adjusting the exploration range based on the patient's physiological indicators. When sensitive states (e.g., elevated heart rate or decreased blood oxygen) are detected, the exploration range is automatically narrowed or paused. A tiered exploration strategy has also been designed, employing different exploration strategies for different parameters: allowing for a wider range of exploration for parameters with low safety sensitivity (such as fine-tuning of vibration frequency), while adopting a more conservative exploration strategy for parameters with high safety sensitivity (such as negative pressure intensity). Through these technologies, while ensuring patient safety, we can continuously explore more optimal treatment parameter combinations, gradually refine treatment strategies, and achieve an optimal balance between exploration and utilization.

[0139] Step S605: Based on the optimal strategy network and the exploration-exploitation balance strategy, the recommended parameters are calculated according to the current patient status in the real-time feedback data, and the safety constraint check and smoothing processing are applied to dynamically adjust the working parameters to obtain the optimized cough-assisted treatment plan.

[0140] In this step, the learned strategy is converted into actual treatment parameter adjustment to achieve closed-loop control. First, according to the real-time collected patient status data s, the strategy network is optimized. Calculate the recommended parameter set: . This calculation process is a forward propagation, which inputs the current state vector and outputs the recommended treatment parameters, including vibration frequency, vibration amplitude, negative pressure intensity, etc. According to the exploration-utilization balance strategy, appropriate noise may be added to the output parameters to promote exploration. After obtaining the preliminary recommended parameters, multi-layer safety constraint checks are applied to ensure that the parameters are within a safe range. Safety constraints include hard constraints (absolute upper and lower limits of parameters) and soft constraints (limits on the rate of change of parameters). Hard constraints ensure that the parameters do not exceed the device support range or the patient safety threshold, such as the vibration frequency is limited to 5-200Hz and the negative pressure intensity does not exceed Soft constraints limit the rate of change of parameters to prevent sudden changes in parameters from causing discomfort to patients, such as the vibration frequency change rate not exceeding 10Hz / minute, and the negative pressure intensity change not exceeding / minute, etc. Parameter correlation checks are also implemented to ensure that the parameter combination conforms to physiological principles. For example, high-frequency vibration usually requires a smaller amplitude to ensure safety. Parameters that pass the safety check also need to be smoothed. A low-pass filter is applied to smooth the parameter changes to avoid short-term fluctuations. The formula is , where α is a smoothing coefficient, typically 0.1-0.3. A hierarchical parameter adjustment strategy is adopted, dividing the parameters into a fine-tuning layer (low-risk parameters that are adjusted frequently), a moderate adjustment layer (medium-impact parameters that are adjusted regularly), and a large-scale adjustment layer (high-impact parameters that are adjusted only when necessary). Parameters at different levels have different adjustment frequencies and amplitude limits. An emergency response mechanism is also designed to quickly respond to changes in patient status or equipment abnormalities, such as immediately reducing treatment intensity when patient discomfort is detected. Ultimately, a complete optimized cough-assisted treatment plan is generated, including detailed parameter settings, time arrangements, regional sequences, and operating guidelines. This plan is continuously optimized through a closed-loop control mechanism and can be adjusted in real time based on treatment progress and patient responses, achieving truly precise and personalized treatment.

[0141] In step S601, the reward function is constructed based on the short-term sputum cough effect index, including:

[0142] Step S6011: Use overlapping sliding windows to continuously analyze the treatment process data in the real-time feedback data, evaluate the degree of sputum loosening through respiratory sound energy ratio, sputum sound changes and sound conduction improvement, and obtain a sputum loosening index.

[0143] In this step, a method for continuously evaluating the effect of sputum loosening was designed to monitor the progress of treatment in real time through changes in acoustic characteristics. First, the overlapping sliding window technology was used to perform segmented analysis on the continuously collected audio data. The window length is usually set to 30-60 seconds, which is sufficient to capture the changes in acoustic characteristics of multiple respiratory cycles; the overlap rate of adjacent windows is set to 50-75% to ensure the continuity and smooth transition of the analysis. The Hanning window function is applied to the data in each window to reduce the spectral leakage effect and improve the accuracy of spectral analysis. In each analysis window, the respiratory sound energy ratio is first calculated, that is, the ratio of the energy in different frequency bands The low frequency band (100-300Hz) mainly reflects the vibration of the large airways, while the high frequency band (800-2000Hz) is related to the friction between the small airways and sputum. As the sputum loosens, the high frequency components usually increase. By tracking The changing trend of the phlegm wheezing is used to evaluate the progress of sputum loosening. Secondly, the changing characteristics of the wheezing sound are analyzed. The wheezing sound is a characteristic sound produced when sputum partially blocks the airway, usually manifested as a narrowband peak in the spectrum. The frequency, intensity and duration of the wheezing sound are identified by the spectral peak detection algorithm, and the changes in these characteristics are tracked. As the treatment progresses, the wheezing sound will usually first increase (the sputum begins to loosen but has not yet been discharged) and then decrease (the sputum is successfully discharged or moves to a larger airway). Third, the improvement of sound conduction is evaluated. The presence of sputum will change the conduction characteristics of sound in the lungs, which is manifested as a decrease in the sound correlation between different collection points. The cross-correlation coefficient between the sound signals collected by sensors at different positions is calculated. As the sputum loosens and decreases, the cross-correlation coefficient usually increases, indicating that the sound conduction characteristics have improved. These three indicators are weighted together to calculate the sputum loosening index: , where Δ represents the change from baseline, 、 and is a weighting factor, determined through clinical validation. To accommodate individual patient differences, an adaptive baseline mechanism was implemented, using acoustic signatures from the initial phase of treatment as a personalized baseline to assess relative changes rather than absolute values. Finally, the calculated loosening index L was normalized to a range of 0–100, providing a visual representation of the degree of sputum loosening and providing a key input for the subsequent reward function calculation.

[0144] Step S6012: Based on acoustic event statistics, negative pressure system feedback and cough effectiveness evaluation, the sputum output estimation index is obtained by weighted summation of the number of acoustic events, negative pressure recovery estimation and cough energy.

[0145] In this step, the system attempts to address a clinical challenge: how to objectively quantify the volume of sputum expelled. Traditional methods often rely on subjective estimates by healthcare professionals or require collection of expelled material for measurement, making real-time assessment difficult. This system innovatively combines multiple indirect indicators to achieve real-time estimation of sputum expelled volume. First, the system performs acoustic event counting to identify and count characteristic sound events associated with sputum movement. These events include sudden changes in sputum sound, bubbling sounds (the sound of bubbles bursting when sputum interacts with airflow), and the characteristic "gurgling" sound produced when sputum passes through larger airways. The system automatically identifies these acoustic events using an acoustic event detection algorithm that combines energy threshold detection and spectral pattern matching. Each identified event is weighted based on its intensity and duration. Events with greater intensity and longer duration generally correspond to greater sputum movement. Second, the system analyzes feedback data from the vacuum suction system. Modern vacuum suction devices are typically equipped with flow and pressure sensors that monitor changes in flow and resistance in the suction line. By analyzing the vacuum system's flow curve, the system identifies flow fluctuation patterns caused by sputum entering the suction line; by analyzing the vacuum system's pressure curve, it identifies pressure fluctuations caused by sputum obstruction. Combining flow integral and pressure change pattern, the system is able to estimate the amount of sputum recovered by the negative pressure system. Third, the system evaluates the effectiveness of coughing, which is a natural mechanism for expectoration, and its effectiveness is closely related to the amount of sputum expelled. The system analyzes the intensity (energy), spectral characteristics and acoustic changes after coughing to evaluate the effectiveness of each cough. In particular, the system compares the changes in sputum sound before and after coughing. A significantly weakened sputum sound usually indicates effective sputum expulsion. The system combines these three indicators to calculate the sputum volume estimate: ,in is the weighted acoustic event number, is the negative pressure recovery estimate, It is cough energy, 、 and The weighting coefficient is calibrated using clinical data. The system also implements a time-accumulation mechanism to track cumulative sputum output throughout treatment and compare it against the target to assess progress. Ultimately, the system outputs a standardized sputum output estimate, providing an objective basis for treatment efficacy assessment and parameter optimization.

[0146] Step S6013: Comprehensively consider the fluctuation of physiological indicators, discomfort reaction detection and subjective discomfort score, calculate the basic score minus the weighted influence of various discomfort factors, and obtain the patient comfort index.

[0147] In this step, the system comprehensively evaluates the patient's comfort during treatment, which is a key factor in treatment compliance and long-term effectiveness. The system adopts a multi-dimensional evaluation method, combining objective physiological indicators and subjective feedback. First, the system monitors the fluctuation of key physiological indicators, including heart rate (HR), blood oxygen saturation, , respiratory rate (RR) and blood pressure (BP), etc. Abnormal fluctuations in these indicators usually reflect physiological stress reactions and may indicate patient discomfort. The system calculates the fluctuation of each indicator, which is defined as the ratio of the maximum deviation from the baseline during treatment to the baseline value. For example, the fluctuation of heart rate is . The system pays special attention to sharp fluctuations in a short period of time, which are usually associated with acute discomfort. Secondly, the system detects the patient's discomfort response in a variety of ways. For patients who can cooperate, the system may be equipped with a simple feedback button that allows the patient to mark when they feel uncomfortable. For all patients, the system monitors the frequency of body movements through an accelerometer. Frequent body movements usually indicate discomfort or irritability. For systems equipped with cameras, facial expression analysis technology can also be used to identify expressions of pain or discomfort, such as frowning and drooping corners of the mouth. The system also analyzes changes in breathing patterns. Irregular or rapid breathing may indicate discomfort. Third, the system integrates subjective discomfort scores entered by medical staff or the patients themselves. This usually uses a 0-10 numerical rating scale (NRS) or a simplified emoticon scale to regularly record the patient's subjective feelings. The system combines these three types of information to calculate a comfort index:

[0148]

[0149] in, is the basic comfort score (usually set to 100), It is a comprehensive index of fluctuations in physiological indicators. is the discomfort reaction index, It is a subjective rating. 、 and is the weighting coefficient. The system also implements an individualized calibration mechanism, learning from specific patient response patterns and adjusting the weights and thresholds of various indicators to improve the accuracy of comfort assessment. The final comfort index output, ranging from 0 to 100, intuitively represents the patient's comfort level and provides an important reference for adjusting treatment parameters, ensuring that the patient's comfort experience is maintained while pursuing treatment effectiveness.

[0150] Step S6014: Compare the current status in the real-time feedback data with the treatment goal, track the changes in the treatment status of each lung area and the degree of goal achievement, evaluate the treatment progress speed per unit time, and obtain the treatment progress index.

[0151] In this step, the system accurately tracks and evaluates the progress of treatment, providing a basis for adjusting treatment strategies. First, the system determines specific treatment goals based on the initial assessment, including target improvement values ​​for sputum density and viscosity in each lung region. These goals are usually expressed as a percentage improvement relative to the initial state, such as "sputum density in the right lower lung region decreased by 70%." The system divides the lungs into multiple anatomical regions (usually 6-8 regions), sets goals for each region, and tracks progress. For each lung region, the system calculates the gap between the current state and the target state in real time: , where S represents the status indicator (such as sputum density or viscosity) and i represents the lung zone number. This formula normalizes the gap to a range of 0-1, where 1 means no improvement (current = initial) and 0 means the goal has been fully achieved. The system draws a progress map for each lung zone and uses color coding to intuitively display the treatment progress of different areas, helping medical staff quickly identify areas that need to be focused on. The system also tracks the time trend of the status of each lung zone and evaluates the rate of improvement through linear or nonlinear regression analysis. For areas with slow progress or stagnation, the system will mark them as "areas of concern", indicating that the treatment strategy may need to be adjusted. To evaluate the overall treatment efficiency, the system calculates the rate of progress per unit time: ,in It is the improvement of progress within a certain period of time. is the time interval. The system compares this rate with the expected rate to assess whether the treatment is progressing as planned. The system also implements a treatment time estimation function, which predicts the remaining time required to reach the target based on the current progress rate. To comprehensively evaluate the overall progress of multiple lung regions, the system calculates a weighted average progress index: ,in The weighting of each lung region is typically determined based on regional importance and initial severity. The system also considers treatment time efficiency, incorporating the ratio of actual to estimated treatment time into its assessment. Ultimately, the system outputs standardized treatment progress indicators that visually demonstrate the overall completion and efficiency of treatment, providing a critical reference for optimizing treatment parameters and allocating resources.

[0152] Step S6015: Based on the sputum looseness index, the sputum discharge volume estimation index, the patient comfort index and the treatment progress index, a comprehensive effect score is constructed by weighted summation and multi-objective balance to obtain the short-term sputum cough effect index.

[0153] In this step, the system integrates the individual indicators obtained in the previous steps into a comprehensive short-term sputum cough effect indicator to provide a unified evaluation standard for the reward function of reinforcement learning. This integration process requires balancing multiple potentially conflicting goals, such as the trade-off between treatment effect and patient comfort. First, the system standardizes each individual indicator to ensure that they are in the same numerical range (usually 0-100) for easy comparison and combination. Standardization uses Min-Max normalization or Z-score normalization, and the appropriate method is selected based on the distribution characteristics of historical data. Then, the system applies a weighted summation method to construct a basic comprehensive score: , where L is the sputum looseness index, is the sputum output estimation index, C is the patient comfort index, P is the treatment progress index, 、 、 and are weight coefficients. These weights reflect the relative importance of each indicator in the overall evaluation and are usually determined through clinical expert opinions and data analysis. The system implements a dynamic weight adjustment mechanism to adjust the weight of each indicator according to the treatment stage and patient condition. For example, in the early stage of treatment, more emphasis may be placed on loosening sputum, and the weight may be Higher; in the later stage, more attention may be paid to sputum discharge, and the weight To deal with the nonlinear relationship and interaction between indicators, the system also introduces cross terms, such as (The synergistic effect of loosening and draining). The system pays special attention to short-term effect changes and adopts a time-weighted mechanism, with recent effect changes having higher weights: , where γ is the time discount factor (typically 0.8-0.95), t is the time step size, is the effect score at time t. This mechanism enables the system to quickly respond to recent changes in treatment effects. The system also implements a multi-objective balance mechanism to ensure that extreme improvements in one indicator do not mask the shortcomings of other indicators. The specific implementation uses a nonlinear combination function or a threshold-based penalty term. For example, when the comfort level C is below a certain threshold, no matter how high the other indicators are, the comprehensive score will be significantly penalized. To evaluate the statistical significance of the effect change, the system performs paired t-tests or non-parametric tests on key indicators to ensure that the observed changes are not random fluctuations. Ultimately, the system outputs a comprehensive short-term cough effect indicator, which comprehensively reflects the immediate effect of the treatment, including multiple dimensions such as sputum loosening, sputum volume, patient comfort, and treatment progress. This indicator is directly used to construct the reward function of reinforcement learning to guide the system to learn the optimal treatment strategy.

[0154] See also Figure 2 , Figure 2 This is a block diagram of the structure of the cough assist control device provided by an embodiment of the present invention. The device includes:

[0155] An audio signal acquisition module 701 is configured to acquire audio signals of breathing sounds and coughing sounds from multiple locations in the patient's chest cavity using an audio sensor array, and to perform noise reduction, segmentation, and feature extraction on the audio signals to obtain a standardized audio feature dataset.

[0156] Sputum characteristic analysis module 702 is configured to construct a dynamic weighted graph model based on the standardized audio feature dataset, apply a random edge sampling algorithm to estimate the average degree of the dynamic weighted graph model, and analyze the spectral characteristic changes of the cough sounds in the standardized audio feature dataset using a parallel batch dynamic algorithm to obtain a sputum viscosity index and a sputum distribution location map;

[0157] A vibration scheme optimization module 703 is configured to map the vibration parameter space into an unweighted disk map based on the sputum viscosity index and the sputum distribution location map, and apply a shortest path algorithm to calculate the optimal vibration frequency, amplitude, and time parameter combination to obtain a personalized vibration treatment plan;

[0158] The multimodal control module 704 is configured to synchronously control the operating parameters of the acoustic vibrator and the negative pressure suction device according to the personalized vibration therapy plan and in combination with the patient's real-time respiratory cycle to obtain coordinated and consistent multimodal therapy execution instructions;

[0159] The parameter optimization module 705 is used to collect real-time feedback data including audio change data and patient physiological indicators during the treatment process, compare and analyze them with the multimodal treatment execution instructions, and dynamically adjust the working parameters based on the reinforcement learning algorithm to obtain an optimized cough-assisted treatment plan.

[0160] The specific implementation process of the cough assisting control method of the present invention is as follows:

[0161] 1. Breathing and coughing audio signal collection and preprocessing:

[0162] The patient's breathing and coughing sound signals are collected based on an audio sensor array. After preprocessing such as signal noise reduction, segmentation, and feature extraction, a standardized breathing and coughing audio feature dataset is output as the basis for subsequent analysis.

[0163] Specifically, based on the patient's chest anatomical structure, a highly sensitive MEMS microphone array is deployed in the chest, lateral chest, and back regions. The acquisition position is optimized through the sensor array topology to output a multi-channel raw audio signal stream. The multi-channel raw audio signal stream is received, and a method combining an adaptive filtering algorithm and a wavelet transform is applied to identify and filter out environmental noise, medical equipment interference, and patient body movement noise, outputting a high signal-to-noise ratio breathing and coughing audio signal. Based on the high signal-to-noise ratio audio signal, the time-domain-frequency domain feature analysis and energy threshold detection algorithm are used to automatically identify and segment respiratory cycles, coughing events, and quiet periods, and output a time-stamped audio event sequence. For the time-stamped audio event sequence, multidimensional acoustic features such as spectral energy distribution, MFCC coefficients, harmonic structure, and time domain features are extracted, and the most discriminative feature set is screened through a feature selection algorithm to output a standardized respiratory sound and cough audio feature dataset.

[0164] 2. Sputum characteristics analysis based on graph theory:

[0165] The audio feature data output in step 1 is constructed into a dynamic weighted graph model. The improved random edge sampling algorithm is applied to estimate the average degree of the graph. The parallel batch dynamic algorithm is used to perform graph structure analysis and output the sputum viscosity index and distribution location map.

[0166] Specifically, based on a standardized audio feature dataset, acoustic features collected at different locations are used as graph nodes, and the acoustic correlation between features is used as edge weights to construct a dynamic weighted graph model that characterizes the distribution of sputum in the respiratory tract. An improved random edge sampling algorithm is applied to efficiently estimate the average degree of different regions from the constructed graph model, and the average degree is mapped to a sputum distribution density index, outputting sputum density distribution maps of different lung regions. Based on the sputum density distribution map, a parallel batch dynamic algorithm is applied to process the spectral feature changes of cough sounds, establish a mapping relationship between spectral features and sputum viscosity, and output regional sputum viscosity index. The sputum density distribution and viscosity index are integrated, combined with the patient's lung anatomical structure parameters, and a three-dimensional spatial interpolation algorithm is used to generate an accurate sputum distribution location map, marking key treatment areas and priorities.

[0167] 3. Intelligent optimization of acoustic vibration scheme:

[0168] Based on the sputum characteristic indicators output in step 2, a shortest path model of the relationship between vibration effect and sputum loosening is constructed. The unweighted disk graph shortest path algorithm is applied to calculate the optimal combination of vibration frequency, amplitude, and time parameters, and output a personalized vibration treatment plan.

[0169] Specifically, based on the sputum distribution location map and viscosity index, a relationship model between acoustic vibration parameters (frequency, amplitude, action time) and sputum loosening effect is constructed to form a decision diagram for the multidimensional parameter space. The vibration parameter space is mapped into an unweighted disk graph, with the vibration parameter combinations as points in the graph and the vibration effects as the distances between the points. The optimal shortest path algorithm is applied to calculate the optimal parameter path from the initial state to the target sputum loosening effect. Based on the optimal parameter path and combined with the patient tolerance model, a curve of the vibration parameter changes over time is designed using a dynamic programming algorithm to ensure that the treatment effect is maximized while minimizing patient discomfort. By integrating the vibration parameter dynamic curve and the patient's sputum distribution map, a customized vibration treatment plan for different lung regions is generated, including the regional action sequence, action time, and vibration parameter configuration.

[0170] 4. Multimodal treatment collaborative control:

[0171] According to the vibration treatment plan output in step 3, combined with the patient's real-time physiological state, the working parameters of the ultrasonic vibrator and the negative pressure suction device are synchronously controlled to achieve precise coordination of vibration and negative pressure suction, and output coordinated and consistent multimodal treatment execution instructions.

[0172] Specifically, according to the personalized vibration therapy plan, the vibration parameters are converted into driving signals, and the operating frequency, amplitude and phase of the multi-channel vibrator are precisely controlled to achieve precise vibration effects on the target lung area. Based on the vibration therapy plan and the patient's respiratory cycle, the negative pressure suction parameters coordinated with the vibration parameters are calculated, including the negative pressure intensity, suction time window and pressure change curve, and the negative pressure control instructions are output. By monitoring the patient's respiratory phase in real time and using a prediction algorithm to estimate the optimal time for intervention, the start and stop timing of the vibrator and the negative pressure suction device are precisely controlled to achieve optimal synchronization with the patient's respiratory rhythm. The vibrator drive signal, negative pressure control instruction and respiratory synchronization signal are integrated to generate an execution timing table for precise coordination between devices, ensuring that each modality of treatment forms a synergistic force and outputs coordinated multimodal treatment execution instructions.

[0173] 5. Treatment effect evaluation and dynamic parameter optimization:

[0174] Collect real-time feedback data during the treatment process (including audio changes, patient physiological indicators, etc.), compare and analyze it with the treatment execution instructions in step 4, evaluate the treatment effect, and dynamically adjust and optimize the treatment parameters based on the reinforcement learning algorithm to output an updated treatment plan.

[0175] Specifically, during the treatment process, audio change data, patient physiological indicators (such as blood oxygen saturation and respiratory rate), and device execution parameters are continuously collected to construct a multidimensional treatment process data stream. Based on the multidimensional treatment process data stream, a sliding window analysis method is applied to calculate the coughing effect indicators within a short time window, including sputum looseness, discharge volume estimation, and patient comfort score. Using the treatment parameters as the action space and the coughing effect indicators as the reward function, a parameter optimization model based on reinforcement learning is established, and the treatment strategy is continuously optimized through experience replay and policy gradient methods. Based on the output of the reinforcement learning model, the vibration and negative pressure parameters are adjusted in real time, and the treatment sequence and time allocation are updated to generate an optimized treatment plan adapted to the patient's current state and feedback to the system execution module.

[0176] The coughing auxiliary control method and device of the present invention collects audio signals of breathing sounds and coughing sounds at multiple positions of the patient's chest cavity based on an audio sensor array, performs noise reduction, segmentation and feature extraction on the audio signals to obtain a standardized audio feature data set; constructs a dynamic weighted graph model based on the standardized audio feature data set, applies a random edge sampling algorithm to estimate the average degree of the dynamic weighted graph model, analyzes the spectral feature changes of the coughing sounds in the standardized audio feature data set through a parallel batch dynamic algorithm, and obtains a sputum viscosity index and a sputum distribution position map; based on the sputum viscosity index and sputum Liquid distribution position map, the vibration parameter space is mapped into an unweighted disk map, the shortest path algorithm is applied to calculate the optimal vibration frequency, amplitude and time parameter combination, and a personalized vibration treatment plan is obtained; according to the personalized vibration treatment plan, combined with the patient's real-time breathing cycle, the working parameters of the acoustic vibrator and the negative pressure suction device are synchronously controlled to obtain a coordinated and consistent multimodal treatment execution instruction; real-time feedback data including audio change data and patient physiological indicators during the treatment process are collected, compared and analyzed with the multimodal treatment execution instruction, and the working parameters are dynamically adjusted based on the reinforcement learning algorithm to obtain an optimized cough auxiliary treatment plan. Through the above technical solution, the present invention can make intelligent adjustments based on the patient's real-time respiratory status and sputum characteristics, accurately analyze the sputum characteristics, and form an optimal treatment strategy through the synergy between the various components of the system, thereby improving cough efficiency and reducing patient discomfort.

[0177] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0178] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of patent protection for the present application shall be determined by the appended claims.

Claims

1. A method for assisting in controlling sputum cough, characterized in that: include: The audio sensor array collects audio signals of breathing sounds and coughing sounds at multiple locations in the patient's chest cavity, performs noise reduction, segmentation, and feature extraction on the audio signals, and obtains a standardized audio feature dataset; A dynamic weighted graph model is constructed based on the standardized audio feature dataset, a random edge sampling algorithm is applied to estimate the average degree of the dynamic weighted graph model, and changes in the spectral characteristics of cough sounds in the standardized audio feature dataset are analyzed using a parallel batch dynamic algorithm to obtain a sputum viscosity index and a sputum distribution location map; Based on the sputum viscosity index and sputum distribution location map, the vibration parameter space is mapped into an unweighted disk map, and the shortest path algorithm is applied to calculate the optimal vibration frequency, amplitude and time parameter combination to obtain a personalized vibration treatment plan; According to the personalized vibration therapy plan, combined with the patient's real-time breathing cycle, the working parameters of the sonic vibrator and the negative pressure suction device are synchronously controlled to obtain coordinated and consistent multimodal therapy execution instructions; Real-time feedback data including audio change data and patient physiological indicators are collected during the treatment process, compared and analyzed with the multimodal treatment execution instructions, and the working parameters are dynamically adjusted based on the reinforcement learning algorithm to obtain an optimized cough-assisted treatment plan.

2. The method according to claim 1, characterized in that The step of performing noise reduction, segmentation, and feature extraction on the audio signal to obtain a standardized audio feature dataset includes: Based on the audio signals of the breathing sounds and coughing sounds, an adaptive filtering algorithm and wavelet transform are applied to filter out environmental noise and medical equipment interference to obtain an audio signal with a high signal-to-noise ratio; Based on the high signal-to-noise ratio audio signal, using time-domain-frequency domain feature analysis and energy threshold detection algorithm to identify respiratory cycles, cough events and quiet periods, and obtain a time-stamped audio event sequence; Based on the time-tagged audio event sequence, spectral energy distribution, MFCC coefficients, harmonic structure and time domain features are extracted, and the most discriminative feature set is screened through a feature selection algorithm to obtain the standardized audio feature dataset.

3. The method according to claim 1, characterized in that Applying a random edge sampling algorithm to estimate the average degree of the dynamic weighted graph model includes: Based on the dynamic weighted graph model, edge samples are randomly selected from the edge set using importance sampling technology, so that the selection probability of the edge is proportional to the weight of the edge sample, thereby obtaining a random edge sample set; Based on the random edge sample set, the local degrees of the nodes at both ends of each edge are calculated, and the global average degree of the dynamic weighted graph model is estimated by summation operation to obtain a sputum distribution density index; Based on the sputum distribution density index, the lungs are divided into multiple anatomical regions and the average degree in each region is calculated to obtain sputum density distribution maps of different lung regions.

4. The method according to claim 3, characterized in that The analyzing the spectral feature changes of the cough sound in the standardized audio feature dataset by a parallel batch dynamic algorithm includes: Based on the sputum density distribution map, the cough sounds in the standardized audio feature dataset are grouped according to lung regions and corresponding sub-graphs are constructed to obtain acoustic characteristic representations of each region; For the acoustic characteristic representation of each region, a parallel computing architecture is used to simultaneously extract spectral energy distribution, harmonic components, and time-frequency characteristics to obtain a multi-dimensional acoustic feature vector; Based on the multidimensional acoustic feature vector, a mapping relationship between the spectrum feature and the sputum viscosity is established, and the viscosity index of each region is calculated to obtain the sputum viscosity index; The sputum density distribution map and the sputum viscosity index are integrated, combined with the patient's lung anatomical structure parameters, and an accurate sputum distribution position map is generated through a three-dimensional spatial interpolation algorithm.

5. The method according to claim 1, characterized in that The application of the shortest path algorithm to calculate the optimal vibration frequency, amplitude and time parameter combination includes: Based on the sputum viscosity index and the sputum distribution position map, the frequency parameter and the amplitude parameter are mapped into plane coordinates, and the time parameter is mapped into the size of the point to obtain an unweighted disk map; Based on the unweighted disc graph, the distance between two points is defined as a weighted combination of the inverse of the effect gain and the patient's tolerance, and the Dijkstra algorithm is applied to search for the shortest path from the initial state to the target state to obtain the optimal parameter path; Based on the optimal parameter path, a dynamic programming algorithm is applied to design a curve showing the change of vibration parameters over time, and the time allocation is optimized in combination with the patient tolerance model to obtain the personalized vibration treatment plan.

6. The method according to claim 1, characterized in that According to the personalized vibration therapy program, combined with the patient's real-time breathing cycle, the working parameters of the sonic vibrator and the negative pressure suction device are synchronously controlled to obtain a coordinated multimodal therapy execution instruction, including: Based on the personalized vibration therapy plan, the vibration parameters are converted into driving signals, the operating frequency, amplitude and phase of the acoustic vibrator are controlled, and the vibrator driving instructions are obtained; Based on the personalized vibration therapy program and the patient's real-time breathing cycle, the negative pressure intensity, the suction time window, and the pressure change curve coordinated with the vibration parameters are calculated to obtain a negative pressure control instruction; By monitoring the patient's respiratory phase in real time, using a prediction algorithm to estimate the optimal intervention time, the start and stop sequence of the sonic vibrator and the negative pressure suction device is controlled to obtain a respiratory synchronization signal; Based on the vibrator drive instruction, the negative pressure control instruction and the respiratory synchronization signal, a device collaboration execution timing table is generated to obtain a device collaboration control strategy; The vibrator driving instruction, the negative pressure control instruction, the respiratory synchronization signal and the device collaborative control strategy are integrated to ensure that the acoustic vibration treatment and the negative pressure suction device treatment cooperate with each other and obtain the coordinated multimodal treatment execution instruction.

7. The method according to claim 6, characterized in that The method includes: monitoring the patient's respiratory phase in real time, using a prediction algorithm to estimate the optimal intervention time, controlling the start and stop timing of the acoustic vibrator and the negative pressure suction device, and obtaining a respiratory synchronization signal, including: By fusing the chest and abdominal motion sensors, airflow sensors, and sound sensors to collect respiratory signals, the respiratory cycle is divided into four phases: early inspiration, late inspiration, early expiration, and late expiration, and the respiratory phase recognition results are obtained. Based on the respiratory phase identification result, wavelet transform and peak detection algorithm are applied to capture the phase transition point, and the phase time point of the next cycle is predicted in combination with the pattern of the last 3-5 respiratory cycles to obtain the respiratory phase prediction result; Based on the respiratory phase prediction results, the optimal intervention time is determined to start vibration therapy at the early stage of exhalation and implement negative pressure suction in the middle and late stages of exhalation, and the time difference between vibration and negative pressure is designed to obtain the respiratory synchronization signal.

8. The method according to claim 1, characterized in that Dynamically adjusting the working parameters based on a reinforcement learning algorithm to obtain an optimized cough and sputum auxiliary treatment plan includes: The patient's physiological state, sputum characteristics, and treatment progress are used as the state space, the working parameters and treatment mode selection are used as the action space, and a reward function is constructed based on the short-term sputum cough effect index to obtain a formal representation of the reinforcement learning problem; Based on the formal representation of the reinforcement learning problem, the policy gradient method and the neural network parameterization strategy are used to optimize the treatment strategy through gradient estimation and parameter update to obtain the initial policy network; Maintaining an experience replay buffer to store treatment process data in the real-time feedback data, prioritizing sampling of important experiences based on the size of the temporal difference error, performing batch learning and parameter updates, and continuously updating the initial strategy network parameters to obtain an optimized strategy network; Design an exploration strategy with parameter noise and entropy regularization, conduct planned exploration within a safe parameter range, adjust the exploration degree according to the uncertainty estimated by the value function, and obtain an exploration-exploitation balance strategy; Based on the optimal strategy network and the exploration-exploitation balance strategy, the recommended parameters are calculated according to the current patient status in the real-time feedback data, and the working parameters are dynamically adjusted by applying safety constraint checking and smoothing processing to obtain the optimized cough-assisted treatment plan.

9. The method according to claim 8, characterized in that The reward function is constructed based on the short-term coughing effect indicator, including: Using overlapping sliding windows to continuously analyze the treatment process data in the real-time feedback data, the degree of sputum loosening is assessed by the energy ratio of breath sounds, changes in sputum sounds, and improvement in sound conduction, to obtain a sputum loosening index; Based on acoustic event statistics, negative pressure system feedback, and cough effectiveness evaluation, the sputum output estimation index is obtained by weighted summation of acoustic event counts, negative pressure recovery estimation, and cough energy. The patient comfort index is obtained by combining the fluctuation of physiological indicators, discomfort reaction detection and subjective discomfort score, and calculating the basic score minus the weighted influence of various discomfort factors; Comparing the gap between the current status in the real-time feedback data and the treatment goal, tracking the changes in the treatment status of each lung area and the degree of goal achievement, evaluating the treatment progress speed per unit time, and obtaining a treatment progress indicator; Based on the sputum looseness index, the sputum discharge volume estimation index, the patient comfort index and the treatment progress index, a comprehensive effect score is constructed by weighted summation and multi-objective balance to obtain the short-term sputum cough effect index.

10. A cough assisting control device, characterized in that: include: An audio signal acquisition module is used to collect audio signals of breathing sounds and coughing sounds at multiple locations in the patient's chest cavity based on an audio sensor array, and to perform noise reduction, segmentation, and feature extraction on the audio signals to obtain a standardized audio feature dataset; a sputum characteristic analysis module, configured to construct a dynamic weighted graph model based on the standardized audio feature dataset, estimate the average degree of the dynamic weighted graph model using a random edge sampling algorithm, and analyze the spectral characteristic changes of the cough sounds in the standardized audio feature dataset using a parallel batch dynamic algorithm to obtain a sputum viscosity index and a sputum distribution location map; a vibration scheme optimization module for mapping the vibration parameter space into an unweighted disk map based on the sputum viscosity index and the sputum distribution location map, and applying a shortest path algorithm to calculate the optimal vibration frequency, amplitude, and time parameter combination to obtain a personalized vibration treatment plan; A multimodal control module, configured to synchronously control the operating parameters of the acoustic vibrator and the negative pressure suction device according to the personalized vibration therapy plan and in combination with the patient's real-time respiratory cycle, to obtain coordinated and consistent multimodal therapy execution instructions; The parameter optimization module is used to collect real-time feedback data including audio change data and patient physiological indicators during the treatment process, compare and analyze them with the multimodal treatment execution instructions, and dynamically adjust the working parameters based on the reinforcement learning algorithm to obtain an optimized cough-assisted treatment plan.

Images