Respiratory tract phlegm blockage monitoring method and device

Abstract

The invention provides a respiratory tract phlegm blockage monitoring method and device, and belongs to the technical field of respiratory tract phlegm blockage monitoring, and the monitoring method specifically comprises the following steps: obtaining respiratory tract multi-mode biophysical signals of a target domain patient and a plurality of source domain patients; the method comprises the following steps: acquiring a target domain signal through a micro sensor array, and constructing an optimal source domain set by using a signal processing algorithm formed by a coding module and a classification module; inputting a target domain signal and the optimal source domain set into a global adapter, extracting shallow layer features through a feature extraction module, and realizing conditional distribution alignment through a global adaptation unit; edge distribution alignment is guided through adversarial training containing a gradient inversion layer; and the discriminator and the global adapter parameters are fixed, and a classifier is trained to identify and evaluate the sputum blockage state. According to the method, the sputum blockage monitoring accuracy is improved through multi-mode signal processing and confrontation training.

Description

Technical Field

[0001] This invention relates to the field of respiratory tract sputum obstruction monitoring technology, specifically to a method and device for monitoring respiratory tract sputum obstruction. Background Technology

[0002] Significant progress has been made in airway sputum blockage monitoring technology in clinical scenarios such as intensive care, anesthesia management, and long-term respiratory support. However, due to the dynamic and localized characteristics of airway sputum accumulation and blockage, existing monitoring methods still face challenges in terms of real-time performance and accuracy. For example, a disposable sterile double-tube nasopharyngeal airway, through optimized structural design, independently separates the ventilation slot and the monitoring sampling slot within the silicone tubing. This achieves physical separation between high-flow oxygenation or anesthesia machine ventilation and end-tidal carbon dioxide monitoring, significantly reducing mutual interference between the two and improving the accuracy of monitoring data. Furthermore, this approach, by enlarging the monitoring sampling slot and supplementing it with a side hole structure, further reduces the risk of blockage caused by "sputum and condensation" or "sputum and moisture condensation and accumulation," demonstrating a design approach that mitigates the impact of sputum blockage through physical structural optimization.

[0003] The intelligent respiratory support and airway clearance assist system utilizes advanced information technology and data processing algorithms. By integrating a multi-source data acquisition module, a respiratory analysis unit, a lung imaging data analysis unit, and a ventilator waveform analysis unit, it performs multi-dimensional intelligent monitoring of the patient's respiratory status. Based on the deep fusion and analysis of multi-source data, the system can help identify the potential risk of sputum blockage and determine whether to activate the sputum clearance mode according to preset logic, providing decision support for clinicians. Its advantage lies in its ability to identify the chain reactions that sputum blockage may trigger at the macroscopic physiological data level, providing the possibility for automated assisted intervention.

[0004] Patent CN107569726B discloses a closed suction catheter, including a T-type three-way valve, an injection tube device on the insertion end of the suction catheter connected to the T-type three-way valve, a solenoid valve connected to the injection tube insertion end, and a controller. An annular flow meter is installed at the head of the conical end of the insertion end of the suction catheter. The annular flow meter is a thick annular shape with four radially arranged through holes between the inner and outer circles. A pressure sensor is installed between any two adjacent through holes, and the signal line of the pressure sensor in the annular flow meter is connected to the controller. An anti-backflow device is embedded in the suction catheter body interface connected to the T-type three-way valve. This allows for monitoring of sputum flow during suctioning, controlling the opening of the solenoid valve, and adding medication or flushing solution to dilute difficult-to-suction sputum.

[0005] Patent CN109646004A discloses an intelligent sputum sensing and monitoring device, comprising a sputum sensor device, an A / D data acquisition module, a sputum analysis and processing module, a control module, and a terminal device. The sputum sensor device includes a sensor extension tube and a sensor module. The sensor module includes a CCD miniature image sensor and a sound sensor located at the front end of the sensor extension tube, and an airflow sensor located at the rear end of the sensor extension tube. The sensor module monitors the characteristics of the patient's sputum in real time. After processing by the sputum analysis and processing module, the control module promptly sends instructions to the terminal device, allowing medical staff or family members to take appropriate action upon receiving the information.

[0006] As clinical needs evolve towards greater precision and more timely responses, the aforementioned technical solutions are gradually showing certain limitations in addressing new challenges.

[0007] Specifically, solutions focused on structural optimization emphasize "prevention" and "relief," lacking the ability to perceive the actual accumulation of sputum and the degree of blockage in real time. This means that in the early stages of abnormal sputum accumulation, the device struggles to provide early warning signals; its passive effects only become apparent when the blockage reaches a certain level and affects ventilation or sampling functions. This lag may lead to delays in clinical intervention, increasing the risk of complications for patients, and hindering the precise management of the progression of sputum blockage.

[0008] For intelligent systems, the identification of sputum blockage mainly relies on inferential analysis of indirect indicators such as external imaging data (e.g., X-rays, CT scans) and ventilator waveform data. This indirect data reflects the macroscopic impact of sputum blockage on lung physiological function or respiratory dynamics, rather than directly capturing the physical accumulation of sputum in a localized location in the airway. This indirect analysis model has certain limitations: First, imaging examinations are usually intermittent, making it difficult to provide continuous real-time monitoring; second, changes in ventilator waveforms can be affected by various factors, such as patient position, compliance, and other lung diseases, leading to insufficient specificity in sputum blockage diagnosis and the possibility of misdiagnosis or missed diagnosis. Furthermore, neither imaging data nor ventilator waveforms can provide high-resolution localization of the specific location and degree of sputum blockage, while accurate localization is crucial for guiding targeted suction or other local interventions. Simultaneously, the complex algorithm models and high-end hardware configurations involved in this system inevitably lead to high R&D, deployment, and maintenance costs, and place high demands on the professional skills of operators, creating obstacles to promotion and widespread adoption in primary healthcare institutions and resource-constrained environments.

[0009] In summary, existing respiratory tract sputum blockage monitoring technologies present a dilemma: one type of solution, while simple in structure and low in cost, has limited functionality and lacks active, real-time dynamic monitoring capabilities; the other type, while possessing a certain level of intelligence, relies on indirect data, making it difficult to accurately capture the location and extent of sputum blockage, and is also complex and costly. The fundamental limitation of these solutions lies in the fact that their technical principles all rely on monitoring secondary effects caused by sputum blockage (such as changes in flow resistance and imaging alterations), rather than directly and in situ measuring the physical presence of the sputum itself. This leads to an irreconcilable contradiction between real-time performance, specificity, and precise localization. Summary of the Invention

[0010] The purpose of this invention is to provide a method and device for monitoring respiratory tract sputum blockage, which effectively solves the contradiction between real-time performance, accurate positioning and cost-effectiveness in the monitoring of respiratory tract sputum accumulation and blockage in the prior art. By introducing multimodal biophysical signal data samples and global adaptation units, the sensitivity and accuracy of sputum blockage status identification are improved, while reducing the complexity of the system and the difficulty of promotion.

[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0012] A method for monitoring airway sputum blockage includes the following steps:

[0013] S101: Acquire respiratory multimodal biophysical signal data samples from patients in the target domain and multiple source domains, respectively. The multimodal biophysical signal data samples include ultrasound reflection characteristics, bioelectrical impedance changes, and optical scattering characteristics.

[0014] S102: Collect respiratory signal data of patients in the target domain through a miniature sensor array, and select signal data samples from multiple source domain patients using a preset signal processing algorithm to construct an optimal source domain set. The signal processing algorithm consists of an encoding module and a classification module. Both the encoding module and the classification module consist of three fully connected layers, and each source domain uses a separately trained signal processing algorithm with non-shared parameters.

[0015] S103: Input the signal data of the patient in the target domain and its corresponding optimal source domain set into the global adapter. The global adapter includes a feature extraction module and a global adaptation unit based on a multimodal fusion mechanism. The feature extraction module extracts shallow features of each signal data sample in the target domain and the optimal source domain set respectively, and the global adaptation unit guides the alignment of the conditional distribution between the target domain and the optimal source domain set.

[0016] S104: The shallow features after conditional distribution alignment are used as input to the classifier and discriminator, and the edge distribution between the target domain and the source domain is aligned through adversarial training. The adversarial training uses a gradient inversion layer to force the feature distribution to converge.

[0017] S105: Fix the parameters of the discriminator and the global adapter, and train the classifier to identify and evaluate the sputum blockage state based on the input signal data.

[0018] Furthermore, the feature extraction module described in S103 is a three-tiered dual-path network. The three-tiered dual-path network includes two paths corresponding to the source domain and the target domain, respectively. The signal data samples input to the two paths are of the same size. Each path includes an input layer, a frequency domain convolutional layer, a time domain convolutional layer, a channel convolutional layer, an average pooling layer, and a projection layer connected in sequence. The projection layer is used to project the extracted features into a single-channel embedding sequence with the same dimension, which serves as the shallow feature finally obtained by the feature extraction module.

[0019] Furthermore, in S103, the conditional distribution alignment between the target domain and the optimal source domain set is guided by the global adaptation unit, including:

[0020] The shallow features of the target domain signal data and the shallow features of the signal data in the optimal source domain set are divided into multiple segments of equal length.

[0021] A linear layer is used to generate a key matrix, a query matrix, and a value matrix, where the key matrix and the query matrix are derived from shallow features of signal data in the optimal source domain set, and the value matrix is ​​derived from shallow features of signal data in the target domain.

[0022] Attention queries are performed using multiple attention heads. Each attention head uses a query matrix, a key matrix, and a value matrix to calculate attention weights and generate query results.

[0023] The query results from each attention head are concatenated to obtain shallow features after conditional distribution alignment.

[0024] Furthermore, before inputting the target domain signal data and its corresponding optimal source domain set into the global adapter in S103, a step of enhancing the target domain signal data is also included:

[0025] Generate weight values ​​for the target domain signal data and normalize them. Randomly select two signal data samples and use the normalized weight values ​​to generate a new signal data sample by weighted combination. The label of the new signal data sample adopts the label of the original signal data sample with the larger normalized weight value.

[0026] Furthermore, in S104, when guiding the alignment of the edge distribution between the target domain and the source domain through adversarial training, a gradient inversion layer is used to calculate the gradient penalty. The loss function of adversarial training includes the discriminative loss of the source domain and the target domain, as well as the gradient penalty term. The samples in the random interpolation domain are generated by combining the samples of the source domain and the target domain with random weights.

[0027] Furthermore, the loss function used when training the classifier in S105 includes classification loss and adversarial loss. The classification loss is calculated based on the number of signal data samples and the categories in the source and target domains to determine the prediction error, while the adversarial loss is calculated based on the discriminator's output of shallow features in the source and target domains.

[0028] Furthermore, the process of constructing the optimal source domain set described in S102 includes:

[0029] The source domain signal data is encoded using an encoding module, and then the similarity between each source domain and the target domain is evaluated using a classification module. The source domain signal data samples with the highest similarity are selected to form the optimal source domain set.

[0030] Furthermore, the adversarial training process in S104 includes:

[0031] By backpropagating gradients through a gradient reversal layer, the discriminator is unable to distinguish shallow features between the source and target domains, thereby achieving edge distribution alignment and adversarial training iterations until the loss function converges.

[0032] In addition, the present invention also discloses a respiratory tract sputum obstruction monitoring device, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to perform the respiratory tract sputum obstruction monitoring method described above.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] This invention can issue an early warning at the very early stage of abnormal sputum accumulation (before significant changes in macroscopic physiological parameters occur). Simultaneously, the combination of multimodal signals and sensor arrays enables the system to spatially locate the specific location of the blockage, providing an irreplaceable decision-making basis for precise interventions such as targeted suction in clinical practice. The technical solution provided by this invention relies on biophysical signals directly causally related to sputum, effectively avoiding misjudgments caused by patient position, activity, or other lung diseases. It overcomes the fundamental deficiency of insufficient specificity when analyzing indirect signals such as ventilator waveforms, significantly improving the reliability and clinical trustworthiness of the output results. This invention reduces the need for training data from target patients through an efficient algorithm framework, thereby reducing reliance on expensive, high-power hardware (such as high-end imaging equipment) while maintaining high accuracy. This allows the system to achieve high-performance monitoring with a relatively simple and low-cost configuration, breaking the paradox of "high performance inevitably leading to high complexity and high cost" in existing technologies, greatly improving the accessibility and potential for widespread application in primary healthcare institutions and resource-constrained environments. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the respiratory tract sputum blockage monitoring method of the present invention.

[0037] Figure 2 This is a flowchart of the three-cascade dual-path network of the present invention. Detailed Implementation

[0038] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0039] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0040] Example 1:

[0041] See Figures 1-2 This embodiment discloses a method for monitoring airway sputum blockage, including the following steps:

[0042] S101: Acquire respiratory multimodal biophysical signal data samples from patients in the target domain and multiple source domains, respectively. The multimodal biophysical signal data samples include ultrasound reflection characteristics, bioelectrical impedance changes, and optical scattering characteristics.

[0043] S102: Collect respiratory signal data of patients in the target domain through a miniature sensor array, and select signal data samples from multiple source domain patients using a preset signal processing algorithm to construct an optimal source domain set. The signal processing algorithm consists of an encoding module and a classification module. Both the encoding module and the classification module consist of three fully connected layers, and each source domain uses a separately trained signal processing algorithm with non-shared parameters.

[0044] S103: Input the signal data of the patient in the target domain and its corresponding optimal source domain set into the global adapter. The global adapter includes a feature extraction module and a global adaptation unit based on a multimodal fusion mechanism. The feature extraction module extracts shallow features of each signal data sample in the target domain and the optimal source domain set respectively, and the global adaptation unit guides the alignment of the conditional distribution between the target domain and the optimal source domain set.

[0045] S104: The shallow features after conditional distribution alignment are used as input to the classifier and discriminator, and the edge distribution between the target domain and the source domain is aligned through adversarial training. The adversarial training uses a gradient inversion layer to force the feature distribution to converge.

[0046] S105: Fix the parameters of the discriminator and the global adapter, and train the classifier to identify and evaluate the sputum blockage state based on the input signal data.

[0047] This invention abandons the strategy of inferring sputum blockage indirectly by monitoring secondary physiological effects in existing technologies, and instead directly and multimodally perceives the core physical phenomenon of sputum accumulation online. It directly collects native biophysical signals reflecting the physical characteristics of sputum through a micro-sensor array. These signals are directly related to the intrinsic physical properties of sputum, such as viscosity, density, and dielectric constant, thus ensuring the essential relevance of the data to the target problem from the source. Simultaneously, it introduces a domain adaptation machine learning paradigm to address the model generalization challenge caused by physiological differences between different patients (source and target domains). Through a specially designed global adaptation unit and adversarial training mechanism, this method forces the model to learn patient-independent, deep-level sputum feature representations, achieving effective knowledge transfer from source domain patients with abundant data to target domain patients with scarce data. This ensures the consistency and reliability of the core perception principle across different individuals. Finally, state recognition is completed through a classifier trained with cross-domain adaptation features and inputting directly perceived signals. This forms a closed-loop technology from "physical signal acquisition" to "pathological state assessment," fundamentally improving the determinism and reliability of the output.

[0048] By directly sensing changes in the physical properties of sputum itself, this invention can issue an early warning at the very early stage of abnormal sputum accumulation (before significant changes in macroscopic physiological parameters occur). Simultaneously, the combination of multimodal signals and sensor arrays enables the system to spatially locate the specific location of the blockage, providing an irreplaceable decision-making basis for precise interventions such as targeted suction in clinical practice. The technical solution provided by this invention relies on biophysical signals directly causally related to sputum, effectively avoiding misjudgments caused by patient position, activity, or other lung diseases. It overcomes the fundamental deficiency of insufficient specificity when analyzing indirect signals such as ventilator waveforms, significantly improving the reliability and clinical trustworthiness of the output results. This invention reduces the need for training data on target patients through an efficient algorithmic framework (such as domain adaptation), thereby reducing reliance on expensive, high-power hardware (such as high-end imaging equipment) while maintaining high accuracy. This allows the system to achieve high-performance monitoring with a relatively simple and low-cost configuration, breaking the paradox of "high performance inevitably leading to high complexity and high cost" in existing technologies, greatly improving the accessibility and potential for widespread application of the technology in primary healthcare institutions and resource-constrained environments.

[0049] Furthermore, the feature extraction module described in S103 is a three-tiered dual-path network. The three-tiered dual-path network includes two paths corresponding to the source domain and the target domain, respectively. The signal data samples input to the two paths are of the same size. Each path includes an input layer, a frequency domain convolutional layer, a time domain convolutional layer, a channel convolutional layer, an average pooling layer, and a projection layer connected in sequence. The projection layer is used to project the extracted features into a single-channel embedding sequence with the same dimension, which serves as the shallow feature finally obtained by the feature extraction module.

[0050] Furthermore, in S103, the conditional distribution alignment between the target domain and the optimal source domain set is guided by the global adaptation unit, including:

[0051] The shallow features of the target domain signal data and the shallow features of the signal data in the optimal source domain set are divided into multiple segments of equal length.

[0052] A linear layer is used to generate a key matrix, a query matrix, and a value matrix, where the key matrix and the query matrix are derived from shallow features of signal data in the optimal source domain set, and the value matrix is ​​derived from shallow features of signal data in the target domain.

[0053] Attention queries are performed using multiple attention heads. Each attention head uses a query matrix, a key matrix, and a value matrix to calculate attention weights and generate query results.

[0054] The query results from each attention head are concatenated to obtain shallow features after conditional distribution alignment.

[0055] Furthermore, before inputting the target domain signal data and its corresponding optimal source domain set into the global adapter in S103, a step of enhancing the target domain signal data is also included:

[0056] Generate weight values ​​for the target domain signal data and normalize them. Randomly select two signal data samples and use the normalized weight values ​​to generate a new signal data sample by weighted combination. The label of the new signal data sample adopts the label of the original signal data sample with the larger normalized weight value.

[0057] Furthermore, in S104, when guiding the alignment of the edge distribution between the target domain and the source domain through adversarial training, a gradient inversion layer is used to calculate the gradient penalty. The loss function of adversarial training includes the discriminative loss of the source domain and the target domain, as well as the gradient penalty term. The samples in the random interpolation domain are generated by combining the samples of the source domain and the target domain with random weights.

[0058] Furthermore, the loss function used when training the classifier in S105 includes classification loss and adversarial loss. The classification loss is calculated based on the number of signal data samples and the categories in the source and target domains to determine the prediction error, while the adversarial loss is calculated based on the discriminator's output of shallow features in the source and target domains.

[0059] Furthermore, the process of constructing the optimal source domain set described in S102 includes:

[0060] The source domain signal data is encoded using an encoding module, and then the similarity between each source domain and the target domain is evaluated using a classification module. The source domain signal data samples with the highest similarity are selected to form the optimal source domain set.

[0061] Furthermore, the adversarial training process in S104 includes:

[0062] By backpropagating gradients through a gradient reversal layer, the discriminator is unable to distinguish shallow features between the source and target domains, thereby achieving edge distribution alignment and adversarial training iterations until the loss function converges.

[0063]

[0064] In addition, the present invention also discloses a respiratory tract sputum obstruction monitoring device, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to perform the respiratory tract sputum obstruction monitoring method described above.

[0065] The microsensor array uses a flexible polymer material (such as medical-grade silicone or polyurethane) as a substrate and adopts a patch design, its size and shape suitable for attachment to the corresponding part of the human chest wall. The array integrates multiple sensing units, each containing the following three types of microsensors:

[0066] Miniature ultrasonic transducer: using a center frequency of The piezoelectric ceramic probe has a diameter of no more than 5mm. During operation, the transmitting circuit generates a pulse excitation signal, which the transducer converts into ultrasound waves that are transmitted into the body. The receiving circuit receives the echo signal reflected from the sputum-gas interface in the respiratory tract, with a voltage amplitude... It is proportional to the echo intensity and reflects the acoustic impedance characteristics of sputum.

[0067] Miniature bioelectrical impedance electrode pairs: Measurements were performed using a four-electrode method, with one pair serving as the excitation electrode. The injection frequency into the body surface was:

[0068] and , amplitude One pair provides a constant alternating current; the other pair consists of measuring electrodes used to detect the corresponding voltage signal. Through the formula The electrical impedance modulus of the local tissue was calculated. and The accumulation of sputum can alter local conductivity, leading to a decrease in electrical impedance.

[0069] Miniature photoelectric sensor: Contains two light sources (LEDs) that emit different wavelengths, with wavelengths of... and And a photodetector (such as a photodiode). The light source alternately emits light, and the photodetector receives the light signal scattered by the tissue and converts it into a photocurrent signal. and Calculate the difference in optical density between the two. This value is related to blood oxygen and water content and can be used to help distinguish sputum from other tissues.

[0070] The sensor array comprises three sensing units, positioned to correspond to the following anatomical landmarks on the human body surface: the first unit is located in the suprasternal notch, the second unit is located in the second intercostal space along the right midclavicular line (corresponding to the projection of the right main bronchus), and the third unit is located in the second intercostal space along the left midclavicular line (corresponding to the projection of the left main bronchus). Each sensing unit is connected to the signal acquisition and processing main control module via flexible wires.

[0071] Specific implementation methods of signal preprocessing:

[0072] The raw signals from each sensor need to be preprocessed to convert them into feature data that can be used in subsequent algorithms. Preprocessing is performed in the microprocessor separately for each mode:

[0073] Ultrasonic signal preprocessing: processing the received raw radio frequency echo signal Envelope detection was performed to obtain the signal strength-time series. The sequence is filtered using a moving average (window length of 10 ms to suppress noise). The average value of the sequence is calculated over a complete respiratory cycle (typically 3–5 s). and standard deviation , as a characteristic value characterizing the acoustic properties of sputum.

[0074] Bioelectrical impedance signal preprocessing: Directly read the calculated electrical impedance magnitude. and Calculate the ratio of the impedance at the two frequencies. Due to the electrical properties of sputum, its accumulation can lead to... Characteristic changes occur.

[0075] Optical signal preprocessing: Reading the photocurrent value after conversion by the photodetector and Calculate the ratio of the intensity of the two received wavelengths of light. This ratio is sensitive to blood oxygen and water content and can be used to help differentiate sputum from other tissues.

[0076] The total (calculated from the above three sensing units and three modes) , , , The 12 feature values ​​(3 x 12) are concatenated into a feature vector X∈R in a time series (e.g., one group per second). 12 This results in multimodal biophysical signal data samples that are ultimately input into subsequent signal processing algorithms. All data samples undergo Z-score normalization before being input into the network.

[0077] Specific parameters and training details of the algorithm model:

[0078] In a preferred embodiment, the key component parameters and training strategy of the algorithm are as follows:

[0079] Signal processing algorithm (for constructing the optimal source domain set): The number of neurons in the three fully connected layers of the encoding and classification modules are [512, 256, 128] and [128, 64, 1], respectively. LeakyReLU is used as the activation function. During training, the Adam optimizer is used with a learning rate of 0.0001 to 0.0001 and a binary cross-entropy loss function.

[0080] Feature extraction module (three-tier dual-pathway network):

[0081] Frequency domain convolutional layer: uses one-dimensional convolution, kernel size Step length Number of output channels This is followed by the ReLU activation function.

[0082] Temporal convolutional layer: uses one-dimensional convolution, kernel size Step length Number of output channels This is followed by the ReLU activation function.

[0083] Channel convolutional layer: Uses 1x1 convolution, outputting the number of channels. .

[0084] Average pooling layer: kernel size .

[0085] Projection layer: A fully connected layer that projects the features as... A dimensional vector, used as shallow features.

[0086] Global Adaptation Unit: Number of Attention Heads Dimensions of the key / query / value matrix 16.

[0087] Adversarial training: The gradient reversal layer hyperparameter λ adopts a gradual strategy, starting from 0 and increasing linearly with the number of training iterations, eventually stabilizing at 1.0. The discriminator is a three-layer fully connected network (number of neurons [128, 64, 1]). Adversarial training uses the Adam optimizer with a learning rate of 0.0001.

[0088] Classifier training: After fixing the parameters of the discriminator and global adapter, an SGD optimizer with momentum of 0.9 and a learning rate of 0.001 is used to minimize the total loss function. Training continues until the accuracy on the validation set no longer improves, or the early stopping mechanism is triggered.

[0089] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described below in conjunction with specific embodiments.

[0090] A method for monitoring airway sputum blockage includes the following steps:

[0091] S101: Acquire respiratory multimodal biophysical signal data samples from patients in the target domain and patients from multiple source domains, respectively;

[0092] The multimodal biophysical signal data samples include ultrasound reflection characteristics, bioelectrical impedance changes, and optical scattering characteristics. The signal acquisition process involves attaching a miniature sensor array to the patient's chest wall at a location corresponding to the airway (e.g., trachea, main bronchus). This sensor array integrates three types of miniature sensors:

[0093] Miniature ultrasonic transducer: used to emit ultrasonic waves and receive the signals reflected back from the sputum-gas interface inside the respiratory tract. The intensity of the reflected signal (echo intensity) is related to the acoustic impedance characteristics of the sputum.

[0094] Miniature bioelectrical impedance electrode pairs: used to inject a weak alternating current into the body surface and measure the corresponding voltage signal, calculating the local tissue impedance value. Accumulated sputum alters the conductivity of the area, leading to a decrease in impedance.

[0095] Miniature photoelectric sensors: These emit near-infrared light of specific wavelengths and receive the light signals after they are scattered by tissue. The presence of sputum alters the scattering path and intensity of photons, and this change can be used to infer the density and distribution of sputum.

[0096] S102: Collect respiratory signal data of patients in the target domain through a miniature sensor array, and select signal data samples from multiple source domain patients using a preset signal processing algorithm to construct the optimal source domain set;

[0097] The process of constructing the optimal source domain set is as follows:

[0098] The signal processing algorithm consists of an encoding module and a classification module, both composed of three fully connected layers. For each candidate source domain patient dataset, a separate signal processing algorithm with non-shared parameters is trained. The encoding module maps the raw multimodal signal data of the source domain patients into a low-dimensional representation vector. The classification module then predicts the similarity probability between the source domain and the current target domain based on this representation vector. Finally, the signal data samples of the top K source domain patients with the highest similarity probabilities are selected to form the optimal source domain set for subsequent adaptation processes. This step ensures the quality and relevance of the source domain data in the subsequent transfer learning process.

[0099] S103: Input the signal data of the target domain patient and its corresponding optimal source domain set into the global adapter, which consists of a feature extraction module and a global adaptation unit based on a multimodal fusion mechanism.

[0100] An optional data augmentation step is also included before input:

[0101] Generate a random weight value for each signal data sample in the target domain. , Then, all generated weights are normalized. Two signal data samples are randomly selected from the target domain. and Use its corresponding normalized weights and Perform a weighted linear combination to generate a new signal data sample. .

[0102] The labels for the new samples are adopted and The label of the original sample corresponding to the larger value in the middle. This operation expands the amount of data in the target domain, which helps improve the model's generalization ability in the target domain.

[0103] Subsequently, a feature extraction module is used to extract shallow features from each signal data sample in the target domain and the optimal source domain set, respectively. The feature extraction module is a three-tiered dual-path network. This network includes two paths with identical structures but independent parameters, processing signal data in the source and target domains respectively. Each path includes the following sequentially connected layers:

[0104] Input layer: Receives standardized multimodal signal data.

[0105] Frequency domain convolutional layer: The input signal is transformed to the frequency domain by performing a Fast Fourier Transform (FFT), and frequency domain features are extracted using a one-dimensional convolution kernel to capture specific frequency components related to sputum.

[0106] Temporal convolutional layer: It uses a one-dimensional convolutional kernel to extract the dynamic change features of the signal in the time dimension, and is used to analyze the temporal pattern of sputum accumulation.

[0107] Channel convolutional layer: Uses 1x1 convolutional kernels to fuse feature channels of different modalities (ultrasound, impedance, optical) and learn the correlation between modalities.

[0108] Average pooling layer: Downsamples the features after convolution, reducing computational complexity and retaining the most significant features.

[0109] Projection layer: A fully connected layer that projects the pooled features into a fixed-dimensional (e.g., 128-dimensional) vector, which serves as the shallow feature output of the feature extraction module.

[0110] Then, the global adaptation unit guides the alignment of conditional distributions between the target domain and the optimal source domain set. The specific process is as follows:

[0111] S201:

[0112] The shallow feature vectors of the target domain signal data are respectively and shallow feature vectors of signal data in the optimal source domain set It is divided into multiple segments of equal length.

[0113] S202:

[0114] These feature fragments are processed using three different linear layers (fully connected layers) to generate a key matrix K, a query matrix Q, and a value matrix V. The key matrix K and the query matrix Q are derived from shallow features of the optimal source domain set. The value matrix V comes from shallow features of the target domain. .

[0115] S203:

[0116] Attention queries are performed in parallel using h attention heads (e.g., h=8). The calculation process for the h-th attention head is as follows:

[0117] ;

[0118] ;

[0119] in, It is the first The attention weight matrix calculated by each attention head represents the importance distribution of source domain features to target domain features; , ) are respectively the first The query, key, and value matrices corresponding to each attention head are obtained by partitioning the complete Q, K, V matrices. It is the dimension of the key matrix K. It serves a scaling function to prevent the gradient of the softmax function from vanishing due to an excessively large dot product; It is the first The output matrix of each attention head.

[0120] S204:

[0121] The output matrices of each attention head are concatenated to obtain the shallow feature O after conditional distribution alignment:

[0122] ;

[0123] The output O integrates the discriminative information of the source domain and the structural information of the target domain, achieving alignment of the conditional distribution.

[0124] S104: Use the shallow features O after conditional distribution alignment as input to the classifier and discriminator, and guide the alignment of the edge distributions between the target domain and the source domain through adversarial training.

[0125] During adversarial training, a gradient inversion layer (GRL) is used to force the feature distributions to converge. In forward propagation, GRL passes the input identically, while in backward propagation, it multiplies the incoming gradient by a negative number (-lambda) to update the gradient inversely, making the features generated by the feature extraction module unable to be distinguished by the discriminator.

[0126] The loss function used in this adversarial training is - The expression is as follows:

[0127]

[0128] in,

[0129] Indicates resistance to loss;

[0130] Represents the mathematical expectation;

[0131] This represents a signal data sample originating from the source domain S;

[0132] This represents a signal data sample originating from the target domain T;

[0133] Indicates the distribution of source domain data;

[0134] Indicates the distribution of data in the target domain;

[0135] D stands for Discriminator, which is a neural network composed of fully connected layers used to determine whether the input features come from the source domain or the target domain;

[0136] F represents the combination of the feature extraction module and the global adaptation unit;

[0137] and These are the extracted and adapted shallow features of the source and target domains, respectively.

[0138] It is a hyperparameter used to control the strength of the gradient penalty term;

[0139] It is a gradient penalty term, applied to the discriminator at random interpolation points. The gradient of the output at that point is obtained by taking the L2 norm;

[0140] It is a gradient operator;

[0141] These are random interpolation points, constructed using the following formula: , It is a random number that is uniformly distributed in the interval [0,1].

[0142] This loss function maximizes the discriminant's discriminative power by optimizing the discriminator D, while minimizing the feature extractor F. This results in the edge distribution of the source and target domains. and As similar as possible.

[0143] S105: Fix the parameters of the discriminator and global adapter, and train the classifier to identify and evaluate the state of sputum blockage based on the input signal data.

[0144] The loss function used when training the classifier The expression is as follows:

[0145] ;

[0146] ;

[0147] ;

[0148] in,

[0149] It is the total loss function of the classifier;

[0150] It is the standard cross-entropy classification loss;

[0151] and These are the number of signal data samples in the source domain and the target domain, respectively.

[0152] C represents the number of categories (e.g., normal, mild sputum blockage, severe sputum blockage).

[0153] It is a one-hot vector representing the true class label of the i-th sample. If sample i belongs to class c, then =1, otherwise 0;

[0154] It is the probability predicted by the classifier that the i-th sample belongs to class c;

[0155] It is an adversarial loss, used at this stage to further ensure the domain invariance of features and improve the generalization performance of the classifier on the target domain.

[0156] By minimizing The trained classifier is eventually able to accurately identify and assess the sputum blockage status of patients in the target domain.

[0157] Example 2:

[0158] This embodiment, based on Embodiment 1, further details the specific configuration of the micro-sensor array and its signal preprocessing process.

[0159] The micro-sensor array uses a flexible circuit board as a substrate and employs a patch design for easy attachment to the human chest wall. The array contains three sensing units, corresponding to the projection areas of the suprasternal notch, the left main bronchus, and the right main bronchus on the body surface, respectively. Each sensing unit includes the ultrasonic transducer, a pair of impedance electrodes, and a photoelectric sensor.

[0160] The signal preprocessing process is as follows:

[0161] Ultrasonic signal: Envelope detection is performed on the received raw radio frequency signal to obtain a signal intensity-time series (A-mode ultrasound). This series is then subjected to moving average filtering to suppress noise, and its average intensity over one respiratory cycle is calculated. and standard deviation As a characteristic. Average intensity Positively correlated with sputum acoustic impedance, standard deviation It reflects the unevenness of sputum.

[0162] Impedance signal: Calculate the ratio of impedance at two frequencies. Due to the electrical properties of sputum, its accumulation can lead to... Characteristic changes occur.

[0163] Optical signal: Two light sources with wavelengths of 730nm and 850nm are used. Calculate the ratio of the intensity of the two received wavelengths. This ratio is sensitive to blood oxygen and water content and can be used to help differentiate sputum from other tissues.

[0164] The eigenvalues ​​obtained from the above calculations , , and The data are spliced ​​together according to the time sequence to form the final multimodal biophysical signal data sample that is input into subsequent algorithms.

[0165] Example 3:

[0166] Existing flexible sensors are prone to interface displacement during patient activity or respiratory movements, leading to signal attenuation or motion artifacts, which affects the stability and accuracy of long-term monitoring. Therefore, this embodiment is a further optimization based on embodiment 2. In this embodiment, the micro-sensor array is based on a flexible biomimetic sensing and dynamic adaptive attachment mechanism.

[0167] The micro-sensor array uses medical-grade silicone with an elastic modulus matching that of human skin as its substrate, and its attachment interface is designed with a biomimetic micro-suction cup structure array.

[0168] The micro-suction cup structure is embedded with a variable stiffness unit made of thermally responsive hydrogel. Under the control of a microcontroller, the variable stiffness unit can switch between a low stiffness attachment mode and a high stiffness stable mode through a micro heating element.

[0169] The micro-sensor array also integrates a micro triaxial accelerometer for real-time monitoring of the relative displacement between the sensor and the body surface, and for feedback control of the switching of the variable stiffness unit.

[0170] The workflow of the dynamic adaptive attachment mechanism includes:

[0171] During the initial attachment stage, the variable stiffness unit is controlled to be in a low stiffness mode, making the micro-suction cup structure easy to adhere to the skin surface.

[0172] During the monitoring phase, when the accelerometer data indicates that the patient is in a calm breathing state, the variable stiffness unit is switched to a high stiffness mode to suppress the relative displacement of the interface caused by respiratory motion.

[0173] When the accelerometer detects large-amplitude motion (such as when a patient turns over or coughs), the variable stiffness unit briefly switches back to low stiffness mode to prevent discomfort or damage to the skin. After the motion ends, it switches back to high stiffness mode.

[0174] In practical implementation:

[0175] Substrate material: Medical-grade silicone rubber with a Shore A hardness of 10-15 is used. Its elastic modulus (~100 kPa) matches that of the human dermis (~50-200 kPa) to ensure mechanical compatibility. In this embodiment, "~" is an approximate symbol.

[0176] Bionic micro-suction cups: On the skin-contacting side of the sensor, an array of microstructures resembling octopus suction cups, with a diameter of 500 μm and a height of 200 μm and a spacing of 1 mm, is fabricated using a micro-molding process. This structure effectively increases the contact area and adhesion.

[0177] Variable stiffness unit: A thermoresponsive hydrogel (such as poly(N-isopropylacrylamide), PNIPAM) is injected into the central chamber of each microcup. This hydrogel swells at low temperatures (<32°C), exhibiting a very low modulus (~1 kPa) and remaining soft; at high temperatures (>35°C), it dehydrates and shrinks, its modulus increasing sharply (~1 MPa), becoming rigid. A micro-film resistance heater (50Ω resistance, 1.5V operating voltage) is integrated at the bottom of each chamber.

[0178] Accelerometer: An ultra-small triaxial MEMS accelerometer (such as ADXL363) is used, which is mounted at the center of the sensor array substrate to monitor vibration and displacement.

[0179] Workflow and parameters:

[0180] Initial application: Before application, 0V voltage is applied to all heaters via an external control circuit, bringing the PNIPAM hydrogel to a low-stiffness state at room temperature (25°C). At this point, the sensor array is applied to the target location on the patient's chest wall, and the micro-suction cup structure can easily form a tight, non-invasive fit with the skin.

[0181] Stability Monitoring: After attachment, the monitoring system is activated. The microprocessor continuously reads accelerometer data. When the patient is detected to be in a stable breathing state (accelerometer amplitude change <0.1g), a 1.5V voltage is applied to all heaters for 10 seconds, raising the hydrogel temperature to approximately 37°C and transforming it into a high-stiffness state. This "locks" the sensor array in its current position, greatly suppressing the minute displacements caused by respiratory movements.

[0182] Dynamic adaptation: If the accelerometer detects an acceleration change exceeding 0.5g (indicating the patient may be beginning significant movement), the microprocessor immediately cuts off the heater voltage, allowing the hydrogel to cool and return to a soft state within 30 seconds. This allows for some slippage between the sensor and the skin, preventing skin traction. Once the acceleration data returns to a stable state (change <0.1g for 1 minute), the heater is reactivated, and the system re-enters the high-stiffness stable mode.

[0183] To verify the effectiveness, a comparative experiment was conducted. The sensor array was fixed using conventional medical adhesive (Comparative Example 2), and the adaptive array of this embodiment was tested on an experimental platform simulating respiratory motion.

[0184] Signal stability: During 4 hours of continuous monitoring, the coefficient of variation (CV) of the ultrasonic signal amplitude acquired by the traditional method was 18.7%, while the method in this embodiment significantly reduced it to 5.2%.

[0185] Motion artifacts: When a perturbation simulating a cough is introduced, the bioelectrical impedance signal acquired by the traditional method exhibits a huge spike artifact, while the signal perturbation amplitude is reduced due to the transient unlocking in this embodiment.

[0186] Patient comfort: Volunteer feedback indicates that the array in this embodiment significantly improves long-term wearing comfort.

[0187] This embodiment actively maintains the stability of the signal acquisition interface by drawing on bionics and soft robotics technologies from across disciplines, improving quality from the data source and providing an effective solution to address the core pain points in long-term, dynamic monitoring.

[0188] Comparative Example 1:

[0189] To highlight the inventiveness of this invention, this embodiment provides a comparative example. This comparative example represents a typical prior art solution: inferring sputum blockage indirectly based on ventilator waveform parameters.

[0190] The method includes the following steps:

[0191] The airway pressure and flow waveforms are continuously acquired from the ventilator data interface.

[0192] Characteristic parameters are extracted from the waveform, including: peak airway pressure (Ppeak), plateau pressure (Pplat), airway resistance (Raw), and dynamic compliance (Cdyn).

[0193] Threshold setting: If Raw remains above the threshold (e.g., 15 cmH2O / L / s) and Cdyn remains below the threshold (e.g., 40 mL / cmH2O) for more than 5 minutes, a sputum blockage alarm will be triggered.

[0194] Comparative analysis of effects:

[0195] In the simulation experiment, the same group of patient data were monitored using both the scheme of Example 1 of this invention and the comparative scheme of this invention.

[0196] Specificity: When a patient's raw blood volume increases due to bronchospasm, the comparative protocol incorrectly triggers a sputum blockage alarm (false positive). However, the protocol of this invention, due to its direct sensing of the physical properties of sputum, is unaffected by bronchospasm and does not issue an alarm, demonstrating higher specificity.

[0197] Early warning: In experiments simulating slow sputum accumulation, the present invention's approach provides an early warning approximately 20 minutes before significant changes in Raw and Cdyn occur, by measuring a continuous decrease in electrical impedance and an increase in ultrasound scattering on the body surface. In contrast, the comparative approach only issues an alarm after sputum accumulation has reached a certain level, causing significant changes in airway mechanics parameters, demonstrating a clear lag.

[0198] Location capability: By analyzing the signal differences between different sensor units, this invention successfully determined that sputum accumulation occurred in the right middle lobe of the lung. In contrast, the comparative scheme provided global airway parameters that failed to offer any location information.

[0199] In summary, the solutions provided in Embodiments 1-2 of this invention, based on direct sensing of the biophysical characteristics of sputum, effectively solve the fundamental defects of the existing technologies represented by the comparative figures, such as high false alarm rate, delayed early warning, and inability to locate, by combining multimodal sensing with domain-adaptive machine learning algorithms.

[0200] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0201] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring respiratory tract sputum blockage, characterized in that, Includes the following steps: S101: Acquire respiratory multimodal biophysical signal data samples from patients in the target domain and multiple source domains, respectively. The multimodal biophysical signal data samples include ultrasound reflection characteristics, bioelectrical impedance changes, and optical scattering characteristics. S102: Collect respiratory signal data of patients in the target domain through a miniature sensor array, and select signal data samples from multiple source domain patients using a preset signal processing algorithm to construct an optimal source domain set. The signal processing algorithm consists of an encoding module and a classification module. Both the encoding module and the classification module consist of three fully connected layers, and each source domain uses a separately trained signal processing algorithm with non-shared parameters. S103: Input the signal data of the patient in the target domain and its corresponding optimal source domain set into the global adapter. The global adapter includes a feature extraction module and a global adaptation unit based on a multimodal fusion mechanism. The feature extraction module extracts shallow features of each signal data sample in the target domain and the optimal source domain set respectively, and the global adaptation unit guides the alignment of the conditional distribution between the target domain and the optimal source domain set. S104: The shallow features after conditional distribution alignment are used as input to the classifier and discriminator, and the edge distribution between the target domain and the source domain is aligned through adversarial training. The adversarial training uses a gradient inversion layer to force the feature distribution to converge. S105: Fix the parameters of the discriminator and the global adapter, and train the classifier to identify and evaluate the sputum blockage state based on the input signal data.

2. The method for monitoring respiratory tract sputum blockage according to claim 1, characterized in that: The feature extraction module described in S103 is a three-tiered dual-path network. The three-tiered dual-path network includes two paths corresponding to the source domain and the target domain, respectively. The input signal data samples of the two paths are of the same size. Each path includes an input layer, a frequency domain convolutional layer, a time domain convolutional layer, a channel convolutional layer, an average pooling layer, and a projection layer connected in sequence. The projection layer is used to project the extracted features into a single-channel embedding sequence with the same dimension, which serves as the shallow feature finally obtained by the feature extraction module.

3. The method for monitoring respiratory tract sputum blockage according to claim 1, characterized in that: S103 utilizes global adaptors to guide the conditional distribution alignment between the target domain and the optimal source domain set, including: The shallow features of the target domain signal data and the shallow features of the signal data in the optimal source domain set are divided into multiple segments of equal length. A linear layer is used to generate a key matrix, a query matrix, and a value matrix, where the key matrix and the query matrix are derived from shallow features of signal data in the optimal source domain set, and the value matrix is ​​derived from shallow features of signal data in the target domain. Attention queries are performed using multiple attention heads. Each attention head uses a query matrix, a key matrix, and a value matrix to calculate attention weights and generate query results. The query results from each attention head are concatenated to obtain shallow features after conditional distribution alignment.

4. The method for monitoring respiratory tract sputum blockage according to claim 1, characterized in that: In S104, when guiding the alignment of the edge distribution between the target domain and the source domain through adversarial training, a gradient inversion layer is used to calculate the gradient penalty. The loss function of adversarial training includes the discriminative loss of the source domain and the target domain, as well as the gradient penalty term. The samples in the random interpolation domain are generated by combining the samples of the source domain and the target domain with random weights.

5. The method for monitoring respiratory tract sputum blockage according to claim 1, characterized in that: The loss function used when training the classifier in S105 includes classification loss and adversarial loss. The classification loss is calculated based on the number of signal data samples and the categories in the source and target domains to determine the prediction error, while the adversarial loss is calculated based on the discriminator's output of shallow features in the source and target domains.

6. The method for monitoring respiratory tract sputum blockage according to claim 1, characterized in that: The process of constructing the optimal source domain set as described in S102 includes: The source domain signal data is encoded using an encoding module, and then the similarity between each source domain and the target domain is evaluated using a classification module. The source domain signal data samples with the highest similarity are selected to form the optimal source domain set.

7. The method for monitoring respiratory tract sputum blockage according to claim 1, characterized in that: The adversarial training process in S104 includes: By backpropagating gradients through a gradient reversal layer, the discriminator is unable to distinguish shallow features between the source and target domains, thereby achieving edge distribution alignment and adversarial training iterations until the loss function converges.

8. A respiratory tract sputum blockage monitoring device, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the airway sputum obstruction monitoring method according to any one of claims 1 to 7.

Images