Remote respiratory health monitoring and management system and method

Abstract

The present application relates to the technical field of respiratory health monitoring, and discloses a remote respiratory health monitoring and management system and method. The method comprises collecting original respiratory data sets containing time sequence and frequency domain features through a multi-modal sensor, dividing respiratory cycle segments and extracting key waveform marker points through dynamic segmentation processing, constructing a respiratory state transition network to form a dynamic topology structure. A respiratory environment coupling matrix is generated in combination with environmental parameters, abnormal respiratory segments disturbed are screened, an analysis model is trained to output a deviation score, multi-level early warning is triggered, and a report containing a backtracking path is generated. Through cross-device verification and matching of user activity log timestamps, data is integrated to construct a respiratory behavior correlation graph, the model determination threshold is updated based on the graph, and the monitoring closed loop feedback is completed. The method realizes accurate monitoring and dynamic management of respiratory health, and provides a complete solution for remote respiratory health control.

Description

Technical Field

[0001] This invention relates to the field of respiratory health monitoring technology, specifically to a remote respiratory health monitoring and management system and method. Background Technology

[0002] Respiratory physiological signals, as important indicators reflecting human health status, play an irreplaceable role in disease prevention, chronic disease management, and telemedicine. With the increasing aging population and rising incidence of chronic respiratory diseases, traditional monitoring models relying on medical institution equipment can no longer meet people's needs for home-based, routine health management. Remote respiratory health monitoring technology has gradually become a hot research topic in the industry.

[0003] Existing respiratory monitoring methods mostly use a single sensor to collect data, which can only acquire local features of the respiratory signal and cannot fully reflect the complex changes in respiratory physiology. In the data processing stage, traditional methods often use static segmentation to divide the respiratory cycle, which cannot adapt to the dynamic fluctuations of respiratory rhythm, resulting in insufficient accuracy in extracting key waveform markers and affecting the accuracy of subsequent respiratory pattern analysis. At the same time, most monitoring schemes ignore the interference of environmental factors on respiratory signals. Changes in environmental parameters such as temperature, humidity, and air pressure are easily misinterpreted as respiratory abnormalities, reducing the reliability of the detection results.

[0004] In terms of respiratory pattern analysis, current technologies lack effective capture of the evolutionary patterns of the respiratory cycle, making it difficult to establish dynamic correlations between respiratory states and timely identification of abnormal trends in early respiratory patterns. Early warning mechanisms are relatively simple, often based on fixed thresholds to trigger alarms, failing to consider individual respiratory differences and the influence of environmental and behavioral factors, leading to false alarms or missed alarms. Furthermore, current monitoring systems generally lack in-depth verification of abnormal respiratory segments and fail to correlate respiratory states with users' daily behaviors, making it difficult to identify the triggering factors for abnormal breathing, thus limiting the guiding value of monitoring results for health management.

[0005] Most existing methods lack a complete closed-loop feedback mechanism, and the judgment criteria of the monitoring model remain fixed. They cannot be dynamically optimized based on individual user characteristics and long-term monitoring data. After long-term use, the monitoring accuracy is prone to decline, making it difficult to adapt to the personalized needs of different users or keep up with the dynamic changes in users' health status. This limits the practical application effect of remote respiratory health monitoring technology. Summary of the Invention

[0006] The purpose of this invention is to provide a remote respiratory health monitoring and management system and method to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for remote respiratory health monitoring and management, the method comprising:

[0008] By acquiring users' respiratory physiological signals through multimodal sensors, a raw respiratory dataset containing temporal and frequency domain features is generated.

[0009] The original respiratory dataset is dynamically segmented to divide it into continuous respiratory cycle segments, and key waveform markers are extracted from each respiratory cycle segment.

[0010] A respiratory state transition network is constructed based on key waveform markers, and waveform markers of adjacent respiratory cycle segments are associated to form a dynamic topology that reflects the evolution of respiratory patterns.

[0011] The respiratory rhythm stability index is calculated based on dynamic topology, and a respiratory environment coupling matrix is ​​generated by combining temperature, humidity and air pressure data in the environmental parameter dataset.

[0012] Abnormal breathing segments affected by environmental interference are screened out by the breathing environment coupling matrix, and the spatiotemporal distribution characteristics of the abnormal breathing segments are marked to train the breathing abnormality analysis model and output the breathing pattern deviation score.

[0013] A multi-level early warning mechanism is triggered based on the breathing pattern deviation score, generating an early warning report that includes the backtracking path of abnormal breathing segments;

[0014] Perform cross-device data verification on abnormal breathing segments in the early warning report and match them with the timestamps of behavioral events in the user activity log;

[0015] By integrating and verifying abnormal breathing segments with behavioral event timestamps, a respiratory behavior correlation map is constructed.

[0016] The threshold for judging respiratory abnormalities is updated based on the respiratory behavior correlation map to complete the closed-loop feedback of respiratory health monitoring.

[0017] Preferably, the acquisition of user respiratory physiological signals via a multimodal sensor specifically includes:

[0018] Simultaneously acquire displacement signals output by the chest and abdominal motion sensor, respiratory sound signals collected by the microphone array, and nasal airflow temperature change signals generated by the infrared thermal imager;

[0019] Baseline drift correction is performed on displacement signals, environmental noise suppression is performed on breath sound signals, and dynamic threshold segmentation is performed on temperature change signals.

[0020] The three types of signals were aligned according to sampling time and merged into a raw respiratory dataset with spatiotemporal labels.

[0021] Preferably, the dynamic segmentation of the original respiratory dataset specifically includes:

[0022] Analyze the inflection points between the inspiratory and expiratory phases in the original respiratory data set, using the interval between adjacent inflection points as the basic segmentation unit;

[0023] Calculate the variance contribution rate of the signal within each segment unit, and merge consecutive segment units with variance contribution rates below a threshold;

[0024] Insert respiratory cycle segment separators at the boundaries of the merged segment units.

[0025] Preferably, the construction of the respiratory state transition network based on key waveform marker points specifically includes:

[0026] Identify the amplitude difference between peaks and troughs in a respiratory cycle segment as waveform marker points;

[0027] Establish the transition probability between the waveform marker point of the current respiratory cycle segment and the waveform marker point of the next respiratory cycle segment;

[0028] When the transition probability exceeds the preset connectivity threshold, directed edges are drawn between the waveform marker points to form a dynamic topology.

[0029] Preferably, the step of generating a respiratory environment coupling matrix by combining temperature, humidity, and air pressure data from the environmental parameter dataset specifically includes:

[0030] Extract the coefficient of variation of inspiratory duration from the respiratory rhythm stability index;

[0031] The coefficient of variation is convolved with the temperature and humidity gradients in the environmental parameter dataset.

[0032] Matrix elements in the output convolution result whose absolute value is greater than the critical value are used as identifiers of disturbed abnormal respiratory segments.

[0033] Preferably, the training of the respiratory anomaly analysis model based on the spatiotemporal distribution characteristics of the labeled abnormal respiratory segments specifically includes:

[0034] The spatiotemporal distribution characteristics of abnormal breathing segments are converted into three-dimensional tensor inputs;

[0035] Local anomaly patterns of tensors are extracted using spatiotemporal convolution kernels;

[0036] The propagation characteristics of abnormal patterns over time are captured by a gated loop unit.

[0037] Preferably, the multi-level early warning mechanism triggered based on the breathing pattern deviation score specifically includes:

[0038] When the breathing pattern deviation score is in the first interval, an abnormal breathing rhythm prompt is generated;

[0039] When the respiratory pattern deviation score is in the second interval, a warning of abnormal respiratory muscle function is generated.

[0040] When the respiratory pattern deviation score is in the third interval, an alarm for abnormal respiratory center regulation is generated.

[0041] Preferably, the cross-device data verification of abnormal respiratory segments in the early warning report specifically includes:

[0042] Compare the timestamps of abnormal breathing segments with the time of user position changes recorded by the smart bracelet;

[0043] When the difference between the time of body position change and the time of onset of abnormal breathing segment is less than the set tolerance, it is marked as a position-related respiratory abnormality.

[0044] When the time of body position change does not overlap with the abnormal breathing segment, it is marked as pathological respiratory abnormality.

[0045] Preferably, the construction of the respiratory behavior association graph specifically includes: taking abnormal respiratory segments as nodes and the chronological order of behavioral event timestamps as edges, calculating the coupling weight between the similarity of respiratory parameters and the behavioral time interval between nodes, and pruning edges in the association graph whose weights are lower than the retention threshold according to the coupling weights;

[0046] The specific thresholds for updating the respiratory abnormality analysis model based on respiratory behavior correlation maps include:

[0047] The cluster density of pathological respiratory abnormalities in the respiratory behavior association map is statistically analyzed. When the cluster density exceeds the historical benchmark value, the judgment threshold of the respiratory pattern deviation score is lowered by a preset ratio. When the cluster density is lower than the historical benchmark value, the judgment threshold is reset to the initial default value.

[0048] Preferably, the present invention also includes a remote respiratory health monitoring and management system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the remote respiratory health monitoring and management method described above.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] By acquiring respiratory physiological signals through multimodal sensors, and simultaneously obtaining temporal and frequency domain features, this approach can more comprehensively capture the complex information of respiratory signals compared to single-sensor acquisition. This provides rich data support for subsequent respiratory state analysis and avoids judgment biases caused by incomplete data. The dynamic segmentation processing method can accurately adapt to the natural fluctuations of respiratory rhythm, effectively dividing continuous respiratory cycle segments and ensuring the accuracy of key waveform marker extraction, making respiratory cycle analysis more closely reflect the actual state of human respiration.

[0051] The construction of a respiratory state transition network links discrete respiratory cycle segments through waveform markers. The resulting dynamic topology clearly reflects the evolution of respiratory patterns, facilitating the timely capture of subtle changes in respiratory states and enabling dynamic tracking and in-depth analysis of respiratory patterns. The respiratory environment coupling matrix generated by combining environmental parameters effectively distinguishes between environmental interference and genuine respiratory anomalies, reducing misjudgments caused by environmental factors such as temperature, humidity, and air pressure. This improves the accuracy of abnormal respiratory segment screening and makes the monitoring results more valuable.

[0052] The multi-level early warning mechanism is designed to trigger different levels of warnings based on breathing pattern deviation scores. The warning reports also include a backtracking path of abnormal breathing segments, allowing users to quickly pinpoint the timeline and evolution of the abnormality, providing clear direction for subsequent health assessments and interventions. The cross-device data verification process, by matching timestamps from user activity logs, verifies the authenticity of abnormal breathing segments and establishes a correlation between respiratory status and daily behavior. This helps uncover the triggering factors for abnormal breathing, making respiratory health management more targeted.

[0053] The construction of the respiratory behavior correlation map integrates abnormal respiratory data and user behavior information, intuitively presenting the intrinsic relationship between the two and providing a basis for personalized respiratory health guidance. Based on the dynamic updating of the analysis model's judgment thresholds using this map, a complete closed-loop feedback mechanism is formed, enabling the model to continuously adapt to changes in users' individual characteristics and health status, constantly optimizing monitoring accuracy, and ensuring the stability and reliability of monitoring results during long-term use. Attached Figure Description

[0054] Figure 1 This is a time-series characteristic diagram of respiratory physiological signals and environmental parameters;

[0055] Figure 2 A flowchart for acquiring user respiratory physiological signals using a multimodal sensor;

[0056] Figure 3 A flowchart for dynamically segmenting the original respiratory dataset;

[0057] Figure 4 This is a dynamic segmented diagram of the respiratory cycle. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1 This invention provides a method for remote respiratory health monitoring and management, comprising: acquiring user respiratory physiological signals using multimodal sensors to generate a raw respiratory dataset containing temporal and frequency domain features; dynamically segmenting the raw respiratory dataset to divide it into continuous respiratory cycle segments and extracting key waveform markers from each segment; using these key waveform markers to construct a respiratory state transition network, associating waveform markers from adjacent respiratory cycle segments to form a dynamic topology reflecting the evolution of respiratory patterns; combining the dynamic topology with temperature, humidity, and air pressure data from the environmental parameter dataset to calculate a respiratory rhythm stability index and generate a respiratory environment coupling matrix; using the respiratory environment coupling matrix to filter out abnormal respiratory segments affected by environmental disturbances, marking the spatiotemporal distribution characteristics of these abnormal respiratory segments to train a respiratory anomaly analysis model, and outputting a respiratory pattern deviation score; triggering a multi-level early warning mechanism based on the respiratory pattern deviation score, generating an early warning report containing the backtracking path of the abnormal respiratory segments; performing cross-device data verification on the abnormal respiratory segments in the early warning report, matching them with behavioral event timestamps in the user activity log; and integrating the verified abnormal respiratory segments with the behavioral event timestamps to construct a respiratory behavior association map. The respiratory behavior association map updates the judgment threshold of the respiratory abnormality analysis model, completing the closed-loop feedback of respiratory health monitoring.

[0060] Example 1: See Figure 2 The multimodal sensor system comprises three independent acquisition units: a chest and abdominal motion sensor, a microphone array, and an infrared thermal imager. The chest and abdominal motion sensor, made of piezoelectric thin film material, is attached to the user's chest and abdomen. It converts the rise and fall of the chest and abdomen during respiration into electrical signals, which are continuously varying voltage values. The amplitude of these voltage fluctuations is proportional to the depth of respiration. The microphone array consists of four high-sensitivity microelectromechanical system (MEMS) microphone units, arranged in a diamond pattern in the suprasternal notch area in front of the user's neck. The microphone array acquires broadband sound signals generated as airflow passes through the pharynx and trachea. These sound signals include low-frequency components specific to the inspiratory phase and high-frequency components specific to the expiratory phase. The infrared thermal imager uses an uncooled microcalorimeter detector. Aimed at the outer edge of the user's nostrils, the imager captures the airflow temperature field distribution at the nasal cavity exit at a rate of 30 frames per second. The temperature field distribution data is then processed at the pixel level to generate a nasal airflow temperature change signal.

[0061] Baseline drift correction of the displacement signal is achieved using a digital filter. The digital filter is designed as a high-pass Butterworth filter with a cutoff frequency of 0.5 Hz and an eighth-order filter. After processing by the filter, the low-frequency baseline drift caused by slow body movement is eliminated, retaining the effective motion signal consistent with the breathing frequency. Environmental noise suppression of the breathing sound signal adopts an adaptive beamforming algorithm. The adaptive beamforming algorithm calculates the signal phase difference between each unit of the microphone array to form a directional receiving beam pointing towards the user's throat. The main lobe width of the beam is set to a 45-degree angle, and the sidelobe suppression ratio is greater than 20 decibels. The diffuse field noise in the environment is suppressed by spatial filtering during beamforming. Dynamic threshold segmentation of the nasal airflow temperature change signal adopts a sliding window statistical method. The sliding window duration is set to three seconds. The mean and standard deviation of the temperature data within the window are calculated. The dynamic threshold is set to the mean plus or minus twice the standard deviation. The interval of the temperature signal exceeding the dynamic threshold is determined as the effective breathing airflow interval.

[0062] The processed three types of signals are aligned according to sampling time using a hardware timestamp synchronization mechanism. The chest and abdominal motion sensor, microphone array, and infrared thermal imager are each connected to the GPS timing module, which provides a time reference with microsecond-level accuracy. The three types of signals are frame-aligned in the data acquisition card's buffer memory according to a unified time base. The aligned signals are merged into a raw respiratory dataset with spatiotemporal labels. The spatiotemporal labels include the absolute timestamp of signal acquisition, sensor geographic coordinates, signal type encoding, and signal quality identifier. The raw respiratory dataset is stored in a hierarchical data format. The top layer is the metadata layer, recording the acquisition device serial number and user identifier. The middle layer is the signal data layer, storing the quantized sample values ​​of the three types of signals. The bottom layer is the checksum layer, storing cyclic redundancy check codes for data integrity verification.

[0063] The signal acquisition frequency of the chest and abdominal motion sensor is set to 100 Hz. The analog voltage output of the chest and abdominal motion sensor is digitized by a 16-bit analog-to-digital converter, and the quantized displacement signal has a range of ±5 volts. The sampling frequency of the microphone array is set to 16 kHz. The sound pressure signal of the microphone array is digitized by a 24-bit analog-to-digital converter, and the dynamic range of the digitized breath sound signal reaches 96 dB. The temperature resolution of the infrared thermal imager reaches 0.1 degrees Celsius, and the temperature change signal of the infrared thermal imager is represented by a 14-bit digital value within the measurement range of 0 to 50 degrees Celsius. Data from the multimodal sensors is transmitted via a wireless personal area network (PAN) using the Bluetooth 5.0 protocol. Forward error correction coding is added to the transmitted data packets, and the data transmission interval is fixed at 100 milliseconds.

[0064] The generation of the original respiratory dataset includes a data preprocessing stage. This stage performs amplitude normalization on three types of signals, scaling the maximum amplitude of each signal to a uniform nominal value. The amplitude normalization for displacement signals is based on the maximum respiratory amplitude, for respiratory sound signals on the maximum sound pressure level, and for temperature change signals on the maximum temperature difference. The normalized signals then undergo time interpolation using a cubic spline interpolation algorithm to resample the non-uniformly sampled signals into 100Hz uniformly sampled signals. The original respiratory dataset is stored using a circular buffer with a capacity set to store one hour of continuous data, with new data overwriting the oldest data to achieve continuous updates.

[0065] The operating state of the multimodal sensor is controlled by the power management unit, which dynamically adjusts the sensor power consumption based on the user's activity level. When the user is stationary, the sensor enters a low-power mode, where the sampling frequency is reduced to 50 Hz. When the user is active, the sensor switches to a full-power mode, where all sensors operate at their rated sampling frequency. Sensor data acquisition is triggered by both timed and event-driven methods. Timed triggers initiate acquisition at preset time intervals, while event-driven acquisition is triggered by changes in user posture detected by the accelerometer. The generation of the raw respiratory dataset includes a data integrity check, which verifies the continuity of timestamps for the three types of signals and the integrity of the data packet sequence. A data retransmission mechanism is activated when data loss is detected.

[0066] The calibration of the chest and abdominal motion sensor utilizes a standard displacement generator, which produces periodic displacements of known amplitude. The output signal of the chest and abdominal motion sensor is linearly fitted to the standard displacement amplitude. Microphone array calibration is performed in an anechoic chamber environment. The calibration sound source emits a pure-tone signal at a standard sound pressure level, and the frequency response characteristics of each unit in the microphone array are calibrated and compensated. The infrared thermal imager calibration uses a blackbody radiation source as a reference. The blackbody radiation source provides a surface target with a known temperature, and the output of the infrared thermal imager is mapped to the blackbody temperature. Calibration parameters are stored in the sensor's non-volatile memory and are used to correct systematic errors in the acquired signals.

[0067] The multimodal sensor's mechanical structure comprises a flexible substrate and a protective shell. The flexible substrate is made of polyimide, and the sensor elements are mounted on it to conform to the curves of the human body. The protective shell is made of medical-grade silicone, providing moisture-proof and dust-proof sealing protection. The sensor array is secured to the user's skin via a medical adhesive layer made of hypoallergenic acrylic adhesive, ensuring comfort during extended wear. The signal leads utilize a shielded twisted-pair design, reducing electromagnetic interference on weak signals. The entire multimodal sensor system weighs less than 50 grams, with even weight distribution to avoid interfering with normal user activities.

[0068] The software for generating the raw respiratory dataset runs on an embedded processor with a 32-bit microcontroller architecture, which integrates a digital signal processing coprocessor. The data acquisition program, written in C, implements multi-channel synchronous sampling, real-time signal processing, and wireless data transmission. The storage format of the raw respiratory dataset defines a custom binary protocol, which consists of a file header, data blocks, and a checksum. The file header records the acquisition start time, sampling frequency, and number of channels. The data blocks store interleaved sample values ​​of three types of signals, and the checksum is calculated using a cyclic redundancy check algorithm. The raw respiratory dataset read interface provides an application programming interface (API) that supports querying by time range and streaming data access.

[0069] Example 2: See Figure 3 The original respiratory dataset comes from synchronous signals acquired by a multimodal sensor. Dynamic segmentation processing begins by analyzing the inflection points between the inspiratory and expiratory phases in the original respiratory dataset. Inflection point identification employs multi-scale derivative analysis, which calculates the first and second derivatives of the signal. The inflection point in the inspiratory phase corresponds to the zero-crossing point where the first derivative changes from negative to positive, and the inflection point in the expiratory phase corresponds to the zero-crossing point where the first derivative changes from positive to negative. The interval between adjacent inflection points serves as the basic segmentation unit, with each basic segmentation unit representing half a respiratory cycle. The duration of each basic segmentation unit is dynamically adjusted according to the respiratory rate. The variance contribution rate within each segmentation unit is calculated using sliding window variance analysis, with the sliding window width matching the length of the basic segmentation unit. The variance contribution rate is calculated as the ratio of the signal variance within the window to the overall variance. Consecutive segmentation units with variance contribution rates below a threshold of 0.1 are merged using a region growing algorithm. The merged segmentation units form a complete respiratory cycle segment. Respiratory cycle segment separators are inserted at the boundaries of the merged segment units. These separators use both timestamps and segment numbers for identification. The timestamps record the start and end times of the segment, while the segment numbers provide unique identification.

[0070] Key waveform marker extraction in respiratory cycle segments focuses on the amplitude difference between peaks and troughs. Peak detection is achieved by finding local maxima, and trough detection by finding local minima. The amplitude difference is calculated as the absolute value of the peak amplitude minus the trough amplitude. Waveform markers also include rising and falling slopes. The rising slope is calculated as the difference between the peak and trough amplitudes divided by the rise time, and the falling slope is calculated as the difference between the peak and trough amplitudes divided by the fall time. Waveform marker data is stored as feature vectors, containing four dimensions: amplitude difference, rising slope, falling slope, and duration. A respiratory state transition network is constructed using waveform marker feature vectors as nodes. Node attributes include the feature vector value and the timestamp of the corresponding respiratory cycle segment. The transition probability between waveform markers in the current respiratory cycle segment and those in the next respiratory cycle segment is established. The transition probability is calculated using a conditional probability model, which statistically analyzes the frequency of transitions from a specific waveform marker to another waveform marker in historical data. When the transition probability exceeds a preset connectivity threshold, directed edges are drawn between the waveform marker points to form a dynamic topology. The preset connectivity threshold is set to 0.6, and the weight of the directed edges is the actual transition probability value. The dynamic topology is stored in memory as an adjacency matrix, where the row and column indices of the adjacency matrix correspond to the waveform marker point numbers, and the matrix elements store the transition probability values.

[0071] The inflection point detection step in dynamic segmentation includes signal preprocessing, which employs a wavelet denoising algorithm. This algorithm uses the sym wavelet basis function for five-level decomposition, removing high-frequency noise components and retaining the core respiratory signal. Accurate inflection point location determination combines amplitude thresholds and duration constraints. The amplitude threshold requires a signal change rate greater than 0.1 volts per second at the inflection point, and the duration requires an interval greater than 0.5 seconds between inflection points. Boundary correction of basic segmentation units uses a dynamic time warping algorithm, aligning the actual signal with a standard respiratory template to eliminate segmentation errors caused by respiratory variations. Variance contribution rate calculation incorporates a sliding window overlap mechanism, with 50% overlap to ensure continuity in variance calculation at segment boundaries. Merging low-variance segmentation units uses a hierarchical clustering algorithm, employing Euclidean distance as a similarity measure, merging segments with a distance less than a threshold. The insertion of respiratory cycle segment separators is performed synchronously with the data stream. These separators are embedded in the data stream as special character sequences for easy subsequent parsing and recognition.

[0072] Key waveform marker extraction includes signal normalization, which scales the amplitude to the zero-to-one range to eliminate individual differences and sensor gain effects. Peak and trough detection employs an adaptive thresholding method, dynamically adjusting the threshold based on the amplitude distribution of the previous five respiratory cycles to avoid missed detections due to changes in respiratory depth. The feature vector dimension of the waveform markers is expanded to six dimensions, adding inspiratory and expiratory areas. The inspiratory area is calculated by numerical integration of the area enclosed by the signal curve and the baseline. Feature vector normalization uses a max-min normalization method to map each feature value to a uniform numerical range. Node generation in the respiratory state transition network includes redundant node merging, aggregating nodes with feature vector distances less than a threshold into supernodes to reduce network complexity. Transition probability calculation employs a smoothing estimation technique, introducing a Laplace smoothing factor to avoid zero-probability issues. Directed edge generation includes directionality verification, based on respiratory physiological time-series characteristics, ensuring that edge directions align with the time flow. The visualization of dynamic topology uses the force-directed layout algorithm, which simulates nodes as charges and edges as springs to generate an intuitive network diagram.

[0073] The update mechanism of the respiratory state transition network is synchronized with the real-time data stream. Newly arrived respiratory cycle segments trigger network topology updates, including node addition, edge weight adjustment, and isolated node deletion. Persistent storage of the network topology utilizes a graph database, which supports efficient traversal and query operations. Dynamic topology analysis includes network metric calculations, such as average path length, clustering coefficient, and degree distribution, which quantify the complexity of respiratory patterns. Anomaly detection in the respiratory state transition network is based on topology mutation identification, manifested as abrupt changes in node degree or splitting of connection components; mutation events trigger anomaly alarms. The computational optimization of the network construction process employs an incremental update algorithm, which only recalculates the changed topology, reducing computational overhead. The integration of dynamic segmentation processing and state transition network construction is achieved through a message queue, which buffers intermediate data, ensuring decoupling and fault tolerance between processing modules. The entire implementation adopts a multi-threaded architecture: one thread handles dynamic segmentation processing, another handles network construction, and threads exchange data via shared memory.

[0074] See Figure 4This figure visually illustrates the core process of "dynamic segmentation of the original respiratory dataset." The horizontal axis represents time, and the vertical axis represents chest and abdominal displacement. The curves reflect the respiratory displacement signals collected by the chest and abdominal motion sensors in the multimodal sensor. The circles in the figure mark the inflection points between the inspiratory and expiratory phases detected by multi-scale derivative analysis. The interval between adjacent inflection points constitutes the "basic segmentation unit." The gray bars represent respiratory cycle segments formed after merging "continuous segmentation units with a variance contribution rate below a threshold." This dynamic segmentation method breaks through the limitations of traditional static segmentation and can accurately adapt to the natural fluctuations of respiratory rhythm: first, the basic segmentation boundaries are clarified through inflection point detection, and then segments with weak physiological significance are merged through variance analysis. Finally, continuous respiratory cycle segments are divided, laying the foundation for subsequent extraction of key waveform markers and construction of a respiratory state transition network. This demonstrates the technical advantage of this invention in "fitting the actual state of human breathing" in the respiratory cycle analysis stage, ensuring the accuracy of subsequent respiratory pattern evolution analysis.

[0075] Example 3: The process of calculating respiratory rhythm stability indices based on dynamic topology begins with topology network feature extraction. The node degree centrality of the dynamic topology is used to characterize the stability of the respiratory rhythm. Node degree centrality calculates the number of connections between each waveform marker node and other nodes. The coefficient of variation of the number of connections reflects the degree of fluctuation in the respiratory rhythm. The coefficient of variation of inspiratory duration is derived from the temporal information of respiratory cycle segments. The coefficient of variation of inspiratory duration is equal to the standard deviation of inspiratory duration divided by the mean of inspiratory duration and then multiplied by 100%. The temperature, humidity, and air pressure data in the environmental parameter dataset are collected through the environmental monitoring module. The temperature and humidity data include Celsius temperature values ​​and relative humidity percentage values, and the air pressure data includes atmospheric pressure values ​​in kilopascals. The temperature and humidity gradient is calculated using the time difference method. The temperature and humidity gradient is equal to the difference between the current temperature and humidity values ​​and the previous temperature and humidity values.

[0076] When coupling respiratory rhythm stability indices with environmental parameters, a convolution operation mode is used. The mathematical expression for the convolution operation is described as follows:

[0077]

[0078] The elements of the convolution result matrix This indicates the degree of correlation between the i-th respiratory cycle segment and the j-th environmental parameter. Indicates time The specific index for respiratory rhythm stability is the coefficient of variation of the duration of inspiration. This represents the j-th parameter in the environmental parameter dataset at time j. The gradient value is given by the summation symbol, which represents the accumulation of the product over N time points. The value is set to the width of the convolution window, which is ten sampling points. Elements of the convolution result matrix. The absolute value is compared with a critical value, which is set to 0.35. Matrix elements with an absolute value greater than 0.35 are marked as significant interference regions. The row dimension of the respiratory environment coupling matrix corresponds to the respiratory cycle segment number, the column dimension corresponds to the temperature, humidity, and air pressure environmental parameter types, and the matrix element values ​​store the numerical values ​​of the convolution operation results.

[0079] The spatiotemporal distribution features of abnormal breathing segments are converted into a three-dimensional tensor input. The three dimensions of the three-dimensional tensor represent the time dimension, spatial dimension, and feature dimension, respectively. The time dimension contains the time-series information of continuous breathing cycle segments, the spatial dimension contains the layout and location information of the multimodal sensors, and the feature dimension contains the respiratory parameter features extracted from waveform marker points. The spatiotemporal convolutional kernel adopts a three-dimensional convolutional structure, which performs sliding calculations on the three-dimensional tensor. The sliding stride is set to one step in the time dimension, one step in the spatial dimension, and one step in the feature dimension. The size of the spatiotemporal convolutional kernel is set to five steps in the time dimension, three steps in the spatial dimension, and seven steps in the feature dimension. The weight parameters of the convolutional kernel are optimized through the training process. Local abnormal pattern extraction is achieved through feature mapping of the convolutional layers. Each convolutional kernel generates a feature map, and the highly activated regions of the feature map correspond to abnormal breathing patterns.

[0080] The gated recurrent unit (ROU) network receives the feature sequence output from the spatiotemporal convolutional layer. The internal structure of the ROU includes update and reset mechanisms. The update gate controls the proportion of historical information passed to the current state, while the reset gate controls the contribution of historical information to the current candidate state. The time step of the ROU is set to thirty consecutive respiratory cycles, and the number of hidden layer nodes is set to sixty-four. The propagation characteristics of abnormal patterns in the temporal dimension are captured through the hidden states of the ROU, which contain memory information of historical abnormal patterns. The output layer of the respiratory anomaly analysis model adopts a fully connected structure. The fully connected layer maps the last hidden state of the ROU to a respiratory pattern deviation score, which is set to a continuous value from zero to one hundred.

[0081] The generation of the respiratory environment coupling matrix includes data standardization preprocessing. Environmental parameter data and respiratory rhythm stability indicators are standardized, transforming each parameter into a distribution with a mean of zero and a standard deviation of one. Zero-padding is used for boundary handling in convolution operations, adding zero points at the beginning and end of the environmental parameter data sequence to ensure the size of the convolution output matrix is ​​consistent with the length of the input sequence. Significant interference regions are labeled using a binarization method: matrix elements with absolute values ​​greater than a threshold are labeled as one, and those less than or equal to the threshold are labeled as zero. The labeling results generate a mask layer for the respiratory environment coupling matrix, which is used to filter abnormal respiratory segments affected by environmental interference. The extraction of spatiotemporal distribution features of abnormal respiratory segments includes timestamp alignment, precisely matching the timestamps of abnormal respiratory segments with the environmental parameter acquisition timestamps, achieving millisecond-level matching accuracy.

[0082] The construction of the 3D tensor includes feature normalization, which scales different respiratory parameters to the same order of magnitude, avoiding the impact of differences in feature value ranges on model training. The ReLU function is chosen as the activation function for the spatiotemporal convolutional layer. ReLU retains positive values ​​and suppresses negative values, enhancing feature sparsity. The identification of local abnormal patterns relies on the weight distribution of the convolutional kernels; during training, the convolutional kernels learn the spatiotemporal feature patterns of abnormal breathing. The input dimension of the gated recurrent unit is consistent with the number of output feature maps of the spatiotemporal convolutional layer, and the time step of the gated recurrent unit adopts a sequence-to-sequence mapping method. The training data for the respiratory abnormality analysis model includes labeled normal and abnormal respiratory segments, which are determined by pulmonologists according to clinical standards. The model training uses the backpropagation algorithm, which optimizes the weights of the spatiotemporal convolutional kernels and the parameters of the gated recurrent unit. The mean squared error function is chosen as the loss function.

[0083] The application of the respiratory environment coupling matrix includes matrix factorization, employing singular value decomposition (SVD) to extract the principal components of the matrix. These principal components correspond to the main influence patterns of environmental factors on breathing patterns. The spatiotemporal distribution feature analysis of abnormal breathing segments includes hotspot detection, which identifies regions where abnormal segments cluster in the spatiotemporal distribution, indicating persistent respiratory abnormalities. Data augmentation of the 3D tensor utilizes random pruning and rotation transformations to increase the diversity of training samples and improve the model's generalization ability. Gradient pruning of the gated recurrent unit sets a threshold to prevent gradient explosion during training. The inference phase of the respiratory anomaly analysis model employs a sliding window mechanism, with the sliding window length consistent with that of the training phase, enabling continuous real-time anomaly detection.

[0084] The calculation of respiratory rhythm stability indices includes trend component removal, which employs a multinomial fitting method to eliminate long-term trend changes in respiratory rhythms while retaining short-term fluctuations. Quality control of environmental parameter data includes outlier removal, which uses box plot statistics to identify and exclude outliers in environmental parameters. Parallelization of convolution operations is accelerated using a graphics processing unit (GPU). The GPU's parallel computing core processes multiple convolution windows simultaneously, improving computational efficiency. The visualization of the respiratory-environment coupling matrix uses a heatmap format, where color intensity represents coupling strength, facilitating intuitive analysis of respiratory-environment interactions. The spatiotemporal distribution characteristics of abnormal respiratory segments are stored using a hierarchical data format, supporting rapid querying and retrieval of large-scale spatiotemporal data. The visualization of the hidden states of gated recurrent units uses principal component analysis (PCA), which projects high-dimensional hidden states onto a two-dimensional plane to observe the evolution trajectory of abnormal patterns. The deployment of the respiratory anomaly analysis model employs an embedded system solution, integrating sensor interfaces and data processing modules to achieve edge computing capabilities.

[0085] Example 4: Referring to Table 1, this example details the implementation of a multi-level early warning mechanism triggering and cross-device data verification. The grading threshold for the respiratory pattern deviation score is set based on clinical respiratory pathology standards. The numerical range of the respiratory pattern deviation score is divided into three continuous intervals, each corresponding to a different type of respiratory abnormality and an early warning level. The triggering condition of the multi-level early warning mechanism is determined by the real-time value of the respiratory pattern deviation score. When the respiratory pattern deviation score falls into a specific interval, the corresponding early warning response process is automatically triggered.

[0086] Table 1: Thresholds and Response Measures for Multi-Level Early Warning Mechanism

[0087] Warning level Rating range Warning type Notification method Response measures Level 1 warning 0-30 Abnormal breathing rhythm Mobile app push Display breathing training instructions Level II Warning 31-70 Warning of respiratory muscle dysfunction SMS notification Outpatient follow-up examination recommended Level III Warning 71-100 Alarm of abnormal respiratory center regulation Automatic telephone calling Activate emergency medical response

[0088] The abnormal respiratory rhythm alert includes a detailed description of the abnormal respiratory parameters, listing the degree of deviation from indicators such as respiratory rate, inspiratory time ratio, and tidal volume coefficient of variation. The respiratory muscle function abnormality alert includes an additional respiratory muscle fatigue index assessment, calculated using the nonlinear characteristics of chest and abdominal movement signals. The respiratory center regulation abnormality alert integrates real-time vital sign monitoring data, including blood oxygen saturation and heart rate variability. The alert report generation uses a templated document structure, containing four parts: basic patient information, a timeline of the abnormal respiratory segment, analysis of environmental influencing factors, and recommended measures.

[0089] The construction of the backtracking path for abnormal breathing segments employs time-series graph technology. Time-series graph nodes represent the starting points of abnormal breathing segments, and edges represent the causal relationships between segments. The visualization of the backtracking path utilizes an interactive timeline design, supporting zooming and detailed viewing. The cross-device data verification process initiates data synchronization across multiple wearable devices, including sensor platforms such as smart bracelets, smart belts, and smart patches. User posture change data recorded by the smart bracelet is acquired through a nine-axis inertial measurement unit (IMU), which includes a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer.

[0090] The detection of postural change time employs a posture recognition algorithm, which categorizes inertial measurement unit (IMU) data into five basic postures: standing, sitting, supine, prone, and lateral. A sliding window matching strategy is used to compare the postural change timestamp with the start timestamp of the abnormal respiratory segment. The sliding window width is set to 60 seconds, and the matching tolerance threshold is set to 10 seconds. When the difference between the postural change time and the start time of the abnormal respiratory segment is less than 10 seconds, the system automatically marks it as a posturally related respiratory abnormality. The criteria for judging posturally related respiratory abnormalities include changes in diaphragmatic position and intrathoracic pressure caused by postural changes. When the postural change time and the abnormal respiratory segment do not overlap, the system marks it as a pathological respiratory abnormality. The determination of pathological respiratory abnormalities refers to the duration and fluctuation characteristics of the respiratory pattern deviation score.

[0091] The early warning report distribution mechanism employs a multi-channel notification model, including mobile application push notifications, SMS alerts, and automated voice calls. Mobile application push notifications include a brief early warning summary and a link to the detailed report; SMS alerts include the early warning level and emergency contact number; and automated voice calls use text-to-speech technology to generate voice alert messages. Early warning reports are stored in a structured database, which includes fields such as early warning event number, trigger time, early warning level, processing status, and feedback result.

[0092] The cross-device data verification process includes a data quality check step, which examines the synchronization accuracy and data integrity of the timestamps of each device. The smart bracelet's posture data is synchronized with the respiratory monitoring device using a network time protocol, which synchronizes the clocks of each device to millisecond-level accuracy. Abnormal respiratory segments are classified and labeled using multidimensional feature vectors, which include respiratory parameter features, temporal features, and device coordination features. Verification of posture-related respiratory anomalies includes respiratory pattern analysis, which compares the consistency of respiratory waveform characteristics before and after changes in posture.

[0093] The determination of pathological respiratory abnormalities includes an elimination process, which progressively eliminates non-pathological factors such as environmental interference, equipment errors, and body position. The backtracking path of abnormal respiratory segments in the early warning report is displayed graphically, using line graphs to show the trend of respiratory parameter changes and heatmaps to show the distribution density of abnormal segments. The confidence assessment of cross-device data validation results employs a probabilistic model, which comprehensively considers factors such as time synchronization accuracy, data quality score, and equipment reliability index.

[0094] The multi-level early warning mechanism requires the system to trigger an early warning within five seconds of calculating the breathing pattern deviation score, and the early warning report generation time must be controlled within ten seconds. The parallel processing architecture for cross-device data verification processes data streams from multiple wearable devices simultaneously, employing a distributed computing framework for efficient data comparison. The spatiotemporal tagging of abnormal breathing segments includes geographic coordinate information derived from the device's GPS module, used to analyze the impact of environmental factors on breathing abnormalities.

[0095] The correlation analysis between postural changes and respiratory abnormalities employs a causal model, which calculates the probability correlation between postural change events and respiratory abnormality events. Customizable settings for early warning reports allow healthcare professionals to adjust warning thresholds and notification methods; the customization interface provides slider controls and option menus for parameter configuration. A failure handling mechanism for cross-device data verification includes data retransmission requests; the system automatically initiates a data retransmission command when data synchronization between devices fails. Persistent storage of abnormal respiratory segments utilizes a cloud storage solution, ensuring data security and accessibility, and supporting historical data review and statistical analysis.

[0096] The multi-level early warning mechanism's feedback loop includes an early warning effectiveness evaluation, which is comprehensively assessed based on feedback scores from medical staff and the patient's subsequent health status. The optimized algorithm for cross-device data verification employs machine learning technology, which continuously refines the association rules between postural changes and respiratory abnormalities, improving the accuracy of abnormality labeling. The digital signature mechanism for early warning reports uses asymmetric encryption technology, ensuring the integrity and non-repudiation of early warning reports and complying with medical data security regulations.

[0097] Specialized management of position-related respiratory abnormalities includes postural adaptation training recommendations, which provide personalized breathing adjustment plans based on the specific type of postural change. Emergency management procedures for pathological respiratory abnormalities include automatically generated referral recommendations, which recommend appropriate specialist medical institutions based on the severity of the abnormality and the patient's historical medical history. The entire multi-level early warning mechanism and cross-device data verification system's operational status monitoring utilizes a health check interface. This interface periodically checks the working status of each component to ensure system reliability and stability.

[0098] Example 5: This example describes a specific implementation method for constructing a respiratory behavior association map and updating a respiratory abnormality analysis model. The implementation process is illustrated below with a specific example: Assume a user exhibits three sets of abnormal respiratory segments within a continuous 24-hour monitoring period. The timestamps of these abnormal respiratory segments are T1, T2, and T3, with corresponding respiratory pattern deviation scores of 45, 62, and 58, respectively. User activity logs show a positional change event (from sitting to standing) 5 minutes before time point T1, eating activity 10 minutes before time point T2, and no significant behavioral events recorded near time point T3.

[0099] The respiratory behavior correlation graph is constructed using three abnormal respiratory segments as nodes. Each node stores a complete respiratory parameter feature vector, which includes specific values ​​such as respiratory rate, tidal volume, and inspiratory time ratio. The temporal correlation between the behavioral event timestamp and the abnormal respiratory segment is represented by directed edges, with the edge direction pointing from the behavioral event to the abnormal respiratory segment. The similarity of respiratory parameters between nodes is calculated using a cosine similarity algorithm, comparing the spatial angle between the feature vectors of every two abnormal respiratory segments. The behavioral time interval is taken as the absolute value of the time difference between the start times of adjacent abnormal respiratory segments.

[0100] The coupling weight is calculated by combining two factors: respiratory parameter similarity and behavioral time interval. The specific formula is: Coupling Weight = Respiratory Parameter Similarity × (1 / Behavioral Time Interval). The respiratory parameter similarity ranges from 0 to 1, and the behavioral time interval is in minutes. Assuming the respiratory parameter similarity between segments T1 and T2 is 0.8 and the time interval is 120 minutes, then the coupling weight = 0.8 × (1 / 120) = 0.0067. The retention threshold is set to 0.005; all edges below this value will be pruned.

[0101] In the constructed respiratory behavior association map, the identification of pathological respiratory abnormalities is based on node attributes and edge relationship features. When the abnormal respiratory segment has a weak temporal correlation with the behavioral event, but a high similarity in respiratory parameters, this type of abnormality is marked as pathological respiratory abnormality. The cluster density of pathological respiratory abnormalities in the respiratory behavior association map is calculated using the formula: Cluster density = Actual number of edges between pathological abnormal nodes / Maximum possible number of edges.

[0102] The threshold update mechanism is based on the dynamic changes in cluster density. The historical baseline is the average cluster density of the user over the past 7 days, assumed to be 0.35. The cluster density calculated for the current monitoring period is 0.42, exceeding the historical baseline by 20%. At this time, the system lowers the threshold for the breathing pattern deviation score by a preset ratio, by 5% of the original threshold. Assuming the original threshold was 40, the new threshold will be adjusted to 38.

[0103] The respiratory anomaly analysis model is updated using an incremental learning algorithm, and the new threshold values ​​take effect immediately for subsequent anomaly detection. The system records a complete threshold adjustment log, including detailed information such as adjustment time, threshold before adjustment, threshold after adjustment, and reason for adjustment. It also retains model parameters from previous versions for rollback when needed.

[0104] The closed-loop feedback mechanism is achieved through continuous monitoring of model performance metrics. The system tracks anomaly detection accuracy, recall, and other metrics in real time after the judgment threshold is adjusted, ensuring that model updates do not lead to performance degradation. When a significant change in model performance is detected, the system automatically triggers a model recalibration process to maintain the stability and reliability of the respiratory health monitoring system.

[0105] The entire implementation process demonstrates a complete closed loop of respiratory behavior association map construction and model parameter updates, showcasing the entire process from data acquisition and map construction to model optimization through specific examples. This implementation method ensures that the respiratory abnormality detection system can adapt to changes in individual user characteristics, improving the accuracy and practicality of monitoring.

[0106] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for remote respiratory health monitoring and management, characterized in that, The method includes: By acquiring users' respiratory physiological signals through multimodal sensors, a raw respiratory dataset containing temporal and frequency domain features is generated. The original respiratory dataset is dynamically segmented to divide it into continuous respiratory cycle segments, and key waveform markers are extracted from each respiratory cycle segment. A respiratory state transition network is constructed based on key waveform markers, and waveform markers of adjacent respiratory cycle segments are associated to form a dynamic topology that reflects the evolution of respiratory patterns. The construction of the respiratory state transition network based on key waveform marker points specifically includes: Identify the amplitude difference between peaks and troughs in a respiratory cycle segment as waveform marker points; Establish the transition probability between the waveform marker point of the current respiratory cycle segment and the waveform marker point of the next respiratory cycle segment; When the transition probability exceeds the preset connectivity threshold, directed edges are drawn between the waveform marker points to form a dynamic topology. The respiratory rhythm stability index is calculated based on dynamic topology, and a respiratory environment coupling matrix is ​​generated by combining temperature, humidity and air pressure data in the environmental parameter dataset. The specific steps of generating the respiratory environment coupling matrix by combining temperature, humidity, and air pressure data from the environmental parameter dataset include: Extract the coefficient of variation of inspiratory duration from the respiratory rhythm stability index; The coefficient of variation is convolved with the temperature and humidity gradients in the environmental parameter dataset. Matrix elements in the output convolution result whose absolute value is greater than the critical value are used as identifiers of disturbed abnormal breathing segments; Abnormal breathing segments affected by environmental interference are screened out by the breathing environment coupling matrix, and the spatiotemporal distribution characteristics of the abnormal breathing segments are marked to train the breathing abnormality analysis model and output the breathing pattern deviation score. A multi-level early warning mechanism is triggered based on the breathing pattern deviation score, generating an early warning report that includes the backtracking path of abnormal breathing segments; Perform cross-device data verification on abnormal breathing segments in the early warning report and match them with the timestamps of behavioral events in the user activity log; By integrating and verifying abnormal breathing segments with behavioral event timestamps, a respiratory behavior correlation map is constructed. The threshold for judging respiratory abnormality analysis model is updated based on respiratory behavior correlation map to complete the closed-loop feedback of respiratory health monitoring; The construction of the respiratory behavior association graph specifically includes: taking abnormal respiratory segments as nodes and the order of the timestamps of the behavioral events as edges, calculating the coupling weight between the similarity of respiratory parameters and the time interval of the behavior between nodes, and pruning the edges in the association graph whose weights are lower than the retention threshold according to the coupling weights; The specific thresholds for updating the respiratory abnormality analysis model based on respiratory behavior correlation maps include: The cluster density of pathological respiratory abnormalities in the respiratory behavior association map is statistically analyzed. When the cluster density exceeds the historical benchmark value, the judgment threshold of the respiratory pattern deviation score is lowered by a preset ratio. When the cluster density is lower than the historical benchmark value, the judgment threshold is reset to the initial default value.

2. The method for remote respiratory health monitoring and management according to claim 1, characterized in that, The acquisition of user respiratory physiological signals via multimodal sensors specifically includes: Simultaneously acquire displacement signals output by the chest and abdominal motion sensor, respiratory sound signals collected by the microphone array, and nasal airflow temperature change signals generated by the infrared thermal imager; Baseline drift correction is performed on displacement signals, environmental noise suppression is performed on breath sound signals, and dynamic threshold segmentation is performed on temperature change signals. The three types of signals were aligned according to sampling time and merged into a raw respiratory dataset with spatiotemporal labels.

3. The method for remote respiratory health monitoring and management according to claim 1, characterized in that, The dynamic segmentation of the original respiratory dataset specifically includes: Analyze the inflection points between the inspiratory and expiratory phases in the original respiratory data set, using the interval between adjacent inflection points as the basic segmentation unit; Calculate the variance contribution rate of the signal within each segment unit, and merge consecutive segment units with variance contribution rates below a threshold; Insert respiratory cycle segment separator markers at the boundaries of the merged segment units.

4. The method for remote respiratory health monitoring and management according to claim 1, characterized in that, The training of the respiratory anomaly analysis model using the spatiotemporal distribution characteristics of the labeled abnormal respiratory segments specifically includes: The spatiotemporal distribution characteristics of abnormal respiratory segments are converted into three-dimensional tensor inputs; Local anomaly patterns of tensors are extracted using spatiotemporal convolution kernels; The propagation characteristics of abnormal patterns over time are captured by a gated loop unit.

5. The method for remote respiratory health monitoring and management according to claim 1, characterized in that, The multi-level early warning mechanism triggered based on the breathing pattern deviation score specifically includes: When the breathing pattern deviation score is in the first interval, an abnormal breathing rhythm prompt is generated; When the breathing pattern deviation score is in the second interval, a warning of abnormal respiratory muscle function is generated. When the respiratory pattern deviation score is in the third interval, an alarm for abnormal respiratory center regulation is generated.

6. The method for remote respiratory health monitoring and management according to claim 1, characterized in that, The cross-device data verification of abnormal respiratory segments in the early warning report specifically includes: Compare the timestamps of abnormal breathing segments with the time of user position changes recorded by the smart bracelet; When the difference between the time of body position change and the time of onset of abnormal breathing segment is less than the set tolerance, it is marked as a position-related respiratory abnormality. When the time of body position change does not overlap with the abnormal breathing segment, it is marked as pathological respiratory abnormality.

7. A remote respiratory health monitoring and management system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the remote respiratory health monitoring and management method according to any one of claims 1 to 6.

Images