Methods, systems, and apparatuses for respiratory signal monitoring and analysis

Abstract

The application belongs to the technical field of medical devices and artificial intelligence cross-fusion, and specifically discloses a method, system and device for respiratory signal monitoring and analysis. The method comprises data acquisition, data preprocessing and normalization processing, graph construction and spectrum diffusion modeling, state space-based time series modeling, state evaluation and early warning. The system is based on the method and comprises a sensor module, a data processing module, a core processing module and an output and feedback module. The device comprises a sensor, a storage, a processor and a computer program stored on the storage and executable on the processor, and the processor is electrically connected with the sensor and the storage. The application combines graph diffusion modeling and state space prediction mechanism, fully utilizes the learning ability of graph spectrum diffusion neural network on structure information and the efficiency advantage of state space sequence modeling module on time pattern modeling, and realizes continuous, accurate and visual monitoring.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical devices and artificial intelligence cross-fusion, and particularly relates to a method, system and device for respiratory signal monitoring and analysis. BACKGROUND

[0002] Chronic Obstructive Pulmonary Disease (COPD) is a group of progressive diseases characterized by airflow limitation, with high prevalence, high morbidity and high mortality. Early identification and long-term dynamic monitoring are the key to control the development of chronic obstructive pulmonary disease and delay the deterioration of the disease. However, the current mainstream respiratory monitoring methods in clinical practice mainly include lung function detector, blood oxygen saturation detector and portable flow meter measurement, etc. These methods have several shortcomings: high dependence on patient compliance, forced exhalation required for some tests, difficult to perform continuously; large device size or poor data transmission capability, difficult to adapt to long-term monitoring in daily scenarios; only shallow static parameters can be obtained, and deep dynamic changes and potential pathological patterns cannot be analyzed.

[0003] In recent years, deep learning has been gradually introduced into respiratory pattern recognition research. Traditional algorithms such as support vector machine, decision tree and random forest have shown strong classification performance in the early stage, but have problems such as weak time series modeling ability, dependence on hand-designed features for feature extraction, and difficulty in dealing with complex multi-dimensional physiological data. With the development of deep learning, convolutional neural networks (CNN) and long short-term memory networks (LSTM) have been gradually used for automatic feature extraction and anomaly detection of respiratory process data. However, CNN can only extract local spatial features, and LSTM has low computational efficiency in long sequence modeling and is difficult to handle in parallel, which cannot meet the requirements of model accuracy and real-time performance.

[0004] Especially in chronic obstructive pulmonary disease, which requires long-term continuous tracking and is highly sensitive to the "slow change + sudden deterioration" pattern of disease management, existing models lack the ability to jointly model the temporal structure and high-order correlation between features. SUMMARY

[0005] In view of the above problems, the purpose of the present application is to provide a method, system and device for respiratory signal monitoring and analysis, which realizes high-precision, low-delay and strong-explanation dynamic monitoring and abnormal warning of the respiratory state of chronic obstructive pulmonary disease patients.

[0006] The technical solution of the present application is: a method for respiratory signal monitoring and analysis, comprising:

[0007] Obtaining respiratory process data of a chronic obstructive pulmonary disease patient to form an original signal set.

[0008] Data preprocessing is performed on the respiratory process data in the original signal set, and then normalization processing is performed to obtain a normalized signal tensor , wherein is the time length, is the feature dimension, represents the real number field, represents a row column matrix, each row of the matrix represents a certain time point, and each column represents a certain respiratory-related feature channel.

[0009] The normalized respiratory process data is regarded as a graph signal, and a node set is constructed at each time step, a weighted adjacency matrix is established, a graph spectrum diffusion operation is performed based on a graph Laplacian operator, and a graph structure time sequence feature tensor is output based on a graph spectrum diffusion neural network .

[0010] A state space model based on a state space sequence modeling module processes the graph structure time sequence feature tensor Z and outputs a chronic obstructive pulmonary disease state score or an abnormal risk index at time .

[0011] Based on the chronic obstructive pulmonary disease state score and the abnormal risk index at time , a time graph of a patient's respiratory age curve and risk index is constructed, the respiratory age is used to reflect the course stage, and the risk index is used to reflect the distance from the acute exacerbation threshold of chronic obstructive pulmonary disease, and a warning is given when the risk index is greater than the threshold.

[0012] Further, the respiratory process data includes respiratory airflow rate, inspiratory pressure, respiratory heat and humidity change amount, respiratory frequency, tidal volume, and ventilation volume.

[0013] After the respiratory process data is collected, an original signal set is formed, and the original signal set is represented as:

[0014] , wherein represents the respiratory airflow rate, represents the inspiratory pressure, represents the respiratory heat and humidity change amount, represents the respiratory frequency, represents the tidal volume, represents the ventilation volume, represents a multi-dimensional vector at the same time, representing a respiratory parameter set.

[0015] Further, the normalized signal tensor is obtained according to , wherein, , denotes a feature i in a time series.

[0016] Further, the weighted adjacency matrix is obtained according to ; wherein, denotes the th element in the weighted adjacency matrix i,j ; , denotes a feature i in a time series, , denotes a feature j in a time series. denotes a feature between ; denotes a decay rate of the control similarity function.

[0017] Further, the graph Laplacian is , wherein, , denotes a reserved structure of a node itself in the graph; , denotes a weighted adjacency matrix; , D denotes the sum of all edge weights connected to the node i , i.e., the degree of the node; denotes a symmetric normalized adjacency matrix.

[0018] Further, the state space sequence modeling module is represented as a recursion of a hidden state sequence .

[0019] .

[0020] wherein, denotes a system state vector, denotes input graph structure time series features , A 1, B 1, C 1, D 1 each denote a trainable state transition matrix group.

[0021] By continuous state transition optimization, a nonlinear dynamic evolution pattern of the graph structure time series features is learned, and a chronic obstructive pulmonary disease state score or abnormal risk index at a time is obtained according to the following formula.

[0022] wherein, , represents the COPD prediction score result at the current time, which is a scalar real value; is a mapping function defined by the state space sequence modeling module, representing a prediction mechanism from the input graph structure time series features to the output risk score .

[0023] Further, the respiratory age function corresponding to the respiratory age curve is: wherein, represents the COPD prediction score result at the current time.

[0024] A system for respiratory signal monitoring and analysis based on the method, comprising:

[0025] a sensor module for acquiring respiratory process data of a COPD patient;

[0026] a data processing module, comprising: a data preprocessing unit for noise removal and normalization processing of the respiratory process data collected by the sensor module; a feature extraction unit for extracting respiratory-related feature channels from the data after noise removal and normalization processing;

[0027] a core processing module for analyzing and detecting abnormalities of the respiratory pattern based on the respiratory-related feature channels;

[0028] an output and feedback module, comprising: a real-time monitoring and analysis unit for applying the trained model to real-time data monitoring; an alarm unit for early warning when the risk index is greater than a threshold value.

[0029] An apparatus for respiratory signal monitoring and analysis, comprising a sensor for acquiring respiratory process data of a COPD patient, a storage, a processor, and a computer program stored on the storage and executable on the processor, the processor being electrically connected with the sensor and the storage, for receiving the respiratory process data and executing the computer program to implement the method.

[0030] Compared with the prior art, the present application has the advantages that the present application combines graph diffusion modeling and state space prediction mechanism, fully utilizes the learning ability of graph spectrum diffusion neural network on structure information and the efficiency advantage of state space sequence modeling module on time pattern modeling, realizes continuous, accurate and visual monitoring, and has important medical practical value and engineering innovation.

[0031] The present invention innovatively introduces a graph construction mechanism, maps multi-source physiological data into a graph with a topological structure between nodes, and realizes diffusion modeling and joint extraction of respiratory features in the spectral domain through a graph spectrum diffusion neural network. A lightweight state space modeling framework state space sequence modeling module is adopted, based on a linear dynamic system, and long-term dependencies are modeled through a recursive structure, which significantly improves the prediction speed and resource adaptability, and is suitable for continuous monitoring environments. A hybrid neural network architecture that integrates the graph spectrum diffusion neural network and the state space sequence modeling module structure is designed, and the concept of respiratory age is proposed to quantify the individual disease course status, and a respiratory deterioration risk index is constructed to achieve early prediction and intervention of acute exacerbation events of chronic obstructive pulmonary disease. Intelligent monitoring of the entire respiratory process from "multimodal sensing-deep feature modeling-state assessment-personalized prediction" is particularly suitable for early screening, long-term tracking and graded warning of diseases such as chronic obstructive pulmonary disease, asthma, and sleep apnea, and has broad application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flowchart of Example 1 of the present invention. DETAILED DESCRIPTION

[0033] The following combination Figure 1 , a detailed description of the specific embodiments of the present invention is provided. In the description of the present invention, it should be understood that the terms "center," "upper," "lower," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or location relationships, are based on the positions or location relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed or operate in a specific orientation, and therefore should not be construed as limiting the present invention.

[0034] The terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more of such features; in the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0035] It should be noted that the circuit connections involved in the present invention all adopt conventional circuit connection methods and do not involve any innovation.

[0036] In recent years, graph neural networks (GNN) have become an important method for processing structured spatio-temporal data. The graph spectrum diffusion neural network, GSDNet, is a graph model from the perspective of spectral diffusion, which has the ability to model graph structure data in the frequency domain and is suitable for multi-modal graph structure representation of multi-sensor respiratory data. The state space sequence modeling module (Mamba) is a new state space modeling framework that has higher long sequence modeling capability, inference efficiency, and data memory capacity compared to traditional RNN and Transformer architectures. The fusion of the two brings a new solution to the representation and prediction of respiratory time series data.

[0037] In summary, how to combine advanced graph frequency domain modeling and efficient time series modeling framework to realize real-time and high-precision monitoring and early warning of the respiratory state of patients with chronic obstructive pulmonary disease is a key scientific problem that needs to be solved in the current field of intelligent medical care and respiratory disease management.

[0038] Embodiment 1

[0039] As shown in a method for respiratory signal monitoring and analysis, comprising: Figure 1

[0040] S1, data acquisition

[0041] Obtain the respiratory process data of patients with chronic obstructive pulmonary disease to form an original signal set.

[0042] S2, data preprocessing and normalization processing

[0043] The respiratory process data in the original signal set is preprocessed, and after Min-Max normalization processing, a normalized signal tensor is obtained , wherein is the time length, is the feature dimension, represents the real number field, represents a row column matrix, each row in the matrix represents a certain time point, and each column represents a certain respiratory-related feature channel. It should be noted that the signal tensor is a mathematical quantity that can describe multiple variables. After normalizing the respiratory process data into a signal tensor, it can be better processed by a computer. Since the respiratory process data includes multiple variables, it needs to be tensorized to normalize the respiratory process data into a signal tensor.

[0044] S3, graph construction and spectral diffusion modeling ​

[0045] The normalized respiratory process data is regarded as a graph signal, and a node set is constructed at each time step , a weighted adjacency matrix is established , a graph spectrum diffusion operation is performed based on a graph Laplacian operator, and a spectrum-encoded graph structure time series feature tensor is output based on a graph spectrum diffusion neural network .

[0046] S4, state space-based time series modeling

[0047] The state space sequence modeling module processes the graph structure time series feature tensor Z to output a chronic obstructive pulmonary disease state score or an abnormal risk index at time .

[0048] S5, state assessment and early warning

[0049] Based on the chronic obstructive pulmonary disease state score and the abnormal risk index at time , a time graph of the patient's respiratory age curve and risk index is constructed, the respiratory age is used to reflect the disease stage, and the risk index is used to reflect the distance from the chronic obstructive pulmonary disease acute exacerbation threshold. When the risk index is greater than the threshold, an early warning is given.

[0050] At present, traditional models usually process gas flow, temperature and humidity, pressure and other sensor data in series, ignoring the structural correlation between these features, resulting in low feature fusion efficiency and poor pattern recognition ability. The embodiment innovatively introduces a graph construction mechanism, which maps multi-source physiological data into a graph with topological structure between nodes, and realizes diffusion modeling and joint extraction of respiratory features in the spectral domain through a graph spectrum diffusion neural network.

[0051] The existing LSTM, Transformer and other time series models have high computational complexity, high training difficulty and low inference efficiency, and are difficult to deploy on edge devices for execution. The embodiment adopts a lightweight state space modeling framework, a state space sequence modeling module, which is based on a linear dynamic system, models long-term dependencies through a recursive structure, significantly improves the prediction speed and resource adaptation ability, and is suitable for continuous monitoring environments.

[0052] Traditional methods are mostly limited to respiratory abnormality recognition, lack the ability to model the evolution of individual disease progression, and cannot achieve personalized intervention decisions. The embodiment designs a hybrid neural network architecture that integrates a graph spectrum diffusion neural network and a state space sequence modeling module structure, proposes a respiratory age concept to quantify the individual disease state, and constructs a respiratory exacerbation risk index to realize early prediction and intervention of chronic obstructive pulmonary disease acute exacerbation events.

[0053] The embodiment provides an intelligent respiratory monitoring solution from the whole process of "multi-modal sensor, deep feature modeling, state evaluation and individualized prediction", which is especially suitable for early screening, long-term tracking and hierarchical early warning of chronic obstructive pulmonary disease, asthma, apnea and the like, and has wide application prospect and popularization value.

[0054] Preferably, the respiratory process data includes respiratory airflow rate, inspiratory pressure, respiratory heat and moisture change, respiratory frequency, tidal volume and ventilation volume.

[0055] The respiratory process data is collected to form an original signal set, and the original signal set is represented as:

[0056] wherein, represents the respiratory airflow rate, represents the inspiratory pressure, represents the respiratory heat and moisture change, represents the respiratory frequency, represents the tidal volume, represents the ventilation volume, represents a multi-dimensional vector at the same time, representing the respiratory parameter set.

[0057] Preferably, the normalized signal tensor is obtained according to , wherein, , represents the feature i vector in the time sequence.

[0058] Preferably, the weighted adjacency matrix is obtained according to ; wherein, , represents the th item in the weighted adjacency matrix i,j , representing the connection strength between the node i and the node j . , represents the feature i vector in the time sequence, , represents the feature j vector in the time sequence. represents the Euclidean distance between the features and , for measuring the difference degree of them in the time dimension; is a scale factor, representing the decay rate of the similarity function, represents that the function is monotonically decreasing; denotes an exponential function, ensuring the range of the adjacency weight values .

[0059] Preferably, the Turbopascal operator is wherein, , is an identity matrix, representing the reserved structure of the node itself; , is a weighted adjacency matrix, the element in the matrix i,j denotes the connection strength between node and node i ; j , , is a degree matrix, which is a diagonal matrix, wherein the diagonal elements are: ; D represents the sum of all edge weights connected to node i , i.e., the degree of the node; is a symmetric normalized adjacency matrix, used to ensure edge weight normalization, uniform feature scale, and avoid training bias caused by node degree difference.

[0060] Preferably, the state space sequence modeling module is represented as a recursion of a hidden state sequence .

[0061] .

[0062] wherein, denotes a system state vector, denotes input graph structure time series features , A 1, B 1, C 1, D 1 each denote a trainable state transition matrix group.

[0063] Through continuous state transition optimization, the nonlinear dynamic evolution pattern of the graph structure time series features is learned, and the chronic obstructive pulmonary disease state score or abnormal risk index at time is obtained according to the following formula.

[0064] wherein, , denotes the chronic obstructive pulmonary disease prediction score result at the current time, which is a scalar real value; is a mapping function defined by the state space sequence modeling module model, representing the prediction mechanism from the input graph structure time series features to the output risk score .

[0065] Preferably, the respiratory age function corresponding to the respiratory age curve is: wherein, represents the chronic obstructive pulmonary disease prediction score result at the current moment.

[0066] Preferably, the risk index corresponding to the risk index time graph is wherein, represents the chronic obstructive pulmonary disease prediction score result at the current moment.

[0067] The method for respiratory signal monitoring and analysis proposed in this embodiment has an advanced algorithm structure, realizes deep dynamic modeling and high-precision identification, and specifically has the following performances:

[0068] The double-module fusion structure breaks the performance bottleneck of a single model: the graph spectrum diffusion neural network (GSDNet) is innovatively combined with the state space sequence modeling module (Mamba), the former focuses on structural modeling between features, and the latter is good at processing dynamic dependencies over a long time span, so that the model has both local graph frequency extraction capability and global time prediction performance.

[0069] The high-dimensional spatiotemporal signal modeling capability is strong: by constructing a graph Laplacian matrix and a spectral filtering kernel to graph and encode the respiratory multi-modal signal, the problem of difficult processing of feature heterogeneity and strong nonlinearity in traditional models is overcome, and the sensitivity of the model to slight respiratory abnormalities is improved.

[0070] Embodiment 2

[0071] A system for respiratory signal monitoring and analysis, based on the method proposed in embodiment 1, includes a sensor module, a gas flow sensor, a data processing module, a core processing module, and an output and feedback module.

[0072] The sensor module is used to acquire the respiratory process data of the chronic obstructive pulmonary disease patient; the sensor module includes a gas flow sensor, a temperature and humidity sensor, and a pressure sensor.

[0073] The gas flow sensor is used to measure the flow of respiratory airflow and convert it into an electrical signal output, the temperature and humidity sensor is used to measure the temperature and humidity of respiratory airflow and convert it into a readable electrical signal, and the pressure sensor is used to measure the pressure signal of respiratory airflow and convert the pressure signal into a usable electrical signal output.

[0074] The data processing module includes a data preprocessing unit and a feature extraction unit, the data preprocessing unit is used to remove noise and normalize the respiratory process data collected by the sensor module, and the feature extraction unit is used to extract respiratory-related feature channels from the noise-removed and normalized data.

[0075] The core processing module is configured to analyze and detect abnormalities of the respiratory pattern based on the feature channels related to the respiration, and the core processing module is a random forest model.

[0076] The output and feedback module includes a real-time monitoring and analysis unit and an alarm unit, the real-time monitoring and analysis unit is configured to apply the trained model to real-time data monitoring, and the alarm unit is configured to give an early warning when the risk index is greater than a threshold value.

[0077] Preferably, the output and feedback module further includes a result display unit, the result display unit is configured to output the monitoring data and analysis results to a display device or upload to a cloud server.

[0078] Preferably, the system further includes a data storage and transmission module, the data storage and transmission module includes a data storage unit and a data transmission unit, the data storage unit is configured to store the collected raw respiratory data and analysis results, and the data transmission unit is configured to upload the respiratory data to a cloud server through a wireless network

[0079] Embodiment 3

[0080] An apparatus for respiratory signal monitoring and analysis, comprising a sensor, a storage, a processor, and a computer program stored on the storage and executable on the processor, the sensor is configured to acquire respiratory process data of a chronic obstructive pulmonary disease patient, the processor is electrically connected with the sensor and the storage respectively, and is configured to receive the respiratory process data and execute the computer program to realize the method proposed in Embodiment 1.

[0081] The sensor of the present embodiment is a sensing subsystem, which includes a MEMS airflow sensor, a digital temperature and humidity module, and a miniature piezoresistive pressure sensor, respectively collecting multi-modal physiological parameters such as respiratory flow rate, respiratory pressure, and respiratory temperature and humidity.

[0082] The processor adopts an integrated Arm Cortex-M7 processor, deploys a quantized version of the GSDNet+state space sequence modeling module fusion neural network, and supports local inference.

[0083] Preferably, the apparatus further includes a communication module and a display and feedback module. The communication module includes a Bluetooth BLE and a Wi-Fi dual mode, and the communication module is configured to be responsible for data uploading and doctor remote access. The display and feedback module adopts an OLED screen and a buzzer, and can display a state index and give an alarm prompt in real time.

[0084] It should be noted that: the sampling frequency of the apparatus of the present embodiment is set to 50Hz, and after the patient wears the apparatus, the real-time data sequence is as follows:

[0085] .

[0086] The five-order wavelet denoising algorithm is used for pretreatment, and the graph structure adjacency matrix is constructed according to the following formula:

[0087] wherein, is an empirical scale parameter; the input dimension of the graph structure is .

[0088] The GSDNet spectrum diffusion model is configured as:

[0089] The pre-trained GSDNet model contains three graph convolution layers, and the activation function is ELU, and the graph Laplacian adopts a symmetric normalized form: .

[0090] The graph diffusion kernel is a Chebyshev polynomial third-order approximation: wherein, , , and the output feature dimension of the model is .

[0091] The state space sequence modeling module state space modeling and prediction output are:

[0092] The graph encoding features are sent to the lightweight state space sequence modeling module sequence model, and the state transition equation is: wherein, the state dimension is expanded by using causal convolution, and finally the patient respiratory risk index is output.

[0093] Respiratory age and early warning mechanism:

[0094] The system calculates the respiratory age by the following formula.

[0095] wherein, .

[0096] And introduce the sliding window mean detection strategy: .

[0097] If , an alarm is triggered, a red icon is displayed on the device screen, a buzzer sounds, and alarm data is uploaded to the remote doctor platform through Wi-Fi.

[0098] The device for monitoring and analyzing respiratory signals provided in the embodiment has the advantages of flexible deployment, friendly system resources, and adaptation to multiple scene use requirements, which are specifically manifested in:

[0099] Lightweight model structure, edge device deployable: The GSDNet used and the state space sequence modeling module are both efficient neural network structures. Compared with models such as the Transformer, the number of parameters is reduced by about 47%, significantly improving the deployment convenience of the model, and the model can be run in real time in low-power chips and mobile devices, and is suitable for multiple scenes such as home, community and bedside.

[0100] Strong real-time performance, excellent delay control: Tests show that the average end-to-end delay of the model from sensor input to prediction output is less than 200 ms, which can meet the rapid identification and alarm requirements of respiratory event levels such as apnea and shallow and slow syndrome.

[0101] In addition, the device also supports personalized analysis and visual output to adapt to clinical diagnosis and treatment needs, which is specifically manifested in:

[0102] Respiratory age and risk index double-dimensional index system: The device constructs a patient-level respiratory health evaluation system, outputs including respiratory age curve and abnormal risk index curve, and directly reflects the disease development trend to assist doctors in intervention evaluation and individual treatment adjustment.

[0103] Adaptive threshold and early warning module: The alarm threshold can be dynamically adjusted according to the patient's historical state to avoid false positives and false negatives and improve the practicality of the monitoring system.

[0104] Data support remote transmission and doctor terminal access: Combined with the Bluetooth / Wi-Fi module integrated in the device, the model output can be uploaded to the cloud platform for real-time viewing by doctors, and can also be connected with the hospital electronic medical record system for unified management.

[0105] The GSDNet deep learning fusion time sequence COPD analysis device of the embodiment includes four functional modules, namely, a multi-modal data acquisition module, a graph structure construction and spectral diffusion network, a state space sequence modeling module (Mamba), and a real-time monitoring and early warning system, forming an intelligent respiratory monitoring framework combining "graph frequency diffusion + state evolution", which is especially suitable for dynamic monitoring of chronic obstructive pulmonary disease.

[0106] It should be noted that the device proposed in Embodiment 3 can also be extended to sleep apnea detection, high-altitude hypoxia ventilation monitoring, postoperative respiratory function recovery evaluation and other scenes. Without modifying the core algorithm, only the acquisition parameters and alarm thresholds need to be adapted.

[0107] In addition to chronic obstructive pulmonary disease (chronic obstructive pulmonary disease), the technical framework proposed by the device has good transferability and scalability in the following diseases and scenarios: 1. Asthma attack detection and monitoring; 2. Automatic identification of sleep apnea syndrome (OSA); 3. Respiratory dynamics analysis of high-risk patients in ICU; 4. Respiratory mode early warning of new viral pneumonia; 5. Detection of daily respiratory behavior of special groups (elderly, infants). The device is adaptable and can be widely promoted to other chronic disease and respiratory disease management.

[0108] Based on Embodiment 3, the device also has good adaptability and scene migration ability and can be extended to the monitoring tasks of other respiratory-related pathological states.

[0109] This adaptability is due to the structural universality, parameter adjustability, and interface standardization of the GSDNet+state space sequence modeling module model. Without changing the core neural network structure, only by adjusting the input channel configuration, preprocessing rules, and alarm threshold settings, the device can quickly adapt to specific application scenarios such as Table 1.

[0110] Table 1: Application scenario cases that can be adapted based on Embodiment 3

[0111]

[0112] Experimental Example

[0113] The device proposed in Embodiment 3 was tested in 100 patients with chronic obstructive pulmonary disease at different stages, with a test period of 7 consecutive days. The results are shown in Table 2.

[0114] Table 2: Test results of patients with chronic obstructive pulmonary disease at different stages

[0115]

[0116] Conclusion: As shown in Table 2, verification on the experimental data set (>1000 cases of chronic obstructive pulmonary disease patients) shows that the device has an identification accuracy of 94.6%, with a mean absolute error (MAE) of less than 0.12, which is much better than the performance of the simple LSTM model (~88%).

[0117] The above specific embodiments further illustrate the purpose, technical solutions, and benefits of the present application. It should be understood that the above description is only a specific embodiment of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application is included in the protection scope of the present application.

Claims

1. A method for respiratory signal monitoring and analysis, comprising: obtaining respiratory process data of a chronic obstructive pulmonary disease patient, forming a raw signal set; Data preprocessing is performed on the respiratory process data in the original signal set, and then normalized to obtain a normalized signal tensor wherein T is a time length, and n is a feature dimension, denotes a real number field, denotes a T-row n-column matrix, each row of the matrix represents a certain time point, and each column represents a certain respiratory-related feature channel; The normalized respiratory process data is regarded as a graph signal, and a node set V = {v1, v2,..., v n} is constructed at each time step, and a weighted adjacency matrix The graph Laplacian operator is used for graph spectrum diffusion operation, and the graph structure time sequence feature tensor is output based on the graph spectrum diffusion neural network Wherein, d represents the output feature dimension of the graph spectrum diffusion neural network, Indicates a T-row d-column matrix; processing a graph structure time series feature tensor Z based on a state space sequence modeling module state space model, outputting a chronic obstructive pulmonary disease state score or an abnormal risk index at time t; based on the chronic obstructive pulmonary disease state score and the abnormal risk index at time t, constructing a patient's respiratory age curve and a time graph of risk index, the respiratory age is used to reflect the course stage, and the risk index is used to reflect the distance from the acute exacerbation threshold of chronic obstructive pulmonary disease, and a warning is given when the risk index is greater than the threshold.

2. A method for respiratory signal monitoring and analysis as claimed in claim 1, wherein, The respiratory process data includes respiratory airflow rate, inspiratory pressure, respiratory humidification change, respiratory frequency, tidal volume and ventilation volume; After the respiratory process data is collected, a raw signal set is formed, and the raw signal set is represented as: where Q(t) represents the respiratory airflow rate, P(t) represents the inspiratory pressure, H(t) represents the respiratory heat and moisture change, B(t) represents the respiratory frequency, V T (t) represents the tidal volume, MV(t) represents the minute ventilation, represents a multi-dimensional vector at the same time, representing a set of respiratory parameters.

3. A method for respiratory signal monitoring and analysis as claimed in claim 2, wherein, The normalized signal tensor is obtained according to wherein, x i denotes the vector of feature i in the time series.

4. A method for respiratory signal monitoring and analysis as claimed in claim 1, wherein, The weighted adjacency matrix A is given by σ > 0, where A ij represents the i,jth entry of the weighted adjacency matrix A; x i represents the vector of feature i in the time series, x j represents the vector of feature j in the time series;‖x i -x j ‖ 2 represents the Euclidean distance between feature x i and x j ; and σ represents a decay rate that controls the similarity function.

5. A method for respiratory signal monitoring and analysis as claimed in claim 1, wherein, The graph Laplacian is L = I - D -1 / 2 AD -1 / 2 ; wherein, I represents a reserved structure of a node itself in a graph, represents an n-row n-column matrix; A represents a weighted adjacency matrix; D represents the sum of all edge weights connected to node i, i.e., the degree of the node; D -1 / 2 AD -1 / 2 represents a symmetric normalized adjacency matrix.

6. A method for respiratory signal monitoring and analysis as claimed in claim 1, wherein, The state space sequence modeling module is represented as a recursion of an implicit state sequence ε(t): Wherein, ε(t) represents a system state vector, x(t) represents an input graph structure time series feature Z(t), A1, B1, C1, D1 all represent trainable state transition matrix groups, and y(t) represents a chronic obstructive pulmonary disease state score or an abnormal risk index output at time t; Through continuous state transition optimization, the nonlinear dynamic evolution mode of the graph structure time series feature Z(t) is learned, and the chronic obstructive pulmonary disease state score or the abnormal risk index at time t is obtained according to the following formula: wherein, represents the current time point's chronic obstructive pulmonary disease prediction score result; F Mamba (·) represents a prediction mechanism from the input graph structure time series feature Z(t) to the output risk score .

7. A method for respiratory signal monitoring and analysis as claimed in claim 1, wherein, The respiratory age function corresponding to the respiratory age curve is: Wherein, The chronic obstructive pulmonary disease prediction score result at the current time, and a represents a proportional coefficient.

8. A system for respiratory signal monitoring and analysis, characterized by, The method according to any one of claims 1-7, comprising: a sensor module for obtaining respiratory process data of a chronic obstructive pulmonary disease patient; a data processing module, comprising: a data preprocessing unit for removing noise and normalizing the respiratory process data collected by the sensor module; a feature extraction unit for extracting respiratory-related feature channels from the noise-removed and normalized data; a core processing module for analyzing and detecting abnormalities of the respiratory pattern based on the respiratory-related feature channels; an output and feedback module, comprising: a real-time monitoring and analysis unit for applying the trained model to real-time data monitoring; an alarm unit for warning when the risk index is greater than the threshold.

9. An apparatus for respiratory signal monitoring and analysis, characterized by A sensor, a storage, a processor and a computer program stored on the storage and executable on the processor, the sensor is used to obtain respiratory process data of a chronic obstructive pulmonary disease patient, the processor is electrically connected with the sensor and the storage, respectively, for receiving respiratory process data and executing the computer program to realize the method according to any one of claims 1-7.

Images