Portable lung function monitoring method and device based on artificial intelligence

Abstract

The invention discloses a portable lung function monitoring method and device based on artificial intelligence, and relates to the technical field of medical monitoring. The method comprises the following steps: acquiring sparse time sequence data of at least one physiological parameter, and performing inverse problem solving on a preset respiratory system physiological model of which the control equation is a partial differential equation by using a first neural network model so as to identify model parameters representing individual dynamic characteristics in real time; performing forward prediction by using a pre-trained second neural network model based on the identified model parameters and the current physiological state, and generating a rolling prediction sequence of a complete space-time physical field quantity representing the lung function state; a closed-loop correction mechanism is introduced, a physical residual error of a control equation is calculated based on a rolling prediction sequence, and small-step optimization is performed on the prediction sequence according to the physical residual error so as to perform self-adaptive correction. According to the method, the problems that the monitoring dimension of existing portable equipment is single, and long-term prediction of a pure data driving model is unstable are solved.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a portable lung function monitoring method and device based on artificial intelligence. Background Technology

[0002] Lung function monitoring is crucial for the diagnosis, management, and prognostic assessment of respiratory diseases such as chronic obstructive pulmonary disease (COPD) and asthma. Traditional lung function tests, such as spirometry, are typically performed in specialized medical settings like hospitals or clinics, relying on large and complex equipment, making it difficult to continuously and dynamically track a patient's daily physiological state. With the development of the Internet of Things (IoT) and wearable technologies, portable lung function monitoring devices have emerged, offering new possibilities for home-based health management.

[0003] However, existing portable monitoring devices, such as portable spirometers or pulse oximeters, primarily provide macroscopic, single physiological indicators such as forced vital capacity (FVC), forced expiratory volume in one second (FEV1), or oxygen saturation (SpO2). While these indicators have clinical value, they only reflect the overall state of lung function and cannot provide detailed spatiotemporal physical field information regarding airflow, pressure distribution, and lung tissue biomechanics. Furthermore, the data collected by these devices is often temporally sparse and discrete, making it difficult to use to construct comprehensive models capable of predicting the future dynamic evolution of lung function.

[0004] In the field of computational modeling, researchers have attempted to use numerical simulation methods such as computational fluid dynamics (CFD) to construct high-fidelity physiological models of the respiratory system. While these methods can accurately simulate the complex physical processes of the lungs, their computational costs are extremely high, requiring high-performance computing resources, and they have stringent requirements for the completeness of model boundary conditions, making them completely unsuitable for real-time monitoring scenarios in portable devices. On the other hand, purely data-driven artificial intelligence methods, such as recurrent neural networks (RNNs) or long short-term memory (LSTMs), can directly process time-series data collected by sensors, but they typically operate as "black box" models, lacking adherence to the underlying physiological laws of respiratory breathing. This leads to the models being prone to producing results that deviate from physical reality due to error accumulation when making long-term predictions, especially when dealing with sparse and noisy data, making it difficult to guarantee the stability and reliability of their predictions.

[0005] Therefore, a significant technical challenge exists in existing technologies: how to effectively combine sparse physiological data collected by portable devices with complex physical models of the respiratory system to develop a monitoring method that ensures both physical consistency and high computational efficiency. Currently, there is a lack of a technical solution that can infer the spatiotemporal physical field of lung function from limited portable sensor data in real time and accurately, and achieve stable long-term prediction. Summary of the Invention

[0006] This invention provides a portable lung function monitoring method and device based on artificial intelligence, which solves the problems of single monitoring dimensions, unstable prediction and high computing cost in the existing technology.

[0007] In a first aspect, embodiments of the present invention provide a portable lung function monitoring method based on artificial intelligence, the method comprising: Acquire sparse time-series data of at least one physiological parameter collected by a portable sensing device; Based on sparse time series data, a first neural network model is used to solve the inverse problem of a pre-defined respiratory system physiological model whose governing equation is a partial differential equation, and a set of model parameters characterizing the dynamic characteristics of an individual respiratory system are identified in real time. Based on the real-time identified model parameters and the current physiological state, a pre-trained second neural network model for solving the control equations is used to perform forward prediction in an autoregressive manner, generating a rolling prediction sequence of complete spatiotemporal physical field quantities characterizing lung function status covering a future time domain; and During the forward prediction process, the physical residuals of the governing equations are calculated based on the rolling prediction sequence, and small-step optimization is performed on the rolling prediction sequence according to the physical residuals to perform adaptive correction.

[0008] In one alternative implementation, the step of identifying a set of model parameters characterizing the dynamic properties of an individual's respiratory system in real time includes: A physical information neural network PINN is constructed as the first neural network model, and the control equation of the respiratory system physiological model is added as a regularization term to its loss function; PINN was trained using sparse time series data to identify parameters that serve as unknown boundary conditions in the physiological model.

[0009] In one alternative implementation, the method further includes: In the step of performing small-step optimization for adaptive correction, the parameters of the identified unknown boundary conditions are used as a known input required to calculate the physical residual. When evaluating the residuals of the governing equations, the parameters are substituted into the boundary terms of the governing equations, thereby making the small-step optimization process subject to the physical constraints of the parameter identification results.

[0010] In one alternative implementation, the step of generating a rolling prediction sequence of complete spatiotemporal physical quantities characterizing lung function status covering a future time domain includes: At the current time step At this point, the second neural network model is used to predict the spatiotemporal physical field quantities. time derivative ; The predicted time derivative and the current time step status The input is fed into an ODE integrator of an ordinary differential equation; the ODE integrator calculates the next time step by performing one or more numerical integration steps. status In order to achieve forward progression in time sequence; in, For spacetime physical field quantities, For the current time step, For the next time step, Integer index of the time step, To step at the current time The time derivative of the predicted spatiotemporal physical field quantity.

[0011] In one alternative implementation, the second neural network model is a time-varying neural operator (TNO) network, which includes: An input branch network for processing time-varying parameters; A trunk network for processing spatiotemporal coordinates; An instantaneous state branch network for processing time series of current physiological states; and A decoder is configured to perform Hadamard product fusion on the outputs of the input branch network, the trunk network, and the transient state branch network, and map the fused representation to the output solution space.

[0012] In one alternative implementation, the spatiotemporal physical field is a multi-physics coupled field, which includes at least the gas velocity field, pressure distribution field, and stress-strain field of the lung tissue within the respiratory tract. Furthermore, the second neural network model is configured to instantaneously reconstruct the global distribution of the multi-physics coupled field from the input of previously collected single-point or sparse physiological parameters through a single forward inference.

[0013] In one alternative implementation, a continuous-time transfer learning strategy for the mobile source problem is employed to train or fine-tune the second neural network model online. Among these methods, non-stationary physiological events are modeled as moving pressure or flow sources in the respiratory physiological model to improve the model's robustness in predicting such dynamic events.

[0014] In one alternative implementation, the step of performing small-step optimization on the rolling prediction sequence based on the physical residuals is achieved by minimizing the physical residuals as defined by the following loss function:

[0015] The loss function is in the residual form of the Navier-Stokes equation. This is the gas velocity vector; For pressure; The density of the gas; Kinematic viscosity; This is a physical activity. For time, For gradient operators; For the Laplace operator; This indicates computation in the space-time domain. Norm; and, perform small-step optimization as one or more steps of gradient descent on the weights of the second neural network model or its output field.

[0016] In one alternative implementation, the respiratory system physiological model includes multiple subdomains composed of different biological tissues; The first neural network model and / or the second neural network model are subjected to interface continuity constraints during training or in structure. When crossing the interfaces of different biological tissues, the constraints force physical quantities to meet the continuity or preset transition conditions.

[0017] Secondly, embodiments of the present invention provide a portable lung function monitoring device based on artificial intelligence, the device comprising: The data acquisition module is used to acquire sparse time-series data of at least one physiological parameter collected by a portable sensing device. The parameter identification module is used to solve the inverse problem of a pre-defined respiratory system physiological model whose governing equation is a partial differential equation based on sparse time series data using a first neural network model, and to identify a set of model parameters that characterize the dynamic characteristics of an individual respiratory system in real time. The forward prediction module, based on real-time identified model parameters and the current physiological state, utilizes a pre-trained second neural network model for solving the control equations to perform forward prediction in an autoregressive manner, generating a rolling prediction sequence of complete spatiotemporal physical field quantities characterizing lung function status covering a future time domain; and The online correction module is used to calculate the physical residuals of the governing equations based on the rolling prediction sequence during the forward prediction process, and to perform small-step optimization on the rolling prediction sequence based on the physical residuals for adaptive correction.

[0018] A third aspect of this invention provides an electronic device, which includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by at least one processor, such that the at least one processor can perform the method proposed in the first aspect of the present invention.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention.

[0020] Beneficial effects: By acquiring sparse physiological parameters in real time from portable sensing devices and using a first neural network model to solve inverse problems to identify individualized respiratory system model parameters, this method combines limited, easily accessible home monitoring data with complex biophysical models, achieving real-time calibration of individual physiological characteristics. This overcomes the limitations of traditional clinical testing, which cannot continuously track data, and the fact that existing portable devices only provide macroscopic indicators. A closed-loop correction mechanism under physical information constraints is introduced. During the forward prediction process, the physical residuals of the control equations are continuously calculated and small-step optimization is performed to ensure that the generated spatiotemporal physical field prediction sequence always follows the basic laws of respiratory dynamics. This mechanism effectively suppresses model divergence and physical distortion caused by error accumulation in long-term predictions of purely data-driven models, significantly improving the stability and reliability of prediction results. Based on this high-fidelity prediction sequence, this method can reconstruct high-dimensional physiological information such as the gas flow field and pressure distribution within the respiratory tract from sparse inputs, achieving a leap from "point" to "field" monitoring capabilities. Meanwhile, this method utilizes the efficient reasoning capabilities of neural networks, avoiding the high computational costs required by traditional numerical simulations (such as computational fluid dynamics), enabling it to be deployed in portable devices and providing a novel technical approach for achieving low-cost, high-precision personalized dynamic assessment and early warning of lung function. Attached Figure Description

[0021] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the steps of a portable lung function monitoring method based on artificial intelligence, provided in an embodiment of the present invention. Figure 3 This is a functional unit diagram of a portable lung function monitoring device based on artificial intelligence provided in an embodiment of the present invention. Detailed Implementation

[0022] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is merely some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of aiding understanding the invention, but does not constitute a limitation on the invention. The solution of the present invention will be further described below in conjunction with the accompanying drawings.

[0023] Reference Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0024] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0025] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0026] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a data storage module, a network communication module, a user interface module, and electronic programs.

[0027] In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the portable lung function monitoring device based on artificial intelligence stored in the memory 1005 through the processor 1001 and executes the portable lung function monitoring method based on artificial intelligence provided in the embodiment of the present invention.

[0028] The core of this invention lies in constructing a two-stage, closed-loop monitoring framework consisting of "inverse problem parameter identification" and "forward physical field prediction and correction." In the first stage, a Physics-Informed Neural Network (PINN) is used to inversely solve for key physiological model parameters characterizing the unique properties of an individual's respiratory system from sparse, noisy physiological data collected from portable devices, thus completing personalized modeling. In the second stage, the identified parameters are input into a pre-trained advanced neural operator network designed specifically for solving time-varying partial differential equations (PDEs). Using a "predicting change rather than state" approach, combined with an Ordinary Differential Equation (ODE) integrator, it efficiently and stably predicts the full-spatiotemporal physical field of lung function over a future period. Crucially, this forward prediction process incorporates an online adaptive correction mechanism. This mechanism continuously calculates the physical residual between the predicted field and the underlying physical governing equations (such as the Navier-Stokes equations), and uses this residual to drive small-step gradient optimization of the predicted sequence. This real-time correction of accumulated errors ensures the physical consistency and accuracy of long-term predictions. This framework deeply integrates data-driven approaches with physical mechanisms, addressing the challenges of existing technologies in terms of real-time performance, fidelity, and personalization.

[0029] Reference Figure 2 This embodiment provides a portable lung function monitoring method based on artificial intelligence. The specific process of the method includes the following steps S101 to S104: S101: Acquire sparse time-series data of at least one physiological parameter collected by a portable sensing device.

[0030] This step aims to collect physiological signals reflecting a user's respiratory activity in their daily life using a non-invasive or minimally invasive portable device.

[0031] S1011: Deployment of portable sensing devices. The portable sensing device may be a combination of one or more devices, such as: a wearable respiratory flow meter, similar to a miniature spirometer, which can be worn over the mouth and nose or integrated into a mask to measure the rate and volume of respiratory airflow; an acoustic sensor, a high-sensitivity microphone array, attached to a specific location on the chest or back (auscultation area), to collect acoustic signals such as breath sounds, cough sounds, and wheezing sounds; a bioimpedance sensor, which monitors changes in thoracic impedance with respiratory motion by applying a weak current to the chest wall and measuring voltage changes, thereby estimating respiratory volume and frequency; and motion and posture sensors, such as accelerometers and gyroscopes, to record the user's physical activity and posture (standing, sitting, lying down) as contextual information for analyzing respiratory signals.

[0032] S1012: Acquisition of physiological parameters. The at least one physiological parameter is a physical quantity that can indirectly or directly reflect the dynamic process of the respiratory system, such as: respiratory airflow rate, directly measured by a respiratory flow meter; chest wall vibration signal, respiratory sound data acquired by an acoustic sensor; pleural volume change, calculated from measurements taken by a bioimpedance sensor; and respiratory rate, extracted from any of the above signals using signal processing techniques.

[0033] S1013: Data Sparsity and Time Series Characteristics. Because portable devices need to consider power consumption, wearing comfort, and data transmission bandwidth, their data acquisition is usually not continuous, but rather sampled at specific time points or at a lower frequency, forming "sparse time series data." For example, a 5-second breath sound might be collected every minute, or the average flow rate of a respiratory cycle might be recorded every 5 seconds. These timestamped data points form the basis for subsequent model analysis.

[0034] S102: Based on sparse time series data, a first neural network model is used to solve the inverse problem of a pre-defined respiratory system physiological model whose governing equation is a partial differential equation, and a set of model parameters characterizing the dynamic characteristics of an individual respiratory system are identified in real time.

[0035] This step is the core of personalized monitoring, and its goal is to deduce model parameters that describe an individual's internal physiological characteristics from limited external observation data.

[0036] S1021: Presupposition of the physiological model of the respiratory system. First, a mathematical model needs to be established to describe the key physical phenomena in the respiratory process. This model usually consists of a set of partial differential equations (PDEs).

[0037] In one alternative implementation, the model can couple fluid dynamics and solid mechanics to simulate gas flow within the airway and deformation of lung tissue. For example, the incompressible Navier-Stokes equations can be used to describe gas flow within the airway; simultaneously, linear elastic or hyperelastic constitutive models can be used to describe the mechanical behavior of the lung parenchyma and chest wall tissues. These equations together constitute a physiological model of the respiratory system.

[0038] S1022: Definition of Model Parameters. The model parameters are the undetermined coefficients or boundary condition functions in the aforementioned PDEs. They vary among individuals, reflecting individualized physiological characteristics. For example, parameters related to airway resistance, such as equivalent tube diameter and wall roughness, affect the viscous terms or boundary conditions in the Navier-Stokes equations. Parameters related to lung compliance, such as the elastic modulus and Poisson's ratio of lung tissue, are material constants in the governing equations of solid mechanics. Boundary condition parameters, for example, in simulating forced exhalation, the functional form of intrathoracic pressure changing with time, or its key parameters, are unknown and can be used as boundary conditions to be identified. This invention can equate intrathoracic driving pressure to an undetermined boundary parameter, which can be determined by solving an inverse problem.

[0039] S1023: Construction of the First Neural Network Model. To solve the aforementioned inverse problem from sparse data, this invention employs a Physics-Informed Neural Network (PINN) as the first neural network model in one optional implementation. PINN is a deep learning framework unique in that it directly encodes physical laws (i.e., PDEs) into the loss function of the neural network.

[0040] Construct a fully connected feedforward neural network whose input is spatiotemporal coordinates (e.g., time). and spatial location The output is the physical field quantity (e.g., flow velocity) at that point in spacetime. and pressure ).

[0041] PINN's loss function is a composite function, including at least data loss. and physical residual loss Two parts: Data loss :Measure the physical quantities predicted by the network Sparse time series data collected in S101 The difference between them is usually defined using the mean squared error (MSE):

[0042] in, It is the total number of sparse data points. It is the first The spatiotemporal coordinates corresponding to each data point.

[0043] Physical residual loss : Measure network output To what extent do the pre-defined control equations (PDEs) of the respiratory physiological model satisfy? This is determined by sampling within the computational domain. Collocation points This is achieved by calculating the squared L2 norm of the PDE residuals at these points. If the governing equation is the Navier-Stokes equation, then the loss term can be expressed as:

[0044] in, This is the gas velocity vector field predicted by the network. For the pressure field predicted by the network, The density of the gas (usually a known constant or a known function of temperature and pressure). The kinematic viscosity (usually a known constant or a known function), For physical exertion (such as gravity, which is often negligible), For time, For gradient operators, For the Laplace operator, express The square of the norm. All derivative terms (such as...) , , , All calculations are performed precisely based on the network output using Automatic Differentiation (ADE) technology.

[0045] Total loss function :

[0046] in, and These are hyperparameters used to balance the weights of the data fitting term and the physical constraint term in the total loss, respectively.

[0047] S1024: Inverse Problem Solving and Parameter Identification. This involves identifying the model parameters to be identified (e.g., unknown boundary condition parameters). The weights (W) and biases (b) are also used as trainable variables in the neural network and are optimized together with the network's weights (W) and biases (b). The training process involves minimizing the total loss function using a gradient descent algorithm (such as the Adam optimizer). When training converges, the network not only learns an approximate solution for the physical field that satisfies the laws of physics and fits the observed data, but also obtains the optimal unknown model parameters. This process is called real-time parameter identification.

[0048] In one alternative implementation, the step of identifying a set of model parameters characterizing the dynamic properties of an individual's respiratory system in real time includes: A physical information neural network PINN is constructed as the first neural network model, and the control equation of the respiratory system physiological model is added as a regularization term to its loss function; PINN was trained using sparse time series data to identify parameters that serve as unknown boundary conditions in the physiological model.

[0049] In one alternative implementation, considering that the respiratory system is a multi-layered, multi-material composite structure composed of various tissues with different physical properties, such as the trachea, bronchi, and alveoli, physical properties (such as elastic modulus and density) can change abruptly at different tissue interfaces. To accurately model this discontinuity, an adaptive interface PINN (AdaI-PINNs) framework can be introduced. This method uses domain decomposition to assign an independent neural network to each subdomain separated by interfaces (such as the airway lumen, tracheal wall, and lung parenchyma). These networks share the same weights and biases but use adaptive activation functions with different (trainable) slopes. Simultaneously, interface continuity constraints (e.g., displacement continuity, stress balance, or heat flow continuity) are added to the loss function, forcing the network to learn the correct transition behavior of physical quantities when crossing interfaces, thereby significantly improving the accuracy and convergence speed of modeling complex laminated structures.

[0050] S103: Based on the real-time identified model parameters and the current physiological state, a pre-trained second neural network model for solving the control equations is used to perform forward prediction in an autoregressive manner, generating a rolling prediction sequence of complete spatiotemporal physical field quantities that characterize the lung function state, covering a future time domain.

[0051] This step aims to utilize the personalized model obtained in S102 to perform efficient and long-term future state predictions, achieving a leap from single-point monitoring to the entire field, and from the current state to future trends.

[0052] S1031: Selection and Pre-training of the Second Neural Network Model. To achieve fast online inference, the second neural network model needs to learn the mapping from the parameter space to the solution function space (i.e., operator learning). In one optional implementation, the second neural network model is a temporal neural operator (TNO) network. The TNO network includes: an input branch network for processing time-varying parameters; a trunk network for processing spatiotemporal coordinates; an instantaneous state branch network for processing the current physiological state time series; and a decoder configured to perform a Hadamard product fusion on the outputs of the input branch network, trunk network, and instantaneous state branch network, and map the fused representation to the output solution space. Specifically, the input branch network receives and encodes functional inputs that do not change or change slowly over time, such as the spatial distribution of individual model parameters identified in S102. (If the parameters are spatially variable), or global scalar parameters. Temporal Branch (t-branch): Receives and encodes a time-varying functional input, i.e., the current moment or a period of time in the past. physiological state field Trunk Network: Receives the spatiotemporal coordinates of the query point. It outputs coordinate-dependent basis functions. Fusion & Decoder: Combines the encoded outputs of the three networks (usually within the same latent space dimension). (Through the Hadamard product) The fused representations are then mapped to the output solution space by a decoder (such as an MLP) to predict the future. The state field at each time step .

[0053] The pre-training process of TNO is as follows: A large-scale simulation dataset is generated offline. This dataset contains tens of thousands of "parameter-initial state-future state sequences" obtained by solving the respiratory system physiological model using a high-precision numerical solver (such as the finite element method) within a wide range of parameters (covering physiological parameters of normal people and patients with various respiratory diseases). Then, the TNO network is trained in a supervised learning manner using this dataset, enabling it to quickly predict the future physical field evolution sequence from any given model parameters and initial state.

[0054] S1032: Implementation of forward prediction. During online monitoring, the individual model parameters identified in real time by S102 and the current physiological state field reconstructed from sensor data are input into the pre-trained TNO model.

[0055] In one alternative implementation, the step of generating a rolling prediction sequence of complete spatiotemporal physical quantities characterizing lung function status covering a future time domain includes: At the current time step At this point, the second neural network model is used to predict the spatiotemporal physical field quantities. time derivative ; The predicted time derivative and the current time step status Input to an ODE integrator for an ordinary differential equation; The ODE integrator calculates the next time step by performing one or more steps of numerical integration. status In order to achieve forward progression in time sequence; in, For spacetime physical field quantities, For the current time step, For the next time step, Integer index of the time step, To step at the current time The time derivative of the predicted spatiotemporal physical field quantity.

[0056] Specifically, this method adopts a framework of "predicting changes, rather than states," and the specific steps are as follows: S10321: Prediction time derivative: at the current time step At this point, the TNO model is used to predict spatiotemporal physical field quantities. time derivative The advantage of doing this is that the model only needs to learn the changes in the state, rather than the complete next state, which is usually a simpler and more stable learning objective.

[0057] S10322: Ordinary Differential Equation (ODE) Integration: The predicted time derivative is input into a standard numerical integrator, such as the fourth-order Runge-Kutta method (RK4). S10323: State Advancement: The ODE integrator performs one or more steps of numerical integration to calculate the next time step. status .

[0058] S10323: State Progression: The ODE integrator performs one or more steps of numerical integration to calculate the next time step. status Taking the fourth-order Runge-Kutta method (RK4) as an example, its update rules are as follows:

[0059]

[0060]

[0061]

[0062]

[0063] in, It is the current time step The state field, Representing the second neural network model (such as TNO) in time State Predicted time derivative , This is the time step used for numerical integration (it needs to be preset or adaptively adjusted). This is the intermediate slope estimate required for calculation by the RK4 method. In this way, combining the powerful fitting ability of neural networks to spatial derivatives with the mature stability of the ODE integrator in time integration significantly reduces error accumulation in long-sequence autoregressive predictions.

[0064] S1033: Generation of the rolling prediction sequence. By repeatedly executing the "prediction derivative-integral propagation" loop in S1032, a rolling prediction sequence consisting of a complete spatiotemporal physical field across multiple time steps, covering a future period (e.g., the next 10 seconds), can be generated. This sequence not only includes the values ​​at the sensor locations but also the full-field distribution of physical quantities such as flow velocity, pressure, and stress throughout the three-dimensional space of the lung.

[0065] In one alternative implementation, the spatiotemporal physical field is a multi-physics coupled field, including at least the gas velocity field, pressure distribution field, and stress-strain field of the lung tissue within the respiratory tract; and the second neural network model is configured to instantaneously reconstruct the global distribution of the multi-physics coupled field from previously collected single-point or sparse physiological parameter inputs through a single forward inference.

[0066] Specifically, a second neural network model (such as TNO or its variant ResUNet-DeepONet) can be configured to simultaneously predict multiphysics. That is, the network output includes not only the flow field but also the temperature field, stress field, and so on. In the context of lung function monitoring, this means the model can infer the gas velocity field, pressure distribution field, and stress-strain field of lung tissue all at once, achieving a performance leap from single-physics to multiphysics.

[0067] In one alternative implementation, a continuous-time transfer learning strategy for the moving source problem is employed to train or fine-tune the second neural network model online.

[0068] Specifically, in one optional implementation, to handle intense, non-stationary physiological events such as coughing and forceful exhalation, these events can be abstracted into moving pressure or flow sources in the model. A continuous-time segmentation or transfer learning strategy specifically designed for moving source problems is then employed to train or fine-tune the second neural network model online. This strategy segments the entire time domain and performs knowledge transfer between segments, enabling the model to robustly handle rapid changes in source location or intensity over time, thereby improving the speed and robustness of the monitoring method under real physiological events.

[0069] S104: During the forward prediction process, the physical residuals of the control equations are calculated based on the rolling prediction sequence, and small-step optimization is performed on the rolling prediction sequence according to the physical residuals to perform adaptive correction.

[0070] This step is a crucial closed-loop link to ensure long-term prediction stability and physical consistency. It introduces a PINN-like physical constraint during the online prediction phase to "correct" the output of the neural operator in real time.

[0071] S1041: Calculation of physical residuals. Rolling prediction sequence is generated in S103. During or after the process, for each predicted time step State field This is then substituted into the control equation (PDE) of the respiratory system physiological model. Because... This is an approximate solution from a neural network. Substituting it into the equation usually results in a result that is not strictly equal to zero; the result is the physical residual. .

[0072] In one alternative implementation, the physical residual is passed through a loss function. To quantify it. Taking the Navier-Stokes equation as an example, the loss function can be defined in the spatiotemporal domain of the current prediction step. Or only at a point in time spatial domain The PDE residuals calculated above Norm square:

[0073] in, and This represents the velocity and pressure fields in the current rolling prediction sequence. The physical meaning of each symbol has been defined in S1023. The derivatives here are also calculated using automatic differentiation.

[0074] S1042: Execution of small-step optimization. Calculate the physical residual loss. Then, it is treated as an optimization objective. The loss function is calculated automatically with respect to the amount that needs to be corrected (e.g., the currently predicted physics field). The trainable parameters of the second neural network model that generates the prediction or the prediction. The gradient of the predicted field is then calculated. Then, one or more gradient descent steps are performed to update these quantities to reduce the physical residuals. For example, if the predicted field itself is corrected:

[0075] in, It is the corrected physical field. It's a small learning rate. The corrected field size. This will be used for predictions at subsequent time steps. If online fine-tuning of model parameters is required:

[0076] in, These are the fine-tuned model parameters used for predictions at subsequent time steps. Alternatively, a combination of both can be used. This update step is called "small step optimization." It does not perform a complete retraining, but rather fine-tunes the prediction results or model state online and in a lightweight manner to make it more consistent with the laws of physics.

[0077] S1043: Closed-loop process for adaptive correction. The corrected state... As a more accurate current state, it is used for the next step ( The prediction process is thus completed. This forms a closed loop of "prediction-residual calculation-optimization correction-re-prediction". This adaptive correction mechanism effectively suppresses and eliminates the inevitable error accumulation in long-term autoregressive predictions, ensuring that the results remain physically reasonable and consistent even after hundreds of time steps of prediction.

[0078] In one optional implementation, the method of the present invention further includes: In the step of performing small-step optimization for adaptive correction, the parameters of the identified unknown boundary conditions are used as a known input required to calculate the physical residual. When evaluating the residuals of the governing equations, the parameters are substituted into the boundary terms of the governing equations, thereby making the small-step optimization process subject to the physical constraints of the parameter identification results.

[0079] Specifically, in one optional implementation, the unknown boundary condition parameters identified in S102 are used as prior information or dynamic constraints when calculating the physical residuals in S1041. For example, when calculating the residuals in the boundary region, the boundary terms of the governing equations are substituted with the identified parameter values. This ensures that the online correction process not only follows universal physical laws but also personalized boundary conditions learned from the user-specific data, thus tightly coupling parameter identification with forward prediction correction and further enhancing the reliability of the overall monitoring process with limited sensor data.

[0080] In summary, this embodiment, through the organic combination of S101 to S104, constructs a complete portable lung function monitoring method that encompasses sparse data acquisition, personalized modeling, and physical consistency full-field prediction. This method not only addresses the pain points of existing technologies in terms of real-time performance, fidelity, and personalization, but also demonstrates significant technical advantages and broad application prospects through innovative model architecture and algorithmic processes (such as solving the PINN inverse problem, learning the TNO operator, predicting changes rather than states, and online physical residual correction).

[0081] This invention also provides a portable lung function monitoring device based on artificial intelligence, referring to... Figure 3 The diagram shows a functional unit diagram of a portable lung function monitoring device 300 based on artificial intelligence according to the present invention. The device may include the following components: The data acquisition module 301 is used to acquire sparse time series data of at least one physiological parameter collected by a portable sensing device. The parameter identification module 302 is used to solve the inverse problem of a preset respiratory system physiological model whose governing equation is a partial differential equation based on sparse time series data using a first neural network model, and to identify a set of model parameters that characterize the dynamic characteristics of an individual respiratory system in real time. Forward prediction module 303 is used to generate a rolling prediction sequence of complete spatiotemporal physical field quantities characterizing lung function status, based on real-time identified model parameters and a current physiological state, using a pre-trained second neural network model for solving the control equations in an autoregressive manner; and based on these parameters, the current physiological state is determined. The online correction module 304 is used to calculate the physical residuals of the control equations based on the rolling prediction sequence during the forward prediction process, and to perform small-step optimization on the rolling prediction sequence according to the physical residuals for adaptive correction.

[0082] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the portable lung function monitoring method based on artificial intelligence of the present invention.

[0083] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0084] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0085] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the portable lung function monitoring method based on artificial intelligence according to embodiments of the present invention.

[0086] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable hardware devices (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0090] Finally, 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. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0091] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An artificial intelligence-based portable lung function monitoring method, characterized by, The method includes: Acquire sparse time-series data of at least one physiological parameter collected by a portable sensing device; Based on the sparse time series data, a first neural network model is used to solve the inverse problem of a preset respiratory system physiological model whose governing equation is a partial differential equation, and a set of model parameters characterizing the dynamic characteristics of an individual respiratory system are identified in real time. Based on the real-time identified model parameters and the current physiological state, a pre-trained second neural network model for solving the control equations is used to perform forward prediction in an autoregressive manner, generating a rolling prediction sequence of complete spatiotemporal physical field quantities characterizing lung function status covering a future time domain; and In the forward prediction process, the physical residuals of the control equations are calculated based on the rolling prediction sequence, and small-step optimization is performed on the rolling prediction sequence according to the physical residuals to perform adaptive correction.

2. The method of claim 1, wherein, The step of identifying a set of model parameters characterizing the dynamic properties of an individual's respiratory system in real time includes: A physical information neural network PINN is constructed as the first neural network model, and the control equation of the respiratory system physiological model is added as a regularization term to its loss function; The PINN is trained using the sparse time series data to identify the parameters that serve as unknown boundary conditions in the physiological model.

3. The method of claim 2, wherein, The method further includes: In the step of performing the small-step optimization for adaptive correction, the parameters of the identified unknown boundary conditions are used as a known input required to calculate the physical residual. When evaluating the residuals of the governing equations, the parameters are substituted into the boundary terms of the governing equations, thereby subjecting the small-step optimization process to the physical constraints of the parameter identification results.

4. The method of claim 3, wherein, The step of generating a rolling prediction sequence of complete spatiotemporal physical quantities characterizing lung function status covering a future time domain includes: At the current time step , a time derivative of the spatio-temporal physical field quantity is predicted using the second neural network model. The predicted time derivative and the current time step status The input is fed into an ordinary differential equation (ODE) integrator; the ODE integrator calculates the next time step by performing one or more steps of numerical integration. status In order to achieve forward progression in time sequence; wherein, is the spatiotemporal physical field quantity, is the current time step, is the next time step, is the integer index of the time step, is the time derivative of the spatiotemporal physical field quantity predicted at the current time step at the next time step.

5. The method according to claim 4, characterized in that, The second neural network model is a time-of-flight neural operator (TNO) network, which includes: An input branch network for processing time-varying parameters; A trunk network for processing spatiotemporal coordinates; An instantaneous state branch network for processing time series of current physiological states; and A decoder is configured to perform Hadamard product fusion on the outputs of the input branch network, the trunk network, and the instantaneous state branch network, and map the fused representation to the output solution space.

6. The method according to claim 5, characterized in that, The spatiotemporal physical field is a multi-physics coupled field, which includes at least the gas velocity field, pressure distribution field and stress-strain field of lung tissue in the respiratory tract. Furthermore, the second neural network model is configured to instantaneously reconstruct the global distribution of the multi-physics coupled field from the collected single-point or sparse physiological parameter inputs through a single forward inference.

7. The method according to claim 6, characterized in that, A continuous-time transfer learning strategy for the mobile source problem is adopted to train or fine-tune the second neural network model online. Non-stationary physiological events are modeled as pressure or flow sources moving within the respiratory system physiological model to improve the model's robustness in predicting such dynamic events.

8. The method according to claim 7, characterized in that, The step of performing small-step optimization on the rolling prediction sequence based on the physical residual is achieved by minimizing the physical residual defined by the loss function: The loss function is in the residual form of the Navier-Stokes equation. This is the gas velocity vector; For pressure; The density of the gas; Kinematic viscosity; This is a physical activity. For time, For gradient operators; For the Laplace operator; This indicates computation in the space-time domain. Norm; and, the small-step optimization is a one-step or multi-step gradient descent of the weights or output fields of the second neural network model.

9. The method according to claim 8, characterized in that, The respiratory system physiological model contains multiple subdomains composed of different biological tissues; The first neural network model and / or the second neural network model are subjected to interface continuity constraints during training or in structure. These constraints force physical quantities to meet continuity or preset transition conditions when crossing interfaces between different biological tissues.

10. A portable lung function monitoring device based on artificial intelligence, characterized in that, The device includes: The data acquisition module is used to acquire sparse time-series data of at least one physiological parameter collected by a portable sensing device. The parameter identification module is used to solve the inverse problem of a preset respiratory system physiological model whose governing equation is a partial differential equation based on the sparse time series data and a first neural network model, and to identify a set of model parameters that characterize the dynamic characteristics of an individual respiratory system in real time. The forward prediction module is used to generate a rolling prediction sequence of complete spatiotemporal physical field quantities characterizing lung function status, based on the real-time identified model parameters and the current physiological state, using a pre-trained second neural network model for solving the control equations in an autoregressive manner; and based on this prediction. An online correction module is used to calculate the physical residual of the control equation based on the rolling prediction sequence during the forward prediction process, and to perform small-step optimization on the rolling prediction sequence according to the physical residual to perform adaptive correction.

Images