Artificial intelligence-based crrt data all-around collection and optimization processing system

Abstract

This invention relates to the field of medical artificial intelligence and discloses a comprehensive data acquisition and optimization system for CRRT based on artificial intelligence. The system includes: an acquisition and calibration module for acquiring CRRT equipment parameters, physiological signals, and laboratory test data to form multi-source data, and performing time alignment and spatial calibration to obtain calibration data; a modeling and estimation module for constructing a spatiotemporal dynamic model of material transport and biochemical reactions during the CRRT treatment process based on stochastic partial differential equations, and estimating state variables using calibration data to generate high-dimensional state data; and a fusion and inference module for performing dimensionality reduction and adversarial denoising on the high-dimensional state data. Through multi-source acquisition and spatiotemporal calibration of CRRT equipment parameters, physiological signals, and laboratory data, a foundation for treatment data is constructed. Based on stochastic partial differential equations, spatiotemporal dynamic modeling of material transport and biochemical reactions is performed, enabling dynamic characterization of the treatment process.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to an AI-based system for comprehensive acquisition and optimization of CRRT data. Background Technology

[0002] Continuous renal replacement therapy (CRRT), a vital life support method in critical care medicine, continuously removes metabolic waste and maintains electrolyte balance through extracorporeal circulation. It is widely used in the treatment of critically ill patients with conditions such as acute kidney injury and sepsis. The CRRT treatment process involves dynamic adjustments to multiple parameters, including blood flow, ultrafiltration rate, and replacement fluid composition. It also requires real-time monitoring of complex changes in the patient's physiological indicators (such as ionized calcium concentration and acid-base balance), demanding extremely high precision and real-time control of treatment parameters.

[0003] Traditional CRRT optimization systems often rely on statistical correlation models (such as regression analysis and machine learning classification algorithms) to train historical treatment data, thereby establishing the correlation between treatment parameters and physiological indicators. However, these models can only reveal the correlation between variables and cannot analyze the causal mechanism. As a result, during parameter adjustment, it is impossible to distinguish between the true causal effect and the interference of confounding factors, leading to unexpected treatment side effects. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based CRRT data acquisition and optimization system, which solves the problem that traditional CRRT optimization systems cannot distinguish between true causal effects and confounding factors during parameter adjustment, leading to unexpected treatment side effects.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a comprehensive CRRT data acquisition and optimization processing system based on artificial intelligence, comprising:

[0006] Acquisition and calibration module: used to acquire CRRT equipment parameters, physiological signals and laboratory test data, form multi-source data, and perform time alignment and spatial calibration to obtain calibration data;

[0007] Modeling and estimation module: used to construct a spatiotemporal dynamic model of material transport and biochemical reactions in the CRRT treatment process based on stochastic partial differential equations, and to estimate state variables by combining calibration data to generate high-dimensional state data;

[0008] Fusion inference module: used to perform dimensionality reduction and adversarial denoising on high-dimensional state data, generate low-dimensional feature vectors, and construct a causal relationship map between treatment parameters and physiological indicators based on the low-dimensional feature vectors to calculate the intervention effect;

[0009] Control strategy module: used to build a model prediction control framework based on spatiotemporal dynamics model, generate basic treatment plans that meet clinical constraints by combining intervention effects, and adjust the parameters of the plan through deep reinforcement learning;

[0010] Command generation module: used to integrate basic treatment plans and personalized adjustment parameters to generate device control commands, execute real-time control at the edge through an edge-cloud collaborative architecture, and perform iterative training of spatiotemporal dynamics models and deep reinforcement learning strategies in the cloud.

[0011] By adopting the above technical solutions, multi-source acquisition and spatiotemporal calibration of CRRT equipment parameters, physiological signals, and laboratory data are performed to construct a treatment data foundation. Based on stochastic partial differential equations, spatiotemporal dynamic modeling of material transport and biochemical reactions is achieved, enabling dynamic characterization of the treatment process. Through dimensionality reduction, noise reduction, and causal relationship mapping, the intervention effect of treatment parameters on physiological indicators is quantified. Combining model predictive control and deep reinforcement learning, treatment plans that are both clinically safe and individually adaptable are generated. Real-time control and model iteration are achieved through an edge-cloud collaborative architecture. Thus, the true action path of treatment parameters and physiological indicators is clarified through causal mapping, avoiding decision-making biases caused by confounding factors. With the help of stochastic partial differential equation modeling and deep reinforcement learning, the adaptability of the model to individual differences and the real-time performance of parameter adjustment are simultaneously improved. This achieves scientific analysis and precise intervention of causal relationships in the CRRT treatment process, solving the problem that traditional CRRT optimization systems cannot distinguish between true causal effects and confounding factors during parameter adjustment, leading to unexpected treatment side effects.

[0012] Preferably, the acquisition and calibration module includes an acquisition unit, a communication unit, and a calibration unit. The acquisition unit is used to acquire CRRT equipment parameters, physiological signals, and laboratory test data through a three-level sensor network to form multi-source data. The communication unit is used to perform real-time interaction of various data through HL7 FHIR interface protocol conversion. The calibration unit is used to perform time alignment and spatial calibration of the multi-source data by using hardware synchronous clock-triggered sampling combined with fractional calculus interpolation algorithm to obtain calibration data.

[0013] Preferably, the time alignment employs fractional-order calculus interpolation, the formula of which is: ,in, To optimize parameters and the range of values ​​is , For gamma function, For the interpolated signal at the target time point, The first derivative of the original signal. For signal acquisition time window, The target time alignment point.

[0014] Preferably, the modeling and estimation module includes a discretization unit, a parameter identification unit, and a state estimation unit. The discretization unit is used to transform the material transport and biochemical reactions in the CRRT treatment process into a weak form of a three-dimensional stochastic partial differential equation. The parameter identification unit is used to optimize the parameters of the three-dimensional stochastic partial differential equation through variational Bayesian inference. The state estimation unit is used to solve the three-dimensional stochastic partial differential equation based on calibration data, estimate state variables, and generate high-dimensional state data.

[0015] Preferably, the stochastic partial differential equation is: , where the state vector Includes ionized calcium concentration field, citrate concentration field, and blood flow velocity field, diffusion matrix. It is a diagonal matrix, with elements representing the diffusion coefficients of each substance and reaction terms. Describe the kinetics of biochemical reactions. The noise intensity matrix is... For Wiener process;

[0016] The optimization involves optimizing the parameters of a three-dimensional stochastic partial differential equation by maximizing the lower bound of evidence. The expression for maximizing the lower bound of evidence is as follows: ,in, It is a variational distribution. As a prior distribution, For observation data, Variational distribution The expected operator below, The divergence is Kullback-Leibler.

[0017] Preferably, the fusion inference module includes a dimensionality reduction unit, a denoising unit, and a causal modeling unit. The dimensionality reduction unit is used to perform manifold dimensionality reduction on high-dimensional state data using a local linear embedding algorithm. The denoising unit is used to denoise the dimensionality-reduced high-dimensional state data based on a conditional generative adversarial network to generate low-dimensional feature vectors. The causal modeling unit is used to construct a causal relationship map between treatment parameters and physiological indicators based on the low-dimensional feature vectors and calculate the intervention effect using Do-Calculus.

[0018] Preferably, the locally linear embedding algorithm reduces dimensionality by solving the following optimization problem: ,in, For low-dimensional manifold eigenvectors, To reconstruct the weight matrix locally, minimize Sure, The original high-dimensional data;

[0019] The formula for calculating the intervention effect is as follows: ,in, For treatment parameters, As a physiological indicator, for The set of parent nodes, for Except External variables, For intervention hour The probability distribution.

[0020] Preferably, the control strategy module includes a basic control unit and an intelligent optimization unit. The basic control unit is used to construct a model predictive control framework based on a spatiotemporal dynamics model, and generate a basic treatment plan that meets clinical constraints by combining the intervention effect. The clinical constraints include blood flow range, ionized calcium concentration threshold, and ultrafiltration rate upper limit. The cost function of the model predictive control framework is... ,in, To predict the time domain, To control the time domain, To predict the state vector, For reference trajectory, To control variables, and These are the weight matrices for the state and control variables, respectively. The intelligent optimization unit is used to personalize the parameters of the basic treatment plan through deep reinforcement learning. The reward function of the deep reinforcement learning is... ,in, To control the smoothing coefficient, To avoid the reward coefficient for complications, For indicator functions, Control variables for the basic treatment plan This is the reference state vector.

[0021] Preferably, the instruction generation module includes a weight fusion unit, a security verification unit, and a collaborative architecture unit. The weight fusion unit is used to fuse the weights of the basic treatment plan and personalized adjustment parameters dynamically adjusted based on the treatment stage, and output parameter adjustment instructions. The security verification unit is used to filter abnormal instructions based on causal constraints and clinical thresholds before generating device control instructions. The collaborative architecture unit is used to execute real-time control at the edge according to the device control instructions through an edge-cloud collaborative architecture, and to perform iterative training of spatiotemporal dynamics models and deep reinforcement learning strategies in the cloud through federated learning.

[0022] The AI-based CRRT data comprehensive acquisition and optimization processing method, applied to the aforementioned AI-based CRRT data comprehensive acquisition and optimization processing system, includes the following steps:

[0023] Data Acquisition and Calibration: Collect CRRT equipment parameters, physiological signals, and laboratory test data to form multi-source data, and perform time alignment and spatial calibration to obtain calibration data;

[0024] Modeling and estimation: Based on stochastic partial differential equations, a spatiotemporal dynamic model of material transport and biochemical reactions in the CRRT treatment process is constructed, and state variables are estimated by combining calibration data to generate high-dimensional state data;

[0025] Fusion reasoning: Dimensionality reduction and adversarial denoising are performed on high-dimensional state data to generate low-dimensional feature vectors, and a causal relationship map between treatment parameters and physiological indicators is constructed based on the low-dimensional feature vectors to calculate the intervention effect;

[0026] Control strategy: A model prediction control framework is constructed based on a spatiotemporal dynamics model, and a basic treatment plan that meets clinical constraints is generated by combining the intervention effect. The parameters are then adjusted for individualization through deep reinforcement learning.

[0027] Command generation: By integrating basic treatment plans and personalized adjustment parameters, device control commands are generated. Real-time control is executed at the edge through an edge-cloud collaborative architecture, and iterative training of spatiotemporal dynamics models and deep reinforcement learning strategies is performed in the cloud.

[0028] This invention provides an artificial intelligence-based system for comprehensive CRRT data acquisition and optimization processing. It offers the following advantages:

[0029] 1. This invention constructs a treatment data foundation by acquiring and spatiotemporally calibrating CRRT equipment parameters, physiological signals, and laboratory data from multiple sources. Based on stochastic partial differential equations, it models the spatiotemporal dynamics of material transport and biochemical reactions to achieve dynamic characterization of the treatment process. Through dimensionality reduction, noise reduction, and causal relationship mapping, it quantifies the intervention effect of treatment parameters on physiological indicators. Combining model predictive control and deep reinforcement learning, it generates treatment plans that are both clinically safe and individually adaptable. Through an edge-cloud collaborative architecture, it achieves real-time control and model iteration, thereby realizing the scientific analysis and precise intervention of causal relationships in the CRRT treatment process.

[0030] 2. This invention introduces stochastic partial differential equations to mathematically describe the transport and biochemical reaction processes of ion calcium diffusion, citrate complexation, and other substances in CRRT treatment. A numerical solution framework is constructed by discretizing the spatiotemporal domain using the finite element method, and variational Bayesian inference is used to optimize model parameters such as diffusion coefficient and reaction rate. This ensures that the equation parameters are accurately matched with the statistical characteristics of the patient's actual treatment data. By combining the stochastic finite element method with Euler-Markov time discretization technology, the dynamic evolution and distribution of state variables such as ion calcium concentration field and blood flow velocity field can be solved in real time, generating high-dimensional state data that characterizes the spatiotemporal characteristics of the treatment process. This better reflects the dynamic changes in substance distribution and physiological indicators in CRRT treatment, achieving refined mathematical modeling of the treatment process.

[0031] 3. This invention uses a locally linear embedding algorithm to perform manifold dimensionality reduction on high-dimensional state data, preserving key state patterns of CRRT treatment. Then, a conditional generative adversarial network is used to remove data noise, constructing a high-fidelity low-dimensional feature vector. Based on the low-dimensional features, a PC algorithm is used to construct a causal relationship map between treatment parameters and physiological indicators, clarifying the direct causal paths and indirect influence mechanisms between variables. Furthermore, the Do-Calculus algorithm quantifies the intervention effect of treatment parameter adjustments on physiological indicators, thereby distinguishing between true causal effects and confounding factors, avoiding parameter adjustment errors caused by spurious associations, providing a traceable physical mechanism basis for treatment decisions, and improving the safety and clinical interpretability of CRRT treatment.

[0032] 4. This invention utilizes model predictive control based on spatiotemporal dynamics and clinical constraints to generate a basic treatment plan that meets safety standards using an optimized cost function, ensuring that treatment parameters are adjusted within clinically acceptable ranges. Deep reinforcement learning uses real-time state bias, control smoothness, and complication avoidance as reward functions to personalize the parameters of the basic plan, learning the optimal treatment parameters corresponding to individual patient characteristics. Thus, model predictive control safeguards the bottom line of treatment safety while deep reinforcement learning uncovers the optimal solution under individual differences, achieving an organic unity of standardization and personalization in treatment plans, and improving the adaptability and efficacy of CRRT treatment for different patients.

[0033] 5. This invention dynamically adjusts the weights of the basic treatment plan and personalized parameters according to the treatment stage, and generates compliant device control commands after filtering abnormal commands by combining causal constraints and clinical thresholds. The edge device drives the CRRT device to execute control in real time through the HL7FHIR protocol interface, ensuring millisecond-level real-time performance of the treatment response. The cloud, based on federated learning technology, associates and trains the real-time treatment data collected at the edge device with spatiotemporal dynamics models and deep reinforcement learning strategies, iteratively optimizing model parameters and control strategies. This not only meets the stringent real-time requirements of CRRT treatment, but also continuously learns to adapt to changes in the patient's condition and new clinical scenarios. In the long term, it can continuously improve its ability to cope with complex treatment environments, promoting the development of CRRT treatment towards intelligence and adaptability. Attached Figure Description

[0034] Figure 1 This is a system architecture diagram of the AI-based CRRT data comprehensive acquisition and optimization processing system proposed in this invention;

[0035] Figure 2 This is a flowchart of the method for comprehensive acquisition and optimization processing of CRRT data based on artificial intelligence proposed in this invention. Detailed Implementation

[0036] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] Please see the appendix Figure 1 This invention provides an artificial intelligence-based CRRT data comprehensive acquisition and optimization processing system, including:

[0038] Acquisition and calibration module: Used to acquire CRRT equipment parameters, physiological signals, and laboratory test data to form multi-source data, and to perform time alignment and spatial calibration to obtain calibration data. The acquisition and calibration module includes an acquisition unit, a communication unit, and a calibration unit. The acquisition unit is used to acquire CRRT equipment parameters, physiological signals, and laboratory test data through a three-level sensor network to form multi-source data. The communication unit is used to perform real-time interaction of various data through HL7 FHIR interface protocol conversion. The calibration unit is used to perform time alignment and spatial calibration of multi-source data by using hardware synchronous clock-triggered sampling combined with fractional calculus interpolation algorithm to obtain calibration data.

[0039] Time alignment uses fractional calculus interpolation, the formula of which is: ,in, To optimize parameters and the range of values ​​is , For gamma function, For the interpolated signal at the target time point, The first derivative of the original signal. For signal acquisition time window, The target time alignment point.

[0040] Specifically, the acquisition and calibration module is used to construct a multi-source data acquisition, transmission, and spatiotemporal calibration link for CRRT, providing standardized input for the modeling and estimation module. The acquisition unit collects data hierarchically through a three-tiered sensor network: the first tier connects to the CRRT equipment, acquiring equipment operating parameters such as pressure, flow rate, and replacement fluid concentration; the second tier connects to physiological monitoring equipment, acquiring physiological signals such as ECG, blood oxygen, and pulse rate; and the third tier connects to the laboratory information system, obtaining test data such as blood gas analysis and ion concentration. This cross-device data acquisition covers all dimensions of CRRT treatment information, forming a multi-source heterogeneous data set.

[0041] The communication unit is based on the HL7 FHIR interface protocol, which converts the multi-source data output by the acquisition unit into a protocol, breaking down the barriers of proprietary protocols of equipment manufacturers and transforming heterogeneous data into a standardized medical data format. At the same time, relying on the real-time interaction mechanism, it initially aligns the timestamps of multi-source data, providing basic data input for the spatiotemporal calibration of the calibration unit and ensuring the real-time performance and compatibility of data transmission.

[0042] The calibration unit employs a strategy combining hardware synchronous clock-triggered sampling and fractional-order calculus interpolation: the hardware synchronous clock provides a unified time reference for multi-source data, triggering each sensor to sample synchronously at a preset time, thus resolving the asynchronous sampling issue between devices; for dynamically changing non-uniform sampling data during CRRT treatment, such as fluctuating blood flow velocity and instantaneously changing ion concentrations, time alignment is achieved using a fractional-order calculus interpolation algorithm, the formula of which is: ,in To optimize parameters, the following values ​​are selected: Adapting to the dynamic characteristics of CRRT data For gamma function, Fit the interpolated signal at the target time point. Characterizing the rate of change of the original signal, A signal acquisition time window is defined. Simultaneously, based on the built-in spatial calibration model, coordinate mapping is performed to account for spatial deployment differences in physiological signal sensors, such as surface electrodes and built-in sensors, transforming multi-source data into a unified spatial coordinate system. After calibration, the output calibration data has a unified sampling benchmark in the time dimension and achieves coordinate normalization in the spatial dimension. This lays a data accuracy foundation for the modeling and estimation module to construct a spatiotemporal dynamic model, effectively reducing model errors caused by data heterogeneity.

[0043] When the acquisition unit acquires the blood flow velocity signal from the CRRT device, this signal exhibits a non-uniform time interval distribution due to the device's sampling mechanism and clinical fluctuations. The calibration unit uses a hardware-synchronized clock to trigger the target time point. Using this as a benchmark, the first derivative of the original blood flow velocity signal is... As input, set optimization parameters Its adaptability to dynamic changes in blood flow velocity and signal acquisition time window It covers recent key fluctuation data, and is used in fractional calculus interpolation formulas. The target time is obtained by fitting the data through integral operations. Blood flow velocity interpolation signal at the location The interpolated signal aligns the time axis of non-uniformly sampled blood flow velocity data with other physiological signals such as ECG and blood oxygenation, providing standardized data for time synchronization for the modeling and estimation module to construct a spatiotemporal dynamic model, and ensuring the spatiotemporal consistency of the estimation of state variables such as ion concentration field and blood flow velocity field.

[0044] The modeling and estimation module is used to construct a spatiotemporal dynamic model of the material transport and biochemical reactions in the CRRT treatment process based on stochastic partial differential equations, and to estimate state variables by combining calibration data to generate high-dimensional state data. The modeling and estimation module includes a discretization unit, a parameter identification unit, and a state estimation unit. The discretization unit is used to transform the material transport and biochemical reactions in the CRRT treatment process into a weak form of three-dimensional stochastic partial differential equations. The parameter identification unit is used to optimize the parameters of the three-dimensional stochastic partial differential equations through variational Bayesian inference. The state estimation unit is used to solve the three-dimensional stochastic partial differential equations based on calibration data, estimate state variables, and generate high-dimensional state data.

[0045] The stochastic partial differential equation is , where the state vector Includes ionized calcium concentration field, citrate concentration field, and blood flow velocity field, diffusion matrix. It is a diagonal matrix, with elements representing the diffusion coefficients of each substance and reaction terms. Describe the kinetics of biochemical reactions. The noise intensity matrix is... For Wiener process;

[0046] The optimization involves maximizing the lower bound of evidence to optimize the parameters of the three-dimensional stochastic partial differential equation. The expression for maximizing the lower bound of evidence is as follows: ,in, It is a variational distribution. As a prior distribution, For observation data, Variational distribution The expected operator below, The divergence is Kullback-Leibler.

[0047] Specifically, the modeling and estimation module receives calibration data output from the acquisition and calibration module, which includes multi-source data such as time-aligned and spatially normalized CRRT device parameters and physiological signals. Through the collaborative work of the discretization unit, parameter identification unit, and state estimation unit, it constructs a spatiotemporal dynamic model of the CRRT treatment process and estimates state variables.

[0048] The discretization unit first transforms the continuous processes of ionized calcium diffusion, citrate distribution, and biochemical reactions such as calcium-citrate complexation during CRRT treatment into a weak form of three-dimensional stochastic partial differential equations. Using the finite element method, the spatial domain is divided into a polyhedral mesh, and the time domain is decomposed into small step sizes, transforming the continuous partial differential equations into a numerically solvable set of discrete equations, laying the foundation for subsequent calculations. The parameter identification unit uses the discretized equation structure as a framework, employing observed values ​​such as ionized calcium concentration and blood flow velocity from the calibration data as... A variational Bayesian inference method is introduced; a prior distribution is defined. For example, assuming the diffusion coefficient follows a normal distribution; and variational distribution. To approximate the posterior distribution of the true parameters, a lower bound for evidence is constructed. Iterative optimization maximizes this lower bound to solve the diffusion matrix in the stochastic partial differential equation. , where represents the diffusion coefficient of each substance and the reaction term. The dynamic parameters and noise intensity matrix This allows the model parameters to match the statistical characteristics of the calibration data, improving the model's accuracy in depicting the CRRT process. The state estimation unit is based on the stochastic partial differential equations optimized by the parameter identification unit, in the form of… ,in The system encompasses ion calcium concentration fields, citrate concentration fields, and blood flow velocity fields, incorporating calibration data as initial conditions and boundary constraints. Using the stochastic finite element method combined with Euler-Markov time discretization, the system solves the equation numerically in the spatiotemporal domain, estimating the distribution and evolution of each state variable to generate high-dimensional state data. This data comprehensively characterizes the distribution and dynamic changes of substances during CRRT treatment, providing accurate state input for dimensionality reduction, noise reduction, and causal analysis in the fusion inference module, and supporting the model foundation for subsequent treatment parameter optimization.

[0049] In this embodiment, the parameter identification unit receives calibration data output by the acquisition and calibration module, including observed values ​​of ionic calcium concentration, citrate concentration, blood flow velocity, etc., at different times, as... The three-dimensional stochastic partial differential equation framework constructed by combining discretized units is in the form of: ,in State vectors characterizing the ionic calcium concentration field, citrate concentration field, and blood flow velocity field; prior distributions are defined. For example, assuming the diffusion matrix The diagonal elements follow a normal distribution, and the reaction term The biochemical reaction rates follow a Gamma distribution; a variational distribution is constructed. Substituting the posterior distribution of the approximation parameters into the formula for the lower bound of evidence. Maximize through iterative optimization Output the optimized diffusion matrix , where represents the diffusion coefficient of each substance and the reaction term. The kinetic parameters, such as the calcium-citric acid complexation rate and the noise intensity matrix. This allows the parameters of the stochastic partial differential equations to be matched with the actual material transport and biochemical reaction characteristics of CRRT. The state estimation unit then calls upon the optimization parameters output by the parameter identification unit, using calibration data as initial conditions and boundary constraints. For example, the initial conditions include the ionized calcium concentration and blood flow velocity at the treatment initiation time, while the boundary constraints include the filter inlet flow rate and replacement fluid concentration. These are then substituted into the stochastic partial differential equations. The state vector in the spatiotemporal domain is solved by discretization using the stochastic finite element method. The dynamic evolution distribution of ion calcium concentration field, citrate concentration field, and blood flow velocity field is obtained, i.e., high-dimensional state data, which enables the spatiotemporal dynamic characterization of material transport and biochemical reactions during CRRT treatment, providing a precise state basis for the fusion reasoning module to analyze the correlation between treatment parameters and physiological indicators.

[0050] The fusion inference module is used to perform dimensionality reduction and adversarial denoising on high-dimensional state data, generate low-dimensional feature vectors, and construct a causal relationship map between treatment parameters and physiological indicators based on the low-dimensional feature vectors to calculate the intervention effect. The fusion inference module includes a dimensionality reduction unit, a denoising unit, and a causal modeling unit. The dimensionality reduction unit is used to perform manifold dimensionality reduction on high-dimensional state data using a local linear embedding algorithm. The denoising unit is used to denoise the dimensionality-reduced high-dimensional state data based on a conditional generative adversarial network to generate low-dimensional feature vectors. The causal modeling unit is used to construct a causal relationship map between treatment parameters and physiological indicators based on the low-dimensional feature vectors and calculate the intervention effect using Do-Calculus.

[0051] The locally linear embedding algorithm reduces dimensionality by solving the following optimization problem: ,in, For low-dimensional manifold eigenvectors, To reconstruct the weight matrix locally, minimize Sure, The original high-dimensional data;

[0052] The formula for calculating the intervention effect is: ,in, For treatment parameters, As a physiological indicator, for The set of parent nodes, for Except External variables, For intervention hour The probability distribution.

[0053] Specifically, the fusion reasoning module receives high-dimensional state data output by the modeling and estimation module, which includes spatiotemporal evolution information of ion calcium concentration field, citrate concentration field, and blood flow velocity field. This data is processed sequentially by the dimensionality reduction unit, denoising unit, and causal modeling unit to provide the control strategy module with low-dimensional feature vectors and causal correlation patterns between treatment parameters and physiological indicators.

[0054] Specifically, the dimensionality reduction unit targets the nonlinear manifold structure of high-dimensional state data and employs a locally linear embedding algorithm: each spatiotemporal sample in the high-dimensional state data is used as the original high-dimensional data. For example, given the combination of ion calcium concentration and blood flow velocity at a certain time and a certain spatial grid point, first calculate the local reconstruction weight matrix between samples. It minimizes The local linear dependencies between samples are determined; then, based on this weight matrix, the low-dimensional manifold eigenvectors are solved. Its optimization objective is This method maps high-dimensional state data to a low-dimensional space, preserving its inherent manifold structure. While reducing the data dimensionality, it retains the key change patterns of material transport and physiological state during CRRT treatment, simplifying calculations for subsequent denoising and causal analysis.

[0055] The denoising unit receives the low-dimensional manifold feature vector output by the dimensionality reduction unit and constructs a conditional generative adversarial network: the generator takes the noisy low-dimensional features as input and combines the physiological constraint priors output by the modeling estimation module, such as the clinically reasonable range of ion calcium concentration, to output denoised feature candidates; the discriminator takes the low-dimensional features of the stable treatment period in the acquisition and calibration module as real samples. These low-dimensional features filter the self-calibration data, distinguish between generated features and real features, and iteratively optimize through adversarial training to make the generator output approximate the real low-dimensional features, remove the noise introduced by dimensionality reduction and the measurement noise of the original data, generate more accurate low-dimensional feature vectors, and improve the reliability of subsequent causal modeling.

[0056] The causal modeling unit extracts the correlation between treatment parameters and physiological indicators based on the denoised low-dimensional feature vectors. In one embodiment, treatment parameters include blood flow and ultrafiltration rate, which are obtained by mapping CRRT device parameters from the acquisition and calibration module. Physiological indicators include ionized calcium concentration and blood gas parameters, extracted from physiological signals and laboratory test data. The unit analyzes the conditional independence between variables using a PC algorithm to construct a causal relationship graph, clarifying the parent and child node relationships between treatment parameters and physiological indicators, such as the direct impact of ultrafiltration rate on ionized calcium concentration and intermediate variables. Do-Calculus is introduced to integrate treatment parameters... As intervention variables, such as adjusting the ultrafiltration rate, physiological indicators can be determined based on causal mapping. set of parent nodes For example, the concentration of ionic calcium, substituted into the intervention effect formula The probability distribution of the impact of quantitative treatment parameter adjustments on physiological indicators provides a causal basis for the control strategy module to optimize treatment plans, avoids the misleading effects of statistical correlations, and improves the targeting of treatment parameter adjustments.

[0057] In this embodiment, the dimensionality reduction unit receives high-dimensional state data output by the modeling and estimation module, covering spatiotemporal sampling information of the ion calcium concentration field, citrate concentration field, and blood flow velocity field, as... First, minimize the linear dependencies of the sample's local neighborhood. Determine the local reconstruction weight matrix ; and then To constrain the optimization problem, solve the optimization problem. Output low-dimensional manifold eigenvectors It retains the key correlation patterns between material transport and physiological state during CRRT treatment, achieves dimensionality reduction of high-dimensional data, and provides a simplified and high-fidelity input for the noise reduction unit.

[0058] The causal modeling unit receives the low-dimensional feature vector output by the denoising unit to extract treatment parameters. For example, ultrafiltration rate; physiological indicators For example, the concentration of ionic calcium is determined by constructing a causal relationship graph using a PC algorithm. set of parent nodes Variables including blood flow velocity and citrate concentration; As an intervention variable, screening Except External variables Substitute into the formula By combining the statistical distribution of low-dimensional feature vectors, the modeling and estimation module data is derived. , computational intervention hour The probability distribution of the treatment parameters is used to quantify the causal effect of treatment parameters on physiological indicators, providing a precise causal basis for the control strategy module to optimize treatment plans.

[0059] The control strategy module is used to construct a model predictive control framework based on a spatiotemporal dynamics model, generate a basic treatment plan that meets clinical constraints by combining intervention effects, and personalize its parameters through deep reinforcement learning. The control strategy module includes a basic control unit and an intelligent optimization unit. The basic control unit is used to construct the model predictive control framework based on the spatiotemporal dynamics model and generate a basic treatment plan that meets clinical constraints by combining intervention effects. Clinical constraints include blood flow range, ionized calcium concentration threshold, and ultrafiltration rate upper limit. The cost function of the model predictive control framework is... ,in, To predict the time domain, To control the time domain, To predict the state vector, For reference trajectory, To control variables, and These are the weight matrices for the state and control variables, respectively. The intelligent optimization unit is used to personalize the parameters of the basic treatment plan through deep reinforcement learning. The reward function for deep reinforcement learning is... ,in, To control the smoothing coefficient, To avoid the reward coefficient for complications, For indicator functions, Control variables for the basic treatment plan This is the reference state vector.

[0060] Specifically, the control strategy module receives the spatiotemporal dynamic model output by the modeling and estimation module and the causal intervention effect from the fusion reasoning module, such as the probability distribution of the effect of ultrafiltration rate adjustment on ion calcium concentration. The spatiotemporal dynamic model describes the state evolution law of ion calcium concentration field, blood flow velocity field, etc. during CRRT treatment. Relying on the collaborative operation of the basic control unit and the intelligent optimization unit, it generates a treatment plan that meets clinical constraints and adapts to individual differences.

[0061] The basic control unit uses a spatiotemporal dynamic model as its prediction core, incorporates the causal intervention effect from the fusion inference module as a constraint, and integrates clinical constraints extracted from the acquisition and calibration module, including blood flow range, ionized calcium concentration threshold, and ultrafiltration rate upper limit, to construct a model prediction and control framework: setting the prediction time domain. With control time domain This is used to match the dynamic adjustment rhythm of CRRT treatment in order to model the predicted state vector output by the estimation module. For example, the predicted distribution of the ion calcium concentration field at future moments, and the reference trajectory for clinical targets. It derives and controls variables from normal physiological indicators of the data acquisition and calibration module. For example, the adjustment parameters of blood flow and ultrafiltration rate are used as inputs, and the cost function is optimized. ,in, , The weight matrix of state deviation and control energy consumption is used to balance the therapeutic effect and the stability of equipment operation, solve the control sequence that meets clinical constraints, generate the basic treatment plan, and ensure the safety and basic effectiveness of the treatment.

[0062] The intelligent optimization unit receives the control variables of the basic treatment plan. and the real-time state vector of the modeling and estimation module It extracts key physiological indicators from high-dimensional state data and introduces deep reinforcement learning: With reference state vector deviation, and The difference is used as a penalty term, and combined with the complication events monitored by the data acquisition and calibration module, the indicator function is driven by physiological signals for judgment. Construct a reward function ,in, Smooth control parameter fluctuations, Incentivize the avoidance of complications; optimize control strategies through iterative learning, and personalize the parameters of the basic treatment plan based on individual patient characteristics, such as underlying diseases and metabolic rate, using multi-source data implicit in the acquisition and calibration module, thereby improving the suitability of the treatment plan for the patient.

[0063] The instruction generation module integrates basic treatment plans and personalized adjustment parameters to generate device control instructions. These instructions are then executed in real-time at the edge via an edge-cloud collaborative architecture, and iteratively trained on the spatiotemporal dynamics model and deep reinforcement learning strategy in the cloud. The module comprises a weight fusion unit, a safety verification unit, and a collaborative architecture unit. The weight fusion unit dynamically adjusts the weights of the basic treatment plan and personalized adjustment parameters based on the treatment stage, merges them, and outputs parameter adjustment instructions. The safety verification unit filters out abnormal instructions based on causal constraints and clinical thresholds before generating device control instructions. The collaborative architecture unit executes real-time control at the edge based on these instructions via the edge-cloud collaborative architecture, and iteratively trains the spatiotemporal dynamics model and deep reinforcement learning strategy in the cloud through federated learning.

[0064] Specifically, the instruction generation module receives the basic treatment plan and personalized adjustment parameters output by the control strategy module, as well as treatment stage information fed back by the acquisition and calibration module, such as the initiation phase and stable phase identifiers. Through weighted fusion, safety verification, and collaborative execution, it generates device control instructions and drives model iteration. The weighted fusion unit dynamically adjusts the fusion weights of the basic treatment plan and personalized parameters based on the treatment progress monitored by the acquisition and calibration module, combined with device runtime and physiological signal trend determination stages: during the initiation phase, priority is given to ensuring the safety constraints of the basic plan, with a high weight; during the stable phase, personalized adaptation is gradually strengthened, increasing the weight of personalized parameters; the adjusted parameters are fused, and parameter adjustment instructions are output, connecting to the safety verification stage. The safety verification unit, upon receiving the parameter adjustment instructions, calls the causal relationship graph from the fusion inference module to extract the causal constraints of "treatment parameters → physiological indicators," such as the path where a sudden change in ultrafiltration rate triggers a decrease in ionized calcium. Combined with clinical thresholds from the acquisition and calibration module, such as ionized calcium ≥1.0 mmol / L, it filters out abnormal instructions that violate causal logic or thresholds; it integrates compliant instructions to generate device control instructions, ensuring that the instructions comply with treatment safety and clinical standards. After receiving device control commands, the collaborative architecture unit adapts to the HL7 FHIR protocol at the edge by acquiring and calibrating the device interface, driving the CRRT device to perform real-time control, such as adjusting ultrafiltration rate and blood flow. Simultaneously, it acquires real-time treatment data at the edge, such as device parameters and physiological feedback, and uploads it to the cloud. Based on federated learning, the cloud associates and trains the real-time data with the spatiotemporal dynamics model of the modeling and estimation module and the deep reinforcement learning strategy of the control strategy module, iteratively optimizing the model and strategy to achieve synergy between edge real-time control and cloud model evolution, thereby improving the system's adaptability to CRRT treatment.

[0065] The data acquisition and calibration module collects equipment parameters, physiological signals, and laboratory data through a three-level sensor network. It combines a hardware-synchronized clock and fractional-order calculus interpolation algorithms to achieve time alignment and spatial calibration of multi-source data, providing a high-precision standardized data foundation for subsequent modeling. The modeling and estimation module performs spatiotemporal dynamic modeling of material transport and biochemical reactions during CRRT treatment based on stochastic partial differential equations. It estimates state variables such as ion calcium concentration field and blood flow velocity field using calibration data, generating high-dimensional state data that characterizes the dynamic properties of the treatment process, thus achieving a mathematical description of the CRRT treatment mechanism. The fusion and inference module performs manifold dimensionality reduction and adversarial denoising on the high-dimensional state data. It constructs a causal relationship map between treatment parameters and physiological indicators using a PC algorithm and quantifies the intervention effect using Do-Calculus, thereby solving... The system analyzes the true causal path between parameter adjustments and physiological changes. The control strategy module combines spatiotemporal dynamics models with causal intervention effects to construct a model prediction control framework that generates a basic treatment plan that meets clinical constraints. Then, it uses deep reinforcement learning to achieve personalized parameter optimization, realizing the synergy between treatment plan safety and individual adaptability. The instruction generation module integrates the basic plan and personalized parameters, and generates equipment control instructions after causal constraints and clinical threshold verification. Through an edge-cloud collaborative architecture, it executes real-time control at the edge, while using federated learning to iteratively optimize the model and strategy in the cloud. This enables scientific analysis and precise intervention of causal relationships in CRRT treatment, solving the problem that traditional CRRT optimization systems cannot distinguish between true causal effects and confounding factors during parameter adjustment, leading to unexpected treatment side effects.

[0066] Please see the appendix Figure 2 The AI-based comprehensive CRRT data acquisition and optimization processing method, applied to the aforementioned AI-based comprehensive CRRT data acquisition and optimization processing system, includes the following steps:

[0067] Data Acquisition and Calibration: Collect CRRT equipment parameters, physiological signals, and laboratory test data to form multi-source data, and perform time alignment and spatial calibration to obtain calibration data;

[0068] Modeling and estimation: Based on stochastic partial differential equations, a spatiotemporal dynamic model of material transport and biochemical reactions in the CRRT treatment process is constructed, and state variables are estimated by combining calibration data to generate high-dimensional state data;

[0069] Fusion reasoning: Dimensionality reduction and adversarial denoising are performed on high-dimensional state data to generate low-dimensional feature vectors, and a causal relationship map between treatment parameters and physiological indicators is constructed based on the low-dimensional feature vectors to calculate the intervention effect;

[0070] Control strategy: A model prediction control framework is constructed based on a spatiotemporal dynamics model, and a basic treatment plan that meets clinical constraints is generated by combining the intervention effect. The parameters are then adjusted for individualization through deep reinforcement learning.

[0071] Command generation: By integrating basic treatment plans and personalized adjustment parameters, device control commands are generated. Real-time control is executed at the edge through an edge-cloud collaborative architecture, and iterative training of spatiotemporal dynamics models and deep reinforcement learning strategies is performed in the cloud.

[0072] Specifically, a three-tiered sensor network is used to collect CRRT equipment operating parameters, physiological monitoring signals, and laboratory test data in layers, forming multi-source heterogeneous data. The communication unit uses the HL7FHIR interface protocol to complete data conversion and preliminary timestamp alignment. The calibration unit triggers synchronous sampling of multiple devices with a hardware synchronization clock. For non-uniform dynamic data such as blood flow velocity fluctuations, a fractional-order calculus interpolation algorithm is introduced to fit the interpolated signal at the target time point, achieving time alignment of multi-source data. At the same time, based on a spatial coordinate mapping model, the differences in sensor deployment are corrected to complete the spatial calibration of the data. The output calibration data provides accurate spatiotemporal synchronization input for subsequent modeling, eliminating the interference of data heterogeneity on the model.

[0073] After receiving the calibration data, the discretization unit transforms the transport and biochemical reactions of substances such as ion calcium diffusion and citrate complexation in CRRT treatment into a weak form of three-dimensional stochastic partial differential equations. The spatiotemporal domain is then discretized using the finite element method to construct a numerical solution framework. The parameter identification unit introduces variational Bayesian inference, setting prior distributions for parameters such as diffusion coefficients and reaction rates. A variational distribution approximates the posterior, and iterative optimization using the lower bound of evidence formula ensures that the model parameters match the statistical characteristics of the calibration data. The state estimation unit substitutes the calibration data as initial and boundary conditions, employing the stochastic finite element method combined with Euler-Markov time discretization to solve for the dynamic evolution distributions of state variables such as the ion calcium concentration field and blood flow velocity field, generating high-dimensional state data to characterize the spatiotemporal dynamics of CRRT treatment.

[0074] For the high-dimensional state data in the modeling estimation, the dimensionality reduction unit employs a locally linear embedding algorithm. It first calculates the local reconstruction weights between samples, then maps the high-dimensional data to a low-dimensional manifold space based on these weights, preserving key state patterns in the CRRT process. The denoising unit constructs a conditional generative adversarial network (GAN), using noisy low-dimensional features as input and combining them with physiological constraint priors from the modeling estimation. Through adversarial training between the generator and discriminator, it removes data noise and outputs accurate low-dimensional feature vectors. The causal modeling unit extracts treatment parameters and physiological indicators from the low-dimensional features, constructs a causal graph using a PC algorithm to clarify the parent-node relationships of variables, and then uses the Do-Calculus formula to quantify the intervention effect of treatment parameter adjustments on physiological indicators, providing a causal basis for the control strategy and avoiding statistical association misleading information.

[0075] Based on the spatiotemporal dynamics model estimated through modeling and the intervention effect fused by inference, the basic control unit constructs a model predictive control framework, sets the prediction and control time domains, and optimizes the cost function using the predicted state vector, clinical reference trajectory, and control variables as inputs to generate a basic treatment plan that meets clinical constraints. The intelligent optimization unit introduces deep reinforcement learning, using the deviation between the real-time state and the reference state, and the difference between the control variables and the basic plan as penalty terms, and constructs a reward function in conjunction with complication indicators to iteratively optimize the control strategy. It personalizes the basic plan according to individual patient characteristics, improving treatment adaptability.

[0076] The system receives the basic control strategy and personalized parameters. The weighted fusion unit dynamically adjusts the weights of both based on the treatment stage feedback from the data acquisition and calibration, and outputs parameter adjustment instructions. The safety verification unit calls the causal graph from the fusion inference and the clinical thresholds from the data acquisition and calibration to filter out abnormal instructions that violate causality or thresholds, and generates compliant device control instructions. The collaborative architecture unit achieves edge-cloud collaboration. The edge device executes real-time control through the device interface from the data acquisition and calibration, simultaneously collecting real-time data and uploading it to the cloud. The cloud, based on federated learning, associates real-time data with the modeled and estimated spatiotemporal dynamics model and the reinforcement learning strategy of the control strategy for training, iteratively optimizing the model and strategy to achieve a closed loop of edge control and cloud evolution, continuously improving the system's optimization capabilities for CRRT treatment.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A comprehensive CRRT data acquisition and optimization processing system based on artificial intelligence, characterized in that, include: Acquisition and calibration module: used to acquire CRRT equipment parameters, physiological signals and laboratory test data, form multi-source data, and perform time alignment and spatial calibration to obtain calibration data; Modeling and estimation module: used to construct a spatiotemporal dynamic model of material transport and biochemical reactions in the CRRT treatment process based on stochastic partial differential equations, and estimate state variables by combining calibration data to generate high-dimensional state data; Fusion inference module: used to perform dimensionality reduction and adversarial denoising on high-dimensional state data, generate low-dimensional feature vectors, and construct a causal relationship map between treatment parameters and physiological indicators based on the low-dimensional feature vectors to calculate the intervention effect; Control strategy module: used to build a model prediction control framework based on spatiotemporal dynamics model, generate basic treatment plans that meet clinical constraints by combining intervention effects, and adjust the parameters of the plan through deep reinforcement learning; Command generation module: used to integrate basic treatment plans and personalized adjustment parameters to generate device control commands, execute real-time control at the edge through an edge-cloud collaborative architecture, and perform iterative training of spatiotemporal dynamics models and deep reinforcement learning strategies in the cloud; The modeling and estimation module includes a discretization unit, a parameter identification unit, and a state estimation unit. The discretization unit is used to transform the material transport and biochemical reactions in the CRRT treatment process into a weak form of a three-dimensional stochastic partial differential equation. The parameter identification unit is used to optimize the parameters of the three-dimensional stochastic partial differential equation through variational Bayesian inference. The state estimation unit is used to solve the three-dimensional stochastic partial differential equation based on calibration data, estimate state variables, and generate high-dimensional state data. The stochastic partial differential equation is: , where the state vector Includes ionized calcium concentration field, citrate concentration field, and blood flow velocity field, diffusion matrix. It is a diagonal matrix, with elements representing the diffusion coefficients of each substance and reaction terms. Describe the kinetics of biochemical reactions. The noise intensity matrix is... For Wiener process; The optimization involves optimizing the parameters of a three-dimensional stochastic partial differential equation by maximizing the lower bound of evidence. The expression for maximizing the lower bound of evidence is as follows: ,in, It is a variational distribution. For the prior distribution, For observation data, Variational distribution The expected operator below, The Kullback-Leibler divergence; The fusion inference module includes a dimensionality reduction unit, a denoising unit, and a causal modeling unit. The dimensionality reduction unit is used to perform manifold dimensionality reduction on high-dimensional state data using a local linear embedding algorithm. The denoising unit is used to denoise the dimensionality-reduced high-dimensional state data based on a conditional generative adversarial network to generate low-dimensional feature vectors. The causal modeling unit is used to construct a causal relationship map between treatment parameters and physiological indicators based on the low-dimensional feature vectors and calculate the intervention effect using Do-Calculus. The locally linear embedding algorithm reduces dimensionality by solving the following optimization problem: ,in, For low-dimensional manifold eigenvectors, To reconstruct the weight matrix locally, minimize Sure, The original high-dimensional data; The formula for calculating the intervention effect is as follows: ,in, For treatment parameters, As a physiological indicator, for The set of parent nodes, for Except External variables, For intervention hour The probability distribution; The control strategy module includes a basic control unit and an intelligent optimization unit. The basic control unit is used to construct a model predictive control framework based on a spatiotemporal dynamics model, and generate a basic treatment plan that meets clinical constraints by combining the intervention effect. The clinical constraints include blood flow range, ionized calcium concentration threshold, and ultrafiltration rate upper limit. The cost function of the model predictive control framework is: ,in, To predict the time domain, To control the time domain, To predict the state vector, For reference trajectory, To control variables, and These are the weight matrices for the state and control variables, respectively. The intelligent optimization unit is used to personalize the parameters of the basic treatment plan through deep reinforcement learning. The reward function of the deep reinforcement learning is... ,in, To control the smoothing coefficient, To avoid the reward coefficient for complications, For indicator functions, Control variables for the basic treatment plan This is the reference state vector.

2. The AI-based CRRT data omnidirectional acquisition and optimization processing system according to claim 1, characterized in that: The acquisition and calibration module includes an acquisition unit, a communication unit, and a calibration unit. The acquisition unit is used to acquire CRRT equipment parameters, physiological signals, and laboratory test data through a three-level sensor network to form multi-source data. The communication unit is used to perform real-time interaction of various data through HL7 FHIR interface protocol conversion. The calibration unit is used to perform time alignment and spatial calibration of the multi-source data by using hardware synchronous clock-triggered sampling combined with fractional calculus interpolation algorithm to obtain calibration data.

3. The AI-based CRRT data omnidirectional acquisition and optimization processing system according to claim 2, characterized in that: The time alignment employs fractional calculus interpolation, the formula of which is: ,in, To optimize parameters and the range of values ​​is , For gamma function, For the interpolated signal at the target time point, The first derivative of the original signal. For signal acquisition time window, The target time alignment point.

4. The AI-based CRRT data omnidirectional acquisition and optimization processing system according to claim 1, characterized in that: The instruction generation module includes a weight fusion unit, a security verification unit, and a collaborative architecture unit. The weight fusion unit is used to fuse the weights of the basic treatment plan and personalized adjustment parameters dynamically adjusted based on the treatment stage, and output parameter adjustment instructions. The security verification unit is used to filter abnormal instructions based on causal constraints and clinical thresholds before generating device control instructions. The collaborative architecture unit is used to execute real-time control at the edge according to the device control instructions through an edge-cloud collaborative architecture, and to perform iterative training of spatiotemporal dynamics models and deep reinforcement learning strategies in the cloud through federated learning.

5. A comprehensive data acquisition and optimization processing method for CRRT based on artificial intelligence, characterized in that: The system for comprehensive acquisition and optimization processing of CRRT data based on artificial intelligence, as described in any one of claims 1-4, includes the following steps: Data Acquisition and Calibration: Collect CRRT equipment parameters, physiological signals, and laboratory test data to form multi-source data, and perform time alignment and spatial calibration to obtain calibration data; Modeling and estimation: Based on stochastic partial differential equations, a spatiotemporal dynamic model of material transport and biochemical reactions in the CRRT treatment process is constructed, and state variables are estimated by combining calibration data to generate high-dimensional state data; Fusion reasoning: Dimensionality reduction and adversarial denoising are performed on high-dimensional state data to generate low-dimensional feature vectors, and a causal relationship map between treatment parameters and physiological indicators is constructed based on the low-dimensional feature vectors to calculate the intervention effect; Control strategy: A model prediction control framework is constructed based on a spatiotemporal dynamics model, and a basic treatment plan that meets clinical constraints is generated by combining the intervention effect. The parameters are then adjusted for individualization through deep reinforcement learning. Command generation: By integrating basic treatment plans and personalized adjustment parameters, device control commands are generated. Real-time control is executed at the edge through an edge-cloud collaborative architecture, and iterative training of spatiotemporal dynamics models and deep reinforcement learning strategies is performed in the cloud.

Images