A radiology examination reservation intelligent sequencing and automatic scheduling system

Abstract

The present application relates to medical information processing and intelligent control technical field, especially to a kind of radiology examination reservation intelligent sequencing and automatic scheduling system, including the following modules: multi-source heterogeneous data acquisition module, for accessing equipment thermodynamics, patient metabolism and queue monitoring data, execute space-time alignment and normalization processing;Stochastic dynamics modeling module, for defining equipment thermodynamics process and patient metabolism process as deterministic drift term based on Ito stochastic process theory, defining environmental thermal fluctuation and individual metabolic difference as random diffusion term, constructing stochastic differential equations and outputting global random state vector.In the present application, through stochastic dynamics modeling and Lyapunov robust control, global optimal scheduling under complex disturbance is realized, thereby improving the traditional scheduling mostly adopts static rule, due to the impact of equipment heat accumulation and random noise is ignored, thereby causing equipment due to sudden overheating forced shutdown leads to system paralysis.

Description

Technical Field

[0001] This invention relates to the field of medical information processing and intelligent control technology, and in particular to an intelligent sorting and automatic scheduling system for radiology examination appointments. Background Technology

[0002] In the scheduling of radiology examinations in large general hospitals, the continuous operational stability of imaging equipment is strictly constrained by thermodynamic physical limits. For example, excessive heat capacity of the X-ray tube can force the equipment to shut down for cooling. Simultaneously, the timing of patient examinations is also constrained by physiological processes such as contrast agent metabolism. However, the actual scheduling environment is extremely complex, filled with random disturbances such as ambient temperature fluctuations, individual metabolic differences, and emergency queue jumping.

[0003] However, traditional scheduling mostly uses static rules, which ignore the impact of equipment heat accumulation and random noise, resulting in equipment being forced to shut down due to sudden overheating, causing system paralysis. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an intelligent sorting and automatic scheduling system for radiology examination appointments. It aims to improve the problem that traditional scheduling mostly uses static rules, which ignore the impact of equipment heat accumulation and random noise, resulting in equipment being forced to shut down due to sudden overheating and causing system paralysis.

[0005] This invention provides the following technical solution: an intelligent sorting and automatic scheduling system for radiology examination appointments, comprising the following modules:

[0006] The multi-source heterogeneous data acquisition module is used to access device thermodynamics, patient metabolism and cohort monitoring data, and perform spatiotemporal alignment and normalization processing;

[0007] The stochastic dynamics modeling module is used to define the thermodynamic process of the equipment and the metabolic process of the patient as deterministic drift terms, and the environmental thermal fluctuations and individual metabolic differences as stochastic diffusion terms, based on Itō's stochastic process theory, to construct a set of stochastic differential equations and output a global stochastic state vector.

[0008] The generalized entropy production functional construction module is used to map heat capacity excess, drug residues and queue backlog to entropy production indicators and construct a value functional with the goal of minimizing the expected cumulative entropy production.

[0009] The HJB neural operator solver module is used to fit the value functional using a physical information neural network, train the network with the goal of minimizing the Hamilton-Jacobi-Bellman equation residuals, and output the gradient vector field of the value function.

[0010] The shadow price extraction module is used to extract co-state variables from the gradient vector field and generate shadow price signals that characterize the marginal cost of resources.

[0011] The Lyapunov robust control module is used to generate a feedback control law based on shadow prices, and to generate the final scheduling instruction under the constraint of Lyapunov energy function decay.

[0012] By adopting the above technical solutions, stochastic dynamics modeling and Lyapunov robust control, global optimal scheduling under complex disturbances is achieved, thereby improving the problem that traditional scheduling mostly uses static rules, which ignores the impact of equipment heat accumulation and random noise, resulting in equipment being forced to shut down due to sudden overheating and causing system paralysis.

[0013] Optionally, the multi-source heterogeneous data acquisition module includes:

[0014] The analog signals of the percentage of anode heat capacity and cooling oil temperature of the X-ray tube of the imaging equipment are read through the underlying communication protocol of the equipment, and the analog signals are converted into discrete time series data.

[0015] By retrieving the patient's electronic medical record through the hospital information system interface, the glomerular filtration rate value and the dosage and type of contrast agent injected were analyzed.

[0016] The current number of people in the queue and the average service time per person are analyzed by using infrared flow counters or camera video streams in the waiting area.

[0017] Using the system master clock as a reference, the discrete time series data, patient physiological and biochemical index data, and queue monitoring data are interpolated and resampled to align the heterogeneous data on the time axis and normalize the various types of data by mapping them to a preset dimensionless numerical range.

[0018] Optionally, the stochastic dynamics modeling module includes:

[0019] The evolution of the X-ray tube heat capacity of imaging equipment is defined as a stochastic process containing deterministic and stochastic components. The natural heat dissipation process based on Newton's law of cooling and the heat accumulation process generated by the inspection task are mapped as deterministic drift terms, and the environmental thermal disturbance of the equipment's heat dissipation system is mapped as a stochastic diffusion term.

[0020] The evolution of tracer concentration in patients is defined as a stochastic process containing deterministic and stochastic components, where the first-order elimination kinetics based on glomerular filtration rate is mapped as a deterministic drift term, and the metabolic rate fluctuations caused by individual hemodynamic differences are mapped as a stochastic diffusion term.

[0021] By combining the deterministic drift terms with the stochastic diffusion terms, a set of stochastic differential equations describing the evolution of the system state is generated, and then numerically integrated to output the global stochastic state vector of the system.

[0022] Optionally, the stochastic dynamics modeling module further includes:

[0023] Treating the waiting queue as a continuous fluid, the difference between the patient arrival rate and the system service rate is mapped as a queue drift term;

[0024] The Poisson jump noise during the arrival and service processes is mapped as a queue diffusion term, and the fluctuation amplitude of this diffusion term is set to be positively correlated with the square root of the current queue length.

[0025] The queue drift term and queue diffusion term are incorporated into the set of stochastic differential equations as components describing the evolution of the congestion state.

[0026] Optionally, the generalized entropy-producing functional construction module includes:

[0027] Construct a heat dissipation penalty function, which calculates the square of the difference between the real-time heat capacity of the device and the safety threshold.

[0028] A physiological violation penalty function is constructed, which calculates the degree of violation of a patient being assigned an examination task when the drug residue concentration has not dropped to a safe level by a quadratic power, and an indicator function is introduced that takes effect only when the assignment action occurs.

[0029] Construct a congestion penalty function that calculates the quadratic length of the waiting queue.

[0030] The instantaneous total entropy productivity is obtained by weighted summation of the heat dissipation penalty function, the physiological violation penalty function, and the congestion penalty function. The value functional is generated by performing mathematical expectation operation on the integral of the instantaneous total entropy productivity over the infinite time domain.

[0031] Optionally, the HJB neural operator solving module includes:

[0032] Construct a generalized Hamiltonian operator, which includes three components: the instantaneous entropy production rate output by the generalized entropy production functional construction module, the dot product of the drift vector of the system's global random state vector and the gradient vector of the value function, and the trace operation of the diffusion matrix of the system's global random state vector and the Hessian matrix of the value function.

[0033] An extreme value condition equation is established that minimizes the generalized Hamiltonian operator. This equation is defined as the Hamilton-Jacobi-Bellman partial differential equation, which is used to describe the evolution of the value function in the stochastic state space.

[0034] Optionally, the HJB neural operator solving module further includes:

[0035] The input layer of the deep neural network receives a global random state vector, and the output layer outputs a scalar value function value.

[0036] Using an automatic differentiation algorithm, based on the forward propagation computation graph of the network, the first-order partial derivative of the output value with respect to the input vector is calculated in reverse to obtain the gradient vector, and the second-order partial derivative is calculated to obtain the Hessian matrix;

[0037] Substitute the gradient vector and Hessian matrix into the Hamilton-Jacobi-Bellman partial differential equation and calculate the difference between the left and right sides of the equation as the physical residual term.

[0038] Using the norm of the physical residual term as the loss function, the weight parameters of the neural network are iteratively updated using the gradient descent optimization algorithm until the physical residual term converges.

[0039] Optionally, the shadow price extraction module includes:

[0040] From the gradient vector field output by the HJB neural operator solver module, the partial derivative component corresponding to the heat capacity state dimension of the equipment is separated and defined as the heat capacity shadow price.

[0041] The partial derivative components corresponding to the patient's physiological state dimension are separated and defined as physiological shadow prices;

[0042] The amplitude of heat capacity shadow price and physiological shadow price is monitored. When the amplitude of either shadow price exceeds the preset marginal cost threshold, a corresponding resource shortage signal is generated.

[0043] Optionally, the Lyapunov robust control module includes:

[0044] Based on the extracted shadow price signal, a search is performed within the allowed value space of the control variables to find the combination of control variables that minimizes the generalized Hamiltonian operator, and this combination is used as the primary feedback control law.

[0045] When the heat capacity shadow price is positive and its magnitude increases, reduce the weight of the control variable that assigns examination tasks to the device; when the physiological shadow price is positive and its magnitude increases, reduce the weight of the control variable that ranks the patient for examination.

[0046] Optionally, the Lyapunov robust control module further includes:

[0047] Construct a quadratic positive definite function of the global random state vector as the Lyapunov energy function;

[0048] Calculate the time derivative of the Lyapunov energy function under the action of the primary feedback control law;

[0049] Check whether the derivative satisfies the decay condition of being less than zero;

[0050] If not satisfied, construct a quadratic programming problem. The objective of this problem is to minimize the Euclidean distance between the modified control law and the primary feedback control law, with the linear constraint that the derivative of the Lyapunov energy function satisfies the decay condition.

[0051] Solving this quadratic programming problem yields a modified control law, which serves as the final feedback control law.

[0052] The present invention has the following beneficial effects:

[0053] 1. In this invention, global optimal scheduling under complex disturbances is achieved through stochastic dynamic modeling and Lyapunov robust control, thereby improving the problem that traditional scheduling mostly adopts static rules, which ignores the impact of equipment heat accumulation and random noise, resulting in equipment being forced to shut down due to sudden overheating and causing system paralysis.

[0054] 2. In this invention, the HJB neural operator solution module is used to fit the value functional using a physical information neural network, thereby overcoming the problem of optimal control solution in high-dimensional state space. This improves the problem that traditional scheduling mostly uses conventional numerical programming algorithms, which cannot meet the real-time dynamic response requirements of radiology departments during peak periods because the computation time increases exponentially with the dimension.

[0055] 3. In this invention, the resource marginal cost signal generated by the shadow price extraction module guides the feedback control law, thereby realizing dynamic risk avoidance based on resource scarcity. This improves the problem that traditional scheduling mostly uses fixed rules for queuing, which lacks the quantification of dynamic risk costs, resulting in high-risk patients occupying high-load equipment and causing medical safety accidents. Attached Figure Description

[0056] Figure 1 This is an architecture diagram of a radiology examination appointment intelligent sorting and automatic scheduling system proposed in this invention;

[0057] Figure 2 This is a detailed flowchart of multi-source heterogeneous data acquisition for an intelligent sorting and automatic scheduling system for radiology examination appointments proposed in this invention;

[0058] Figure 3 This is a flowchart of the HJB neural operator solution and training process for an intelligent sorting and automatic scheduling system for radiology examination appointments proposed in this invention.

[0059] Figure 4 The flowchart shows the Lyapunov robust control logic of a radiology examination appointment intelligent sorting and automatic scheduling system proposed in this invention. Detailed Implementation

[0060] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1:

[0062] In a first embodiment of the present invention, the present invention provides an intelligent sorting and automatic scheduling system for radiology examination appointments, such as... Figures 1-4 As shown, it includes the following modules:

[0063] The multi-source heterogeneous data acquisition module is used to access device thermodynamics, patient metabolism and cohort monitoring data, and perform spatiotemporal alignment and normalization processing;

[0064] Furthermore, the multi-source heterogeneous data acquisition module includes:

[0065] The analog signals of the percentage of anode heat capacity and cooling oil temperature of the X-ray tube of the imaging equipment are read through the underlying communication protocol of the equipment, and the analog signals are converted into discrete time series data.

[0066] By retrieving the patient's electronic medical record through the hospital information system interface, the glomerular filtration rate value and the dosage and type of contrast agent injected were analyzed.

[0067] The current number of people in the queue and the average service time per person are analyzed by using infrared flow counters or camera video streams in the waiting area.

[0068] Using the system master clock as a reference, the discrete time series data, patient physiological and biochemical index data, and queue monitoring data are interpolated and resampled to align the heterogeneous data on the time axis and normalize the various types of data by mapping them to a preset dimensionless numerical range.

[0069] Specifically, the multi-source heterogeneous data acquisition module primarily addresses the technical challenges of inconsistent time bases and significant dimensional differences among physical equipment, physiological individuals, and environmental queues in radiology scheduling. As the system's input front-end, this module is responsible for converting analog signals, text data, and video streams into digital vectors in a unified format, providing a standardized data foundation for subsequent stochastic dynamics modeling.

[0070] The system acquires and discretizes the thermodynamic data of the equipment, directly connecting to the control unit of the imaging equipment via a medical device communication protocol interface. Sensors acquire analog voltage signals of the percentage of anode heat capacity and cooling oil temperature in the X-ray tube. An analog-to-digital converter operates at a preset sampling frequency. Analog signals are sampled to generate discrete time-series data.

[0071] set up The raw value of the X-ray tube heat capacity collected at any time is The original value of the cooling oil temperature is The system converts the analog quantities into digital sequences. and ,in This is the index for the sampling points.

[0072] The patient physiological and biochemical indicator analysis system retrieves the electronic medical record data of the patients through the hospital information system interface. Key physiological parameters are extracted from unstructured text or structured fields using a keyword matching algorithm. The extracted content includes glomerular filtration rate (GFR) values. Dosage of contrast agent to be injected And the type of contrast agent. For missing biochemical indicators, the system uses the average physiological statistics of the population in that age group to fill in the gaps and outputs a physiological feature vector. .

[0073] The system performs visual analysis of queue monitoring data by accessing video streams from cameras in the waiting area. The image processing unit performs head detection or infrared thermal imaging analysis on the video frames to calculate the number of people currently queuing in the waiting area in real time. Simultaneously, the timestamps of each patient entering and leaving the examination room are recorded, and the average service time per person within the sliding window is calculated. .

[0074] Spatiotemporal alignment and interpolation resampling are necessary because the update frequencies of device data, patient data, and queue data are inconsistent, requiring the use of the system master clock. Spatiotemporal alignment is performed based on this.

[0075] The system employs a linear interpolation algorithm to resample low-frequency data, synchronizing it with the sampling time of high-frequency data. For any non-sampling time... The data alignment formula is as follows:

[0076] ;

[0077] in Indicates the time to be requested Interpolated data; and They are the distances from time to time. The most recent previous and next sampling times; and These are the known observations at the corresponding times. Through this step, the system outputs a synchronization state vector that is strictly aligned with the time axis.

[0078] Data normalization is performed to eliminate the interference of different physical dimensions on the weights of subsequent dynamic modeling. The system maps all heterogeneous data to a preset dimensionless numerical range [0,1]. The range standardization method is used, and the normalization calculation formula is as follows:

[0079] ;

[0080] in This represents the dimensionless value after normalization. This represents the original value after interpolation and alignment; This indicates the historical statistical minimum or physical lower limit of this type of data; This indicates the historical maximum value or physical upper limit of this type of data.

[0081] The data input / output process can be summarized as follows: The input terminal receives analog voltage / current signals from the imaging equipment, HL7 format text data from the HIS system, and RTSP video streams from the surveillance cameras. After analog-to-digital conversion, text parsing, and image analysis, the data enters the spatiotemporal alignment unit. After performing interpolation based on the master clock and normalization based on the range, the output terminal generates a globally normalized state vector containing the equipment's thermal state, the patient's metabolic state, and the queue congestion state. It is then transferred to the stochastic dynamics modeling module.

[0082] The stochastic dynamics modeling module is used to define the thermodynamic process of the equipment and the metabolic process of the patient as deterministic drift terms, and the environmental thermal fluctuations and individual metabolic differences as stochastic diffusion terms, based on Itō's stochastic process theory, to construct a set of stochastic differential equations and output a global stochastic state vector.

[0083] Furthermore, the stochastic dynamics modeling module includes:

[0084] The evolution of the X-ray tube heat capacity of imaging equipment is defined as a stochastic process containing deterministic and stochastic components. The natural heat dissipation process based on Newton's law of cooling and the heat accumulation process generated by the inspection task are mapped as deterministic drift terms, and the environmental thermal disturbance of the equipment's heat dissipation system is mapped as a stochastic diffusion term.

[0085] The evolution of tracer concentration in patients is defined as a stochastic process containing deterministic and stochastic components, where the first-order elimination kinetics based on glomerular filtration rate is mapped as a deterministic drift term, and the metabolic rate fluctuations caused by individual hemodynamic differences are mapped as a stochastic diffusion term.

[0086] By combining the deterministic drift terms with the stochastic diffusion terms, a set of stochastic differential equations describing the evolution of the system state is generated, and then numerically integrated to output the global stochastic state vector of the system.

[0087] The stochastic dynamics modeling module also includes:

[0088] Treating the waiting queue as a continuous fluid, the difference between the patient arrival rate and the system service rate is mapped as a queue drift term;

[0089] The Poisson jump noise during the arrival and service processes is mapped as a queue diffusion term, and the fluctuation amplitude of this diffusion term is set to be positively correlated with the square root of the current queue length.

[0090] The queue drift term and queue diffusion term are incorporated into a set of stochastic differential equations as components describing the evolution of congestion states.

[0091] Specifically, the stochastic dynamics modeling module is configured to receive normalized data from the multi-source heterogeneous data acquisition module, and use Iton's stochastic process theory to deconstruct the physical evolution of the radiology system into a deterministic drift part and a stochastic diffusion part, thereby constructing a global stochastic state vector that reflects the uncertainty of the system.

[0092] Thermodynamic evolution modeling of imaging equipment treats the change in the heat capacity of the X-ray tube as a stochastic process. The deterministic drift term corresponds to the heat exchange process that follows physical laws, namely natural heat dissipation based on Newton's law of cooling and active heat generation caused by the inspection task; the stochastic diffusion term corresponds to the thermal fluctuations of the heat dissipation system caused by the slight perturbation of the ambient temperature.

[0093] Equipment heat capacity The stochastic differential equations are constructed as follows:

[0094] ;

[0095] in express Time of the first The percentage of heat capacity of the equipment; This indicates the device's natural heat dissipation coefficient, which is determined by the device's physical characteristics; For scheduling control variables, if and only if in The moment will be the first The patient was assigned to the first The value is 1 if there is a device, otherwise the value is 0. Indicates the first The nominal heat generation power of each patient's corresponding examination item; Indicates the time differential step size; The diffusion coefficient represents the thermal fluctuation of the equipment. The increment of the Wiener process, which follows a standard normal distribution, represents the white noise interference in the thermal environment.

[0096] The patient metabolic kinetics evolution modeling module models the metabolism of tracers or contrast agents in a patient as a first-order kinetic equation perturbed by individual differences. A deterministic drift term describes the ideal decay trajectory of the drug based on renal function, while a stochastic diffusion term describes the fluctuations in metabolic rate caused by individual hemodynamic differences.

[0097] Drug concentration in the patient's body The stochastic differential equations are constructed as follows:

[0098] ;

[0099] in express Time of the first The concentration of residual drugs in the patient's body; This represents the first-order elimination rate constant calculated based on the patient's glomerular filtration rate; Indicates the intensity coefficient of individual metabolic fluctuations; This represents the independent Brownian motion increment, characterizing the random metabolic noise of an individual organism. The diffusion term is in square root form here to ensure that the drug concentration remains non-negative throughout the evolution process.

[0100] The waiting queue fluid approximation modeling module uses a continuous fluid model to approximate the discrete queuing process. The difference between the patient arrival rate and the system service rate is mapped to a deterministic trend of queue length change, i.e., the drift term; the Poisson jump property of arrival interval and service duration is mapped to random fluctuation, i.e., the diffusion term.

[0101] Queue length The stochastic differential equations are constructed as follows:

[0102] ;

[0103] in express The equivalent fluid queue length of the waiting area at any given time; This indicates the average patient arrival rate within the current time window; This represents the current system's overall service rate to patients, and its value depends on the parallel processing capacity of all working devices. The coefficient representing the random oscillation amplitude during the queuing process; This represents the standard Brownian motion increment, which is positively correlated with the square root of the queue length, reflecting the statistical characteristic that the variance of a Poisson process is equal to its mean.

[0104] The data input / output flow for this module is as follows: Inputs include normalized discrete state values ​​of the equipment, patient physiological parameters, and the initial number of people in the queue. The module's internal processor uses the Euler-Maria numerical integration method to discretize and solve the three simultaneous stochastic differential equations. During each time step, the module calculates the increments of the deterministic components and the stochastic increments generated by the pseudo-random number generator, updating the values ​​of each state variable. The module's final output includes... The system's global random state vector, encompassing the thermal state of all devices, the drug concentration state of all patients awaiting examination, and the queue congestion state at any given moment. The vector is then transmitted to the subsequent generalized entropy production functional construction module and HJB solver module.

[0105] The generalized entropy production functional construction module is used to map heat capacity excess, drug residues and queue backlog to entropy production indicators and construct a value functional with the goal of minimizing the expected cumulative entropy production.

[0106] Furthermore, the generalized entropy production functional building blocks include:

[0107] Construct a heat dissipation penalty function, which calculates the square of the difference between the real-time heat capacity of the device and the safety threshold.

[0108] A physiological violation penalty function is constructed, which calculates the degree of violation of a patient being assigned an examination task when the drug residue concentration has not dropped to a safe level by a quadratic power, and an indicator function is introduced that takes effect only when the assignment action occurs.

[0109] Construct a congestion penalty function that calculates the quadratic length of the waiting queue.

[0110] The instantaneous total entropy productivity is obtained by weighted summation of the heat dissipation penalty function, the physiological violation penalty function, and the congestion penalty function. The value functional is generated by performing mathematical expectation operation on the integral of the instantaneous total entropy productivity over the infinite time domain.

[0111] Specifically, the generalized entropy production functional construction module is configured to receive the global random state vector output by the stochastic dynamics modeling module. It utilizes non-equilibrium thermodynamics principles to map the system's physical limits, physiological violations, and queue backlog into a unified scalarized cost index, i.e., entropy production. This module constructs an instantaneous Lagrangian and performs infinite time-domain integration to output a target value functional for optimal control solutions.

[0112] Construction of the heat dissipation penalty function, module definition of heat dissipation penalty function. This function characterizes the increase in system disorder caused by overheating of imaging equipment. It only produces a non-zero value when the equipment's thermal capacity exceeds a physical safety threshold, employing a quadratic structure to impose a non-linear penalty on the extent of the exceedance. The calculation formula is as follows:

[0113] ;

[0114] in Indicates the total number of imaging devices; express Time of the first Real-time heat capacity percentage of the equipment; This indicates the preset heat capacity safety threshold, such as 80%. This represents the heat dissipation weight coefficient, used to adjust the proportion of this item in the overall objective; This indicates a linear rectification operation, ensuring that no penalty is incurred if the limit is not exceeded.

[0115] Construction of the physiological violation penalty function, module definition of physiological violation penalty function. This function represents the risk entropy resulting from forcibly performing examinations in violation of a patient's metabolic patterns. It introduces a scheduling action indicator variable and only takes effect when the system issues an instruction to assign a patient for examination and the patient's drug concentration is below the target level. The calculation formula is as follows:

[0116] ;

[0117] in This indicates the total number of patients awaiting examination. For binary control variables, when the system is The moment to decide the patient Assigned to device The value is 1 if the condition is met, and 0 otherwise. Indicates the patient Current residual concentration of tracer in the body; This indicates the medically permissible safe residual concentration threshold for contrast agents; This represents the physiological violation weighting coefficient.

[0118] The construction of the congestion penalty function, module definition of the congestion penalty function. This characterizes the system efficiency entropy reduction caused by queue backlog. The queue length term is squared, causing the system penalty value to increase parabolically with the number of people in the queue, thus forcing the controller to prioritize handling long queue backlogs. The calculation formula is as follows:

[0119] ;

[0120] in express Real-time queue length in the waiting area; This represents the congestion weighting coefficient.

[0121] The value functional generation module first linearly superimposes the three penalty functions mentioned above to generate the instantaneous total entropy yield, which describes the operating cost of the system at the current moment. :

[0122] ;

[0123] Subsequently, the module defines the value functional. This functional represents the state from the current initial state. Starting from an infinitely long time in the future, we can calculate the mathematical expectation of the instantaneous total entropy productivity after time discounting. The value functional formula is as follows:

[0124] ;

[0125] in It represents the mathematical expectation operator for a random path, used to handle Brownian motion uncertainties in stochastic differential equations; This represents the time discount factor, used to weigh the impact of current costs against future costs. express The system's global state observation value at time 1.

[0126] The data input / output flow is as follows: the input end of this module receives the system's global random state vector from the stochastic dynamics modeling module. This vector contains all devices. All patients and queue length The module processor operates based on preset weight parameters. and threshold parameters The instantaneous total entropy productivity is calculated in real time according to the above formula. The module output does not directly output numerical values. Instead, it transmits the constructed value functional structure and its corresponding instantaneous Lagrange expression to the HJB neural operator solving module as the physical constraint target for neural network training.

[0127] The HJB neural operator solver module is used to fit the value functional using a physical information neural network, train the network with the goal of minimizing the Hamilton-Jacobi-Bellman equation residuals, and output the gradient vector field of the value function.

[0128] Furthermore, the HJB neural operator solver module includes:

[0129] Construct a generalized Hamiltonian operator, which includes three components: the instantaneous entropy production rate output by the generalized entropy production functional construction module, the dot product of the drift vector of the system's global random state vector and the gradient vector of the value function, and the trace operation of the diffusion matrix of the system's global random state vector and the Hessian matrix of the value function.

[0130] An extreme value condition equation is established that minimizes the generalized Hamiltonian operator. This equation is defined as the Hamilton-Jacobi-Bellman partial differential equation, which is used to describe the evolution of the value function in the stochastic state space.

[0131] The HJB neural operator solving module also includes:

[0132] The input layer of the deep neural network receives a global random state vector, and the output layer outputs a scalar value function value.

[0133] Using an automatic differentiation algorithm, based on the forward propagation computation graph of the network, the first-order partial derivative of the output value with respect to the input vector is calculated in reverse to obtain the gradient vector, and the second-order partial derivative is calculated to obtain the Hessian matrix;

[0134] Substitute the gradient vector and Hessian matrix into the Hamilton-Jacobi-Bellman partial differential equation and calculate the difference between the left and right sides of the equation as the physical residual term.

[0135] Using the norm of the physical residual term as the loss function, the weight parameters of the neural network are iteratively updated using the gradient descent optimization algorithm until the physical residual term converges.

[0136] Specifically, the HJB neural operator solver module is configured as the core computing unit, using deep learning methods to numerically solve high-dimensional stochastic optimal control problems. This module approximates the value function through a physical information neural network, transforming the constraints of stochastic differential equations based on physical and physiological laws into the loss function of the neural network, thereby solving for the optimal control gradient field in the global state space without relying on a large amount of pre-labeled data.

[0137] The module for constructing the generalized Hamiltonian operator first builds the generalized Hamiltonian operator based on the principle of dynamic programming. This operator serves as a mathematical bridge connecting the stochastic evolution dynamics of a system with the entropy production optimization objective. The operator is a linear superposition of three parts: an instantaneous entropy production rate term, a deterministic drift coupling term, and a stochastic diffusion coupling term. The mathematical expression of the generalized Hamiltonian operator is as follows:

[0138] ;

[0139] in Represents the global random state vector of the system; Indicates scheduling control variables; This represents the instantaneous total entropy production rate input by the generalized entropy production functional construction module; This represents the deterministic drift vector input from the stochastic dynamics modeling module; Value function Regarding the state vector The first-order gradient vector; This represents the randomness diffusion matrix input from the stochastic dynamics modeling module; Value function Regarding the state vector The second-order Hessian matrix; This represents the matrix trace operation, used to calculate the second-order effect of random disturbances on the system energy.

[0140] The Hamilton-Jacobi-Bellman partial differential equation (HJB) is established based on the Bellman optimality principle, defining the HJB partial differential equation under infinite time-domain discounting costs. This equation describes the optimal control strategy. Under the influence of the value function The necessary spacetime conservation relations must be satisfied. The HJB equations are constructed as follows:

[0141] ;

[0142] in The time discount factor, Indicates within the allowed control domain We search for the infimum that minimizes the Hamiltonian operator.

[0143] The configuration and differential computation of the physical information neural network module constructs a deep neural network with a multilayer perceptron structure. To approximate the unknown value function ,in These are the network weight parameters. The input layer receives a globally random state vector. The dimension corresponds to the sum of the number of devices, the number of patients, and the queue variable. The output layer outputs a single scalar, which is the estimated value of the current state. .

[0144] The module utilizes automatic differentiation to process the forward propagation computation graph of the network. Through the chain rule, it accurately calculates the first-order partial derivative of the output scalar with respect to the input vector without numerical differencing. To obtain the gradient vector and second-order partial derivatives Obtaining the Hessian matrix .

[0145] The physical residual calculation and network training module transforms the HJB equation into an unsupervised learning loss function. The output of the neural network... Substituting its derivative into the HJB equation, the errors on both sides of the equation are calculated, i.e., the physical residual terms. :

[0146] ;

[0147] The loss function is defined as the mean square error of the physical residual on the set of state-space sampling points:

[0148] ;

[0149] in The number of batch sampling points. These are state points randomly sampled from the state space. The module employs gradient descent optimization algorithms such as Adam or L-BFGS, based on... For network parameters Perform iterative updates. When When the network converges to the preset threshold, the training is complete.

[0150] The data input / output flow for this module includes the drift vector provided by the stochastic dynamics modeling module as input. With diffusion matrix And the entropy production function provided by the generalized entropy production functional construction module. The module internally calculates an approximation function that satisfies the HJB equation through repeated iterative training of a neural network. The output does not output the value function itself, but instead outputs the gradient vector field of the value function with respect to the current state in real time. This gradient vector field contains guiding information for each point in the state space pointing towards the direction of minimum entropy production, and is transmitted to the shadow price extraction module for further processing.

[0151] The shadow price extraction module is used to extract co-state variables from the gradient vector field and generate shadow price signals that characterize the marginal cost of resources.

[0152] Furthermore, the shadow price extraction module includes:

[0153] From the gradient vector field output by the HJB neural operator solver module, the partial derivative component corresponding to the heat capacity state dimension of the equipment is separated and defined as the heat capacity shadow price.

[0154] The partial derivative components corresponding to the patient's physiological state dimension are separated and defined as physiological shadow prices;

[0155] The amplitude of heat capacity shadow price and physiological shadow price is monitored. When the amplitude of either shadow price exceeds the preset marginal cost threshold, a corresponding resource shortage signal is generated.

[0156] Specifically, the shadow price extraction module is configured as a signal analysis unit connecting the HJB solver and the controller, responsible for transforming abstract mathematical gradients into resource marginal cost signals with physical and economic significance. This module decouples co-state variables of different dimensions of state variables by analyzing the gradient vector field of the value function, providing a quantitative scarcity reference for scheduling decisions.

[0157] Decoupling and definition of gradient vector field; the module receives the gradient vector field output by the HJB neural operator solution module. In optimal control theory, the gradient of the value function is equivalent to the costate vector. Its physical meaning is the marginal impact of changing a unit state variable on the total future cost when the system is in a specific state. This is due to the input state vector... It is multidimensionally coupled; the module first slices the gradient vector according to the physical properties of the state variables. The gradient vector decomposition formula is as follows:

[0158] ;

[0159] in express The costate vector at time t, i.e., the shadow price vector; The value function is expressed with respect to the first... The partial derivative of the heat capacity of the equipment, where For device indexing, The module defines it as the shadow price of heat capacity. ; The value function is expressed with respect to the first... The partial derivative of the drug concentration in the patient's body is defined by the module as the physiological shadow price. ; This represents the partial derivative of the value function with respect to the queue length, i.e., the congestion shadow price.

[0160] Physical meaning analysis, shadow price of heat capacity The value reflects the sensitivity of increasing the heat load on the device at the current moment to the system's long-term entropy production. A higher value indicates that the device is closer to the edge of overheating and breakdown, and the higher the implicit cost of using the device. (Physiological shadow price) The value reflects the contribution of forcing the patient to undergo examination to the long-term risk of system violations. The higher the value, the less suitable the patient's current physiological state is for immediate examination.

[0161] Resource scarcity signal generation module has a built-in preset marginal cost threshold. The processor monitors the amplitude of various shadow prices in real time. When the shadow price of a certain dimension exceeds a certain threshold, it means that the scarcity of that resource has reached a critical point that threatens system stability. The signal triggering logic is as follows:

[0162] ;

[0163] in Indicates that for the first The equipment is facing resource shortages and binary signals are scarce. This represents the alarm threshold for the marginal cost of heat capacity. Similarly, the physiological scarcity signal for patients can be calculated.

[0164] The data input / output process involves the module directly reading the high-dimensional gradient vector field data calculated by the HJB neural operator solver module. Internally, the module performs a vector index slicing operation, splitting the uniform gradient vector into a device heat capacity shadow price sub-vector and a patient physiological shadow price sub-vector. The module output generates two sets of signals: one is a continuously changing shadow price value. and The first is transmitted as a weighting adjustment factor to the Lyapunov robust control module; the second is discrete resource scarcity signals. This serves as a hard constraint to trigger the controller's emergency avoidance logic.

[0165] The Lyapunov robust control module is used to generate a feedback control law based on shadow prices and generate the final scheduling instruction under the constraint of Lyapunov energy function decay.

[0166] Furthermore, the Lyapunov robust control module includes:

[0167] Based on the extracted shadow price signal, a search is performed within the allowed value space of the control variables to find the combination of control variables that minimizes the generalized Hamiltonian operator, and this combination is used as the primary feedback control law.

[0168] When the heat capacity shadow price is positive and its magnitude increases, reduce the weight of the control variable that assigns examination tasks to the device; when the physiological shadow price is positive and its magnitude increases, reduce the weight of the control variable that ranks the patient for examination.

[0169] The Lyapunov robust control module also includes:

[0170] Construct a quadratic positive definite function of the global random state vector as the Lyapunov energy function;

[0171] Calculate the time derivative of the Lyapunov energy function under the action of the primary feedback control law;

[0172] Check whether the derivative satisfies the decay condition of being less than zero;

[0173] If not satisfied, construct a quadratic programming problem. The objective of this problem is to minimize the Euclidean distance between the modified control law and the primary feedback control law, with the linear constraint that the derivative of the Lyapunov energy function satisfies the decay condition.

[0174] Solving this quadratic programming problem yields a modified control law, which serves as the final feedback control law.

[0175] Specifically, the Lyapunov robust control module is configured as the execution end of the system. Its core logic is to convert the economic signals output by the shadow price extraction module into on / off commands for physical devices and patient scheduling commands. This module introduces the theory of control Lyapunov functions to construct a two-layer control architecture of "primary optimization-secondary correction" to ensure that the generated scheduling commands strictly comply with the boundary constraints of the system's physical stability while satisfying optimality.

[0176] The module first generates the primary feedback control law based on the Pontryagin minimum principle. This involves generating the primary control law for the control variable. Allowable value space Within, the search enables the generalized Hamiltonian operator Control vector that achieves the minimum value .

[0177] Because the Hamiltonian operator contains a drift vector With shadow price vector The inner product terms, and the search logic of the control law, follow the following mathematical relationship:

[0178] ;

[0179] in This represents the primary feedback control law, i.e., the scheduling instruction that has not undergone stability correction; Indicates the feasible region of the control variable, such as ; Indicates the instantaneous entropy production rate; The input matrix represents the direct gain of the control action on the change in system state. This represents the shadow price vector.

[0180] According to the above formula, when the shadow price of the heat capacity of a certain device... When the value is positive and the magnitude increases, the weight of this term in the objective function increases significantly. To minimize the overall objective function, the algorithm will force the corresponding control variable... The value is set to 0, thus mathematically realizing the logic of reducing the task allocation weight of high-risk equipment. Similarly, when the price of physiological shadows... As the level rises, the probability of the patient being selected decreases.

[0181] The construction and derivative calculation of the Lyapunov energy function involves constructing a quadratic positive definite function of the global random state vector as the virtual energy function of the system, i.e., the Lyapunov function. :

[0182] ;

[0183] in Represents the global random state vector of the system; This represents a predefined positive definite symmetric weighted matrix.

[0184] The module uses Itoh's formula to calculate the expected value of the time derivative of the energy function along the system trajectory. :

[0185] ;

[0186] in This represents the gradient vector of the Lyapunov function; This represents the deterministic drift vector of the system; Represents the system's random diffusion matrix; The Hessian matrix representing the Lyapunov function is equal to the matrix in this case. .

[0187] The module sets the system stability decay constraint condition for the quadratic programming correction and final instruction generation. ,in The attenuation coefficient is positive. The module detects the primary control law. Does this condition meet? If not, it means the current optimal strategy may lead to system state divergence. In this case, the module constructs the following quadratic programming problem to forcibly modify the control law: ;

[0188] The constraints are:

[0189] ;

[0190] This represents the final feedback control law after modification; This represents the square of the Euclidean norm.

[0191] The physical meaning of this quadratic programming problem is: under the safety constraint of ensuring the monotonic decay of the system's energy, find a control law that is similar to the primary optimal control law. The control strategy that is closest to the target location.

[0192] The data input / output process involves this module receiving the shadow price signal output by the shadow price extraction module at its input end. and the state vector output by the stochastic dynamics modeling module The module processor first calculates the primary control law, then verifies and corrects it using a Lyapunov stability filter, and finally generates the final scheduling instruction matrix at the output. The instruction matrix is ​​decoded into a specific "time-patient-equipment" matching table and sent to the radiology department appointment management terminal for execution.

[0193] Example 2:

[0194] During the peak morning hours of 10:00 AM in the radiology department of a top-tier hospital, a large number of patients were accumulating in the waiting hall. The ambient temperature fluctuated due to the dense flow of people, affecting the natural cooling efficiency of the imaging equipment. At this time, of the three CT scanners in the department, the CT-A, due to continuous operation, had its X-ray tube heat capacity approaching the 90% overheating alarm threshold. The first patient in the waiting queue was an elderly patient requiring a high-heat-generating "contrast abdominal scan." This patient had weak renal function (low eGFR) and had just completed a previous examination, with the contrast agent not yet fully metabolized. Simultaneously, a patient with multiple injuries from a car accident was suddenly brought in from the emergency department, disrupting the original queuing sequence. The system needed to make real-time decisions under the multiple pressures of the equipment's physical thermal limits, the patient's physiological metabolic safety limits, and the sudden rush of emergency patients: whether to assign the elderly patient to the CT-A, which was on the verge of overheating, allow the equipment to cool down to ensure future system stability, or prioritize the emergency patient.

[0195] In the complex coupled scenarios involving the physical and thermodynamic limits of equipment, patient physiological and metabolic safety constraints, and multi-source random disturbances in the environment and arrival process, existing scheduling systems based on static rules or deterministic algorithms have significant technical shortcomings. Due to the lack of dynamic modeling of the stochastic evolution of the system's global state, existing technologies cannot quantify the cumulative impact of random diffusion terms such as environmental thermal fluctuations and individual metabolic differences on system stability. This results in an inability to find a globally optimal balance between "equipment overheating shutdown risk" and "patient drug residue violation risk." Furthermore, due to the lack of a marginal cost evaluation mechanism based on non-equilibrium entropy production and Lyapunov robust control constraints, the system is prone to making short-sighted decisions that prioritize current local utilization at the expense of long-term system efficiency when faced with Poisson jump noise impacts such as emergency queue jumping. This can lead to system-level collapses such as equipment overheating lockout, patient physiological indicators exceeding limits, or divergent congestion in the waiting queue. To address these problems, this invention provides an intelligent sorting and automatic scheduling system for radiology examination appointments, the structure of which is as follows: Figure 1 As shown. The specific implementation process of this system is as follows:

[0196] The multi-source heterogeneous data acquisition module performs spatiotemporal alignment and normalization, eliminating the dimensional differences between equipment physical signals and medical information data, and establishing a unified benchmark for multi-physics coupled calculations. The stochastic dynamics modeling module, based on Itō's stochastic process theory, deconstructs environmental thermal fluctuations and individual metabolic differences into diffusion terms, achieving a quantitative characterization of the system's stochastic evolution risk. The generalized entropy production functional construction module utilizes non-equilibrium thermodynamics principles, uniformly defining equipment thermal capacity exceeding limits, patient physiological violations, and queue congestion as generalized entropy production, overcoming the shortsightedness of traditional scheduling strategies by minimizing the expected cumulative entropy production in the infinite time domain. The HJB neural operator solving module utilizes physical information neural networks to overcome the problem of solving high-dimensional optimal control, outputting the value function gradient field in real time. The shadow price extraction module uses this to analyze resource marginal cost signals, achieving dynamic early warning of high-risk states. The Lyapunov robust control module introduces an energy function decay constraint, which forces the system state trajectory to converge toward a steady state. This ensures that the global optimality of the scheduling strategy and the absolute stability of the physical system are balanced under complex random disturbances, effectively preventing equipment overheating shutdowns and waiting queue divergence.

[0197] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart sorting and automatic scheduling system for radiology examination appointments, characterized in that, Includes the following modules: The multi-source heterogeneous data acquisition module is used to access device thermodynamics, patient metabolism and cohort monitoring data, and perform spatiotemporal alignment and normalization processing; The stochastic dynamics modeling module is used to define the thermodynamic process of the equipment and the metabolic process of the patient as deterministic drift terms, and the environmental thermal fluctuations and individual metabolic differences as stochastic diffusion terms, based on Itō's stochastic process theory, to construct a set of stochastic differential equations and output a global stochastic state vector. The generalized entropy production functional construction module is used to map heat capacity excess, drug residues and queue backlog to entropy production indicators and construct a value functional with the goal of minimizing the expected cumulative entropy production. The HJB neural operator solver module is used to fit the value functional using a physical information neural network, train the network with the goal of minimizing the Hamilton-Jacobi-Bellman equation residuals, and output the gradient vector field of the value function. The shadow price extraction module is used to extract co-state variables from the gradient vector field and generate shadow price signals that characterize the marginal cost of resources. The Lyapunov robust control module is used to generate a feedback control law based on shadow prices, and to generate the final scheduling instruction under the constraint of Lyapunov energy function decay.

2. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The multi-source heterogeneous data acquisition module includes: The analog signals of the percentage of anode heat capacity and cooling oil temperature of the X-ray tube of the imaging equipment are read through the underlying communication protocol of the equipment, and the analog signals are converted into discrete time series data. By retrieving the patient's electronic medical record through the hospital information system interface, the glomerular filtration rate value and the dosage and type of contrast agent injected were analyzed. The current number of people in the queue and the average service time per person are analyzed by using infrared flow counters or camera video streams in the waiting area. Using the system master clock as a reference, the discrete time series data, patient physiological and biochemical index data, and queue monitoring data are interpolated and resampled to align the heterogeneous data on the time axis and normalize the various types of data by mapping them to a preset dimensionless numerical range.

3. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The stochastic dynamics modeling module includes: The evolution of the X-ray tube heat capacity of imaging equipment is defined as a stochastic process containing deterministic and stochastic components. The natural heat dissipation process based on Newton's law of cooling and the heat accumulation process generated by the inspection task are mapped as deterministic drift terms, and the environmental thermal disturbance of the equipment's heat dissipation system is mapped as a stochastic diffusion term. The evolution of tracer concentration in patients is defined as a stochastic process containing deterministic and stochastic components, where the first-order elimination kinetics based on glomerular filtration rate is mapped as a deterministic drift term, and the metabolic rate fluctuations caused by individual hemodynamic differences are mapped as a stochastic diffusion term. By combining the deterministic drift terms with the stochastic diffusion terms, a set of stochastic differential equations describing the evolution of the system state is generated, and then numerically integrated to output the global stochastic state vector of the system.

4. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The stochastic dynamics modeling module also includes: Treating the waiting queue as a continuous fluid, the difference between the patient arrival rate and the system service rate is mapped as a queue drift term; The Poisson jump noise during the arrival and service processes is mapped as a queue diffusion term, and the fluctuation amplitude of this diffusion term is set to be positively correlated with the square root of the current queue length. The queue drift term and queue diffusion term are incorporated into the set of stochastic differential equations as components describing the evolution of the congestion state.

5. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The generalized entropy-producing functional construction module includes: Construct a heat dissipation penalty function, which calculates the square of the difference between the real-time heat capacity of the device and the safety threshold. A physiological violation penalty function is constructed, which calculates the degree of violation of a patient being assigned an examination task when the drug residue concentration has not dropped to a safe level by a quadratic power, and an indicator function is introduced that takes effect only when the assignment action occurs. Construct a congestion penalty function that calculates the quadratic length of the waiting queue. The instantaneous total entropy productivity is obtained by weighted summation of the heat dissipation penalty function, the physiological violation penalty function, and the congestion penalty function. The value functional is generated by performing mathematical expectation operation on the integral of the instantaneous total entropy productivity over the infinite time domain.

6. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The HJB neural operator solving module includes: Construct a generalized Hamiltonian operator, which includes three components: the instantaneous entropy production rate output by the generalized entropy production functional construction module, the dot product of the drift vector of the system's global random state vector and the gradient vector of the value function, and the trace operation of the diffusion matrix of the system's global random state vector and the Hessian matrix of the value function. An extreme value condition equation is established that minimizes the generalized Hamiltonian operator. This equation is defined as the Hamilton-Jacobi-Bellman partial differential equation, which is used to describe the evolution of the value function in the stochastic state space.

7. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The HJB neural operator solving module also includes: The input layer of the deep neural network receives a global random state vector, and the output layer outputs a scalar value function value. Using an automatic differentiation algorithm, based on the forward propagation computation graph of the network, the first-order partial derivative of the output value with respect to the input vector is calculated in reverse to obtain the gradient vector, and the second-order partial derivative is calculated to obtain the Hessian matrix; Substitute the gradient vector and Hessian matrix into the Hamilton-Jacobi-Bellman partial differential equation and calculate the difference between the left and right sides of the equation as the physical residual term. Using the norm of the physical residual term as the loss function, the weight parameters of the neural network are iteratively updated using the gradient descent optimization algorithm until the physical residual term converges.

8. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The shadow price extraction module includes: From the gradient vector field output by the HJB neural operator solver module, the partial derivative component corresponding to the heat capacity state dimension of the equipment is separated and defined as the heat capacity shadow price. The partial derivative components corresponding to the patient's physiological state dimension are separated and defined as physiological shadow prices; The amplitude of heat capacity shadow price and physiological shadow price is monitored. When the amplitude of either shadow price exceeds the preset marginal cost threshold, a corresponding resource shortage signal is generated.

9. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The Lyapunov robust control module includes: Based on the extracted shadow price signal, a search is performed within the allowed value space of the control variables to find the combination of control variables that minimizes the generalized Hamiltonian operator, and this combination is used as the primary feedback control law. When the heat capacity shadow price is positive and its magnitude increases, reduce the weight of the control variable that assigns examination tasks to the device; when the physiological shadow price is positive and its magnitude increases, reduce the weight of the control variable that ranks the patient for examination.

10. The intelligent sorting and automatic scheduling system for radiology examination appointments according to claim 1, characterized in that, The Lyapunov robust control module also includes: Construct a quadratic positive definite function of the global random state vector as the Lyapunov energy function; Calculate the time derivative of the Lyapunov energy function under the action of the primary feedback control law; Check whether the derivative satisfies the decay condition of being less than zero; If not satisfied, construct a quadratic programming problem. The objective of this problem is to minimize the Euclidean distance between the modified control law and the primary feedback control law, with the linear constraint that the derivative of the Lyapunov energy function satisfies the decay condition. Solving this quadratic programming problem yields a modified control law, which serves as the final feedback control law.

Images