Thyroid disease management and health data statistical analysis system

Abstract

The application relates to the technical field of medical data processing, and discloses a thyroid disease management and health data statistical analysis system. A physiological kinetics model containing a drug metabolism and endogenous secretion mechanism is constructed, discrete detection data are filled in time sequence blanks by using an unscented Kalman filtering algorithm, a continuous physiological panorama conforming to a human mass conservation law and a neural humoral regulation mechanism is generated, a time lag causal tensor construction mechanism with an adaptive causal window is introduced into the system, nonlinear lagging effects of hormone fluctuations on nodule morphological evolution are accurately captured, and deep causal chains of pathological causes and morphological results are restored, finally, an entropy-driven dynamic scheduling module based on information thermodynamics replaces a traditional fixed follow-up mode, a Lyapunov cognitive failure critical point is calculated as an optimal review time by quantizing an uncertainty accumulation process of a disease condition prediction and combining a risk adaptive mechanism.

Description

Technical Field

[0001] This application relates to the field of medical data processing technology, and in particular to a system for the management of thyroid diseases and statistical analysis of health data. Background Technology

[0002] Thyroid diseases, as common and frequently occurring diseases in the field of endocrinology, have complex pathological mechanisms and long courses, encompassing various types such as hyperthyroidism, hypothyroidism, thyroid nodules, and thyroid cancer. In the actual scenarios of clinical diagnosis and long-term health management, doctors' accurate assessment of patients' conditions highly relies on the comprehensive analysis of two types of core data: one is biochemical indicators reflecting the endocrine function status of the thyroid gland, mainly derived from laboratory information management systems, including numerical data such as thyroid-stimulating hormone (TSH), free triiodothyronine (FT3), free thyroxine, and related antibodies; the other is imaging data reflecting changes in the anatomical structure and morphology of the thyroid gland, mainly derived from medical imaging archives and communication systems, especially descriptive textual information in ultrasound reports regarding nodule size, border clarity, echo characteristics, aspect ratio, and blood flow signals. This dual monitoring of functional and morphological data is the cornerstone for achieving accurate subtyping of thyroid diseases, efficacy evaluation, and prognosis, and also the objective basis for developing personalized treatment plans.

[0003] However, under the existing medical information system, these two types of key data are often stored separately in heterogeneous business systems, resulting in serious compatibility barriers in terms of data dimension and time series. Specifically, biochemical indicators are presented as discrete time-series values ​​with clear quantitative characteristics; while imaging reports consist of unstructured natural language text, filled with vague qualitative descriptions. Although existing electronic medical record systems can aggregate and display these two types of data, they are essentially just simple physical stacking, lacking deep logical integration and data cleaning mechanisms. A more critical technical challenge is that the time points for patients to undergo blood tests and ultrasound examinations often do not overlap. This asynchronous sampling frequency makes it impossible to accurately align the two sets of data on the same health timeline. Due to the lack of effective heterogeneous data mapping and time-series coupling technologies, computers cannot directly convert the evolution trend of nodule morphology in imaging reports into calculable quantitative characteristics, let alone establish a mathematical correlation model between it and hormone level fluctuations. This lack of data processing capabilities prevents current auxiliary diagnostic systems from capturing the potential causal relationship between functional abnormalities and morphological deterioration. It makes it difficult to issue early warnings when hormone levels are still within the normal range but nodule morphology has already shown signs of deterioration, thus limiting the practical application effectiveness of intelligent diagnostic and treatment technologies in the whole-cycle management of thyroid diseases. Summary of the Invention

[0004] This application proposes a thyroid disease management and health data statistical analysis system to address the problems mentioned in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution: a thyroid disease management and health data statistical analysis system, comprising: a multimodal image feature quantification module, a physiological dynamics temporal completion module, a time-delay causal tensor construction module, and an entropy-driven dynamic scheduling module, wherein; The multimodal image feature quantization module receives raw thyroid ultrasound image data, performs region of interest extraction on the raw thyroid ultrasound image data to obtain thyroid anatomical region data, performs pixel grayscale normalization processing on the thyroid anatomical region data, calculates texture feature parameters based on pixel grayscale distribution, calculates geometric morphology parameters based on thyroid anatomical region data, and performs vectorization concatenation processing on the texture feature parameters and geometric morphology parameters to generate morphological feature vectors. The physiological dynamics time-series completion module receives thyroid function biochemical test data and drug prescription record data. Using a kinetic model that includes drug metabolism and endogenous secretion terms, it calculates drug intake parameters based on drug prescription record data, calibrates metabolic rate parameters based on thyroid function biochemical test data, and performs interpolation operations between discrete thyroid function biochemical test data time points to generate continuous hormone concentration curves. The time-delay causal tensor construction module receives morphological feature vectors and continuous hormone concentration curves. It uses a preset non-uniform convolution kernel function to perform weighted convolution operations on the continuous hormone concentration curves to obtain hormone cumulative effect features. It then performs tensor concatenation operations on the hormone cumulative effect features and morphological feature vectors to generate a joint feature tensor. The entropy-driven dynamic scheduling module inputs the joint feature tensor into the time series prediction model to generate probability distribution data of the disease status at future time steps, calculates the information entropy value of the probability distribution data of the disease status, and outputs the optimal re-examination time point when the cumulative value of the information entropy value exceeds the preset safety confidence threshold.

[0006] Furthermore, in the multimodal image feature quantization module, the thyroid anatomical region data undergoes pixel grayscale normalization processing. The specific operation for calculating texture feature parameters based on pixel grayscale distribution is as follows: The anisotropic diffusion equation for speckle denoising is used to process the thyroid anatomical region data. The diffusion coefficient function is called to calculate the gradient magnitude of the pixels in the thyroid anatomical region data. Based on the gradient magnitude, the pixels in the flat region and the pixels in the edge region are distinguished. The diffusion smoothing operation is performed on the pixels in the flat region and the diffusion stop operation is performed on the pixels in the edge region to generate pure scattering field data. The pure scattered field data is processed by calling the deep gain inversion algorithm. The physical compensation field data is constructed based on the preset average attenuation coefficient of thyroid tissue. The pure scattered field data and the physical compensation field data are multiplied point by point to generate the physical inversion sound field data that restores the true acoustic impedance characteristics of the tissue. A multi-scale Rayleigh entropy sliding window scan operation is performed on the physically inverted sound field data. The local background mean data of the physically inverted sound field data within the sliding window is calculated. The physically inverted sound field data within the sliding window is divided by the local background mean data to obtain normalized ratio data. The cube value of the normalized ratio data is calculated. The logarithm of the sum of the cube values ​​is calculated. The logarithm is used as the local acoustic entropy yield data and defined as the texture feature parameter.

[0007] Furthermore, in the multimodal image feature quantization module, the geometric morphological parameters are calculated based on the thyroid anatomical region data. The specific operation of vectorizing and concatenating the texture feature parameters with the geometric morphological parameters to generate the morphological feature vector is as follows: The edge contour data of the thyroid anatomical region is extracted, and the edge contour data is smoothed and fitted using a cubic spline curve function to generate continuous and differentiable parameterized curve data. The scale-normalized bending energy data of continuously differentiable parameterized curve data is calculated based on the principle of minimum surface energy in continuum mechanics. The local curvature values ​​of the sampling points on the continuously differentiable parameterized curve data are calculated, the square values ​​of the local curvature values ​​are calculated, the integral operation is performed on the square values ​​along the closed path of the continuously differentiable parameterized curve data to obtain the integral result value, the physical perimeter value of the continuously differentiable parameterized curve data is calculated, the integral result value is multiplied by the physical perimeter value to obtain the scale-normalized bending energy data, and the scale-normalized bending energy data is defined as a geometric morphology parameter. Obtain preset demographic distribution parameter data, perform standard score standardization on texture feature parameters using demographic distribution parameter data, perform standard score standardization on geometric morphology parameters using demographic distribution parameter data, and perform orthogonal vector concatenation operation on the standardized texture feature parameters and standardized geometric morphology parameters to generate morphological feature vector.

[0008] Furthermore, in the physiological dynamics time-series completion module, drug intake parameters are calculated based on drug prescription record data. The specific operation using a dynamic model that includes drug metabolism and endogenous secretion terms is as follows: A gastrointestinal absorption model based on the hydrodynamic hysteresis transport mechanism was constructed, and a random compliance variable following a Bernoulli distribution was generated. The random compliance variable was used to simulate the missed dose behavior and dose deviation behavior of patients during medication. An absorption delay kernel function was constructed using the gamma distribution probability density function. The absorption delay kernel function was used to simulate the drug dissolution process and transmembrane transport process in the gastrointestinal tract. The drug prescription record data, random compliance variable and absorption delay kernel function were convolved to generate exogenous drug entry flux rate data. The exogenous drug entry flux rate data was defined as the drug intake parameter. A coupled differential equation system incorporating the hypothalamic-pituitary-thyroid axis feedback mechanism was constructed. The endogenous secretion term was divided into a Michaelis-Menten kinetic channel controlled by thyroid-stimulating hormone (TSH) concentration and a linear gain channel controlled by autoantibody titers. Thyroid anatomical region volume data from a multimodal image feature quantification module was received. Using the thyroid anatomical region volume data as the basic energy production base, the nonlinear driving effect of TSH concentration data on the Michaelis-Menten kinetic channel and the anomalous driving effect of autoantibody titer data on the linear gain channel were calculated. The outputs of the Michaelis-Menten kinetic channel and the linear gain channel were superimposed to generate total endogenous secretion rate data.

[0009] Furthermore, in the physiological dynamics time-series completion module, metabolic rate parameters are calibrated based on thyroid function biochemical test data. Interpolation is performed between discrete time points of the thyroid function biochemical test data to generate a continuous hormone concentration curve. The specific operation is as follows: We acquire patients' liver and kidney function biochemical index data, construct a physiological elimination mapping function, use liver function biochemical index data to correct the weight of the liver deiodination metabolic pathway in the physiological elimination mapping function, use kidney function biochemical index data to correct the weight of the kidney excretion metabolic pathway in the physiological elimination mapping function, generate a dynamically changing individualized elimination rate constant, and define the individualized elimination rate constant as a metabolic rate parameter, thereby achieving heterogeneous adaptation for different liver and kidney function states. This study utilizes the unscented Kalman filter algorithm to process discrete thyroid function biochemical test data. Based on the law of conservation of mass, a state-space model is constructed. The exogenous drug influx rate data and total endogenous secretion rate data are used as the material inputs of the state-space model, while the product of the individualized elimination rate constant and the current hormone concentration is used as the material output. In the prediction step, coupled differential equations are used to deduce the free thyroxine concentration and thyroid-stimulating hormone concentration at unsampled times. In the update step, the Kalman gain is calculated using real thyroid function biochemical test data. The state vector and covariance matrix are corrected using the Kalman gain, and a continuous hormone concentration curve optimized for the entire time period is output.

[0010] Furthermore, in the time-delay causal tensor construction module, the specific operation of performing weighted convolution operations on continuous hormone concentration curves using a preset non-uniform convolution kernel function to obtain the hormone cumulative effect characteristics is as follows: Extract the morphological feature vectors corresponding to two adjacent imaging examination times, calculate the geometric change rate data of the morphological feature vectors between two adjacent imaging examination times, determine the length value of the adaptive causal influence window based on the geometric change rate data, and use the length value of the adaptive causal influence window to extract the historical time series data before the corresponding examination time from the continuous hormone concentration curve. The preset skewed gamma distribution function is invoked, and a physical singularity protection constant term is introduced into the skewed gamma distribution function. The physical singularity protection constant term is configured to make the denominator of the skewed gamma distribution function non-zero when the backtracking time step is zero. A definite integral operation is performed on the skewed gamma distribution function after introducing the physical singularity protection constant term to obtain the total influence weight value. The total influence weight value is used to perform a division operation on the skewed gamma distribution function to generate a normalized biological response kernel function. The normalized biological response kernel function is applied to historical time series data and discrete weighted convolution is performed. The output of the discrete weighted convolution operation is defined as the hormone cumulative effect feature.

[0011] Furthermore, in the time-delay causal tensor construction module, the specific operation of performing tensor concatenation between the hormone cumulative effect features and the morphological feature vectors to generate a joint feature tensor is as follows: The morphological feature vector is mapped to a preset spatial anatomical state manifold, and the hormone cumulative effect feature is mapped to a preset temporal physiological stress manifold. The Riemann manifold normalization algorithm is called, and the Riemann manifold normalization algorithm is used to perform centering and covariance normalization on the morphological feature vector in the spatial anatomical state manifold and the hormone cumulative effect feature in the temporal physiological stress manifold, respectively. The centering and covariance normalization processes eliminate the dimensional differences and numerical dynamic range differences between the morphological feature vector and the hormone cumulative effect feature, so that the morphological feature vector and the hormone cumulative effect feature have consistent statistical weight contributions in subsequent calculations. The morphological feature vectors after centering and covariance normalization are used as spatial dimension features, and the hormone cumulative effect features after centering and covariance normalization are used as temporal dimension features. Tensor dimension cascade operation is performed on the spatial dimension features and the temporal dimension features, and the coupled feature carrier generated by the tensor dimension cascade operation is defined as the joint feature tensor.

[0012] Furthermore, in the entropy-driven dynamic scheduling module, the specific operation of inputting the joint feature tensor into the time series prediction model to generate the probability distribution data of the disease status at future time steps is as follows: A variational encoder is constructed, and the encoding network of the variational encoder is used to map the joint feature tensor to an initial Gaussian distribution in the latent state space. The initial pathological state micro-vector is obtained by sampling from the initial Gaussian distribution, thereby introducing the initial cognitive uncertainty of observation error and biological noise. A neural stochastic differential equation is constructed as a time series prediction model. In the neural stochastic differential equation, a pathological drift field and a biological diffusion tensor fitted by a neural network are defined. The pathological drift field is used to simulate the deterministic growth trend of lesions under the constraints of anatomical structure, and the biological diffusion tensor is used to simulate the intensity of random fluctuations of lesions under the influence of metabolic microenvironment. Starting with the initial pathological state micro-vector, the evolution trajectory of the pathological state micro-vector at future time steps is deduced using neural stochastic differential equations, generating probability distribution data of disease state that diffuses over time.

[0013] Furthermore, in the entropy-driven dynamic scheduling module, the specific operation for calculating the information entropy value of the disease status probability distribution data is as follows: Based on the second-moment property of the Falk-Planck equation, the bio-diffusion tensor in the neural stochastic differential equation is extracted, the autocorrelation matrix of the bio-diffusion tensor is calculated, and the trace value of the autocorrelation matrix is ​​calculated. Half of the trace value is defined as the instantaneous information entropy generation rate. The instantaneous information entropy generation rate is used to quantify the information dissipation rate of the probability distribution data of the disease status over time. Perform time integration on the instantaneous information entropy generation rate to generate a cumulative information entropy value that monotonically increases with the prediction time length, thereby physically representing the total cumulative uncertainty of the system's perception of the patient's future condition.

[0014] Furthermore, in the entropy-driven dynamic scheduling module, the specific operation of outputting the optimal review time point when the accumulated value of information entropy exceeds the preset safety confidence threshold is as follows: The current pathological state micro-vector is mapped to a malignant risk potential energy value that characterizes the degree of malignancy risk. A hyperbolic risk decay function is constructed, and the basic signal-to-noise ratio tolerance is dynamically adjusted using the malignant risk potential energy value to generate a risk-adaptive safety confidence threshold. This results in a higher malignant risk potential energy value and a lower risk-adaptive safety confidence threshold. Monitor the changes in the cumulative information entropy value, determine the moment when the cumulative information entropy value first exceeds the risk-adaptive safety confidence threshold, and define this moment as the Lyapunov cognitive failure critical point. The maximum physical horizon is set as a fallback constraint. The Lyapunov cognitive failure threshold and the maximum physical horizon are compared, and the smaller value between the Lyapunov cognitive failure threshold and the maximum physical horizon is selected as the optimal review time.

[0015] The beneficial effects of this invention are as follows: This invention constructs a physiological dynamic model encompassing drug metabolism and endogenous secretion mechanisms. It utilizes an unscented Kalman filter algorithm to fill the temporal gaps in discrete detection data, generating a continuous physiological panorama consistent with the law of conservation of mass and neurohumoral regulation. The system introduces a time-delay causal tensor construction mechanism with an adaptive causal window, accurately capturing the nonlinear hysteretic effects of hormone fluctuations on nodule morphological evolution and reconstructing the deep causal chain between pathological causes and morphological outcomes. Finally, an entropy-driven dynamic scheduling module based on information thermodynamics replaces the traditional fixed follow-up model. By quantifying the accumulation of uncertainty in disease prediction and combining it with a risk adaptive mechanism, the Lyapunov cognitive failure threshold is calculated as the optimal time for follow-up examination. This scheme significantly improves the spatiotemporal resolution of thyroid nodule malignancy prediction, achieving a dual maximization of medical resource utilization efficiency and patient safety. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort: Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example like Figure 1 and Figure 2 As shown, this invention discloses a thyroid disease management and health data statistical analysis system, including: a multimodal image feature quantization module, a physiological dynamics temporal completion module, a time-delay causal tensor construction module, and an entropy-driven dynamic scheduling module, wherein; The multimodal image feature quantization module receives raw thyroid ultrasound image data, performs region of interest extraction on the raw thyroid ultrasound image data to obtain thyroid anatomical region data, performs pixel grayscale normalization processing on the thyroid anatomical region data, calculates texture feature parameters based on pixel grayscale distribution, calculates geometric morphology parameters based on thyroid anatomical region data, and performs vectorized concatenation processing on the texture feature parameters and geometric morphology parameters to generate morphological feature vectors.

[0019] This embodiment details the specific execution logic of the multimodal image feature quantification module in solving the two major technical challenges of gain dependence and morphological quantification scale sensitivity of thyroid ultrasound imaging equipment. The multimodal image feature quantification module transforms subjective image visual signals into objective physical quantification indicators through physical inversion sound field construction and topological differential geometric analysis.

[0020] Specifically, in the multimodal image feature quantization module, the thyroid anatomical region data undergoes pixel grayscale normalization processing, and the specific operation for calculating texture feature parameters based on pixel grayscale distribution is as follows: The anisotropic diffusion equation for speckle denoising is used to process the thyroid anatomical region data. The diffusion coefficient function is called to calculate the gradient magnitude of the pixels in the thyroid anatomical region data. Based on the gradient magnitude, the pixels in the flat region data and the pixels in the edge region data are distinguished. A diffusion smoothing operation is performed on the pixels in the flat region, and a diffusion stop operation is performed on the pixels in the edge region to generate clean scattering field data.

[0021] In this embodiment, the multimodal image feature quantization module first addresses the problem of electronic thermal noise and tissue scattering spots mixed in the original ultrasound signal. Existing technologies typically use Gaussian filtering, but Gaussian filtering can lead to blurred lesion edges and loss of the invasive features of malignant tumors.

[0022] The multimodal image feature quantization module uses the speckle denoising anisotropic diffusion equation (SRAD). During execution, a diffusion coefficient function is set. The diffusion coefficient function is a gradient-based nonlinear decay function. In this embodiment, the diffusion coefficient function is configured as the reciprocal of the square of the gradient magnitude.

[0023] The multimodal image feature quantization module calculates the gradient magnitude of each pixel in the thyroid anatomical region data. When the gradient magnitude is lower than the preset noise threshold, the pixel is determined to belong to a flat region. At this time, the diffusion coefficient function outputs a high diffusion coefficient value, and strong isotropic diffusion is performed to smooth electronic thermal noise. When the gradient magnitude is higher than the preset noise threshold, the pixel is determined to belong to an edge region. At this time, the diffusion coefficient function outputs a diffusion coefficient value close to zero, and the multimodal image feature quantization module performs a stop diffusion operation. Through this mechanism, while filtering out background noise, the high-frequency signal of the tissue interface is completely preserved, generating pure scattering field data.

[0024] In this embodiment, the preset noise threshold is set to 15-25 (gradient amplitude unit). The basis for this range is the Rayleigh distribution fitting analysis of the statistical characteristics of thyroid ultrasound speckle noise. Its physical meaning represents the maximum gradient fluctuation limit caused solely by electronic thermal noise. Its function is to act as a logic gate switch to ensure that the diffusion operation is only performed on the non-edge background area, preventing the tiny spicule features at the edge of the lesion from being incorrectly smoothed.

[0025] The pure scattered field data is processed by calling the deep gain inversion algorithm. Based on the preset average attenuation coefficient of thyroid tissue, physical compensation field data is constructed. The pure scattered field data and the physical compensation field data are multiplied point by point to generate physical inversion sound field data that restores the true acoustic impedance characteristics of the tissue.

[0026] In this embodiment, the multimodal imaging feature quantization module aims to eliminate the physical attenuation effect of ultrasound propagation in biological tissues. This effect causes the echo signal of deep tissues to be naturally lower than that of superficial tissues. Existing technology relies on doctors manually adjusting the time gain compensation (TGC) knob, which introduces a great deal of human subjectivity.

[0027] The multimodal image feature quantization module uses a depth gain inversion algorithm and presets the average attenuation coefficient of thyroid tissue. In this embodiment, based on the acoustic physical characteristics of human soft tissue and referring to the typical attenuation coefficient reference value of thyroid soft tissue in clinical ultrasound physics, the average attenuation coefficient of thyroid tissue is preferably set to 0.7dB / cm / MHz.

[0028] The multimodal image feature quantization module acquires the center frequency parameters of the ultrasound probe (usually 7-10MHz) and the geometric depth parameters of the pixels (in centimeters). The multimodal image feature quantization module uses an exponential function to construct physical compensation field data. The value of the physical compensation field data increases exponentially with the increase of geometric depth, and its growth rate is determined by the average attenuation coefficient of thyroid tissue.

[0029] Subsequently, a point-by-point multiplication operation is performed to multiply the pure scattered field data with the physical compensation field data. This operation cancels out the propagation loss of sound waves at the physical level, so that the physically inverted sound field data directly reflects the acoustic impedance characteristics of the thyroid tissue, and realizes the independence of the data from the parameters of the acquisition equipment.

[0030] A multi-scale Rayleigh entropy sliding window scan operation is performed on the physically inverted sound field data. The local background mean data of the physically inverted sound field data within the sliding window is calculated. The physically inverted sound field data within the sliding window is divided by the local background mean data to obtain normalized ratio data. The cube value of the normalized ratio data is calculated. The logarithm of the sum of the cube values ​​is calculated. The logarithm is used as the local acoustic entropy yield data and defined as the texture feature parameter.

[0031] In this embodiment, the multimodal image feature quantification module detects microcalcifications using thermodynamic statistical methods. Malignant microcalcifications physically manifest as strong scattering sources with drastic changes in acoustic impedance, and the signal disorder they cause in local areas is much higher than that of benign colloids.

[0032] Define the sliding window size parameter. In this embodiment, in order to adapt to the characteristic scale of the microcalcification cluster (the typical diameter of a single microcalcification point is usually less than 2 mm), the sliding window size parameter is preferably set to a physical range of 2 mm × 2 mm to ensure that the window can effectively cover the cluster structure composed of one or more microcalcification points.

[0033] The multimodal image feature quantization module calculates the local background mean data within a sliding window and divides the physically inverted sound field data by the local background mean data. This normalization step eliminates the difference in absolute echo intensity and retains only texture fluctuation information.

[0034] Subsequently, the Rayleigh entropy order parameter is introduced. In this embodiment, the Rayleigh entropy order parameter is fixed at 3. The third power value of the normalized ratio data is calculated. The physical basis for choosing the third power is that the non-Rayleigh scattering signal generated by microcalcification points has a significant long-tail effect under high-order moment statistics. The third power operation can specifically amplify the signal weight of strong scattering sources while suppressing the uniform scattering signal of background tissue. The multimodal image feature quantization module calculates the cumulative sum of the third power values ​​and takes the natural logarithm to obtain the local acoustic entropy yield data. This data, as a texture feature parameter, can sensitively quantify the microscopic pathological heterogeneity inside thyroid nodules.

[0035] Furthermore, in the multimodal image feature quantization module, the geometric morphological parameters are calculated based on the thyroid anatomical region data. The specific operation of vectorizing and concatenating the texture feature parameters with the geometric morphological parameters to generate the morphological feature vector is as follows: Edge contour data of the thyroid anatomical region is extracted, and a cubic spline curve function is used to perform smooth fitting on the edge contour data to generate continuous differentiable parametric curve data.

[0036] In this embodiment, the edge contour data is initially composed of discrete pixels. Directly calculating the curvature will produce huge quantization errors. The multimodal image feature quantization module uses a cubic spline curve function to interpolate and fit the edge contour data. The cubic spline curve function ensures that the fitted curve is continuous at the second derivative level. That is, the generated continuous differentiable parameterized curve data has smooth curvature change characteristics. This step transforms the discrete digital image boundary into a geometric curve that conforms to the definition of continuous medium mechanics, laying a mathematical foundation for subsequent accurate differential geometric calculations.

[0037] The scale-normalized bending energy data of continuously differentiable parameterized curve data is calculated based on the principle of minimum surface energy in continuum mechanics. The local curvature values ​​of the sampling points on the continuously differentiable parameterized curve data are calculated, the square values ​​of the local curvature values ​​are calculated, and the integral operation is performed on the square values ​​along the closed path of the continuously differentiable parameterized curve data to obtain the integral result value. The physical perimeter value of the continuously differentiable parameterized curve data is calculated, and the integral result value is multiplied by the physical perimeter value to obtain the scale-normalized bending energy data. The scale-normalized bending energy data is defined as a geometric morphological parameter.

[0038] Specifically, to address the physical limitation of traditional morphological indicators being sensitive to lesion size, a scale invariance analysis based on Noether's general theorem is introduced. The aim is to find an indicator such that a circle with a diameter of 1 cm and a circle with a diameter of 5 cm have the same energy value, while spiky lesions of any diameter have extremely high energy values. To this end, the local curvature value of each sampling point on the continuously differentiable parameterized curve data is calculated. This value characterizes the degree of folding of the boundary at the microscopic level. The system calculates the square of this local curvature value and performs integration along a closed path to obtain the integral result.

[0039] The key is the introduction of a physical perimeter value and a topological normalization constant as joint compensation terms. The physical perimeter value represents the macroscopic length of the lesion outline. In this embodiment, this value is expressed in millimeters as actually measured. The system first performs a multiplication operation, calculates the product of the integral result and the physical perimeter value, and obtains scale-independent raw energy data. It is worth noting that, according to the principle of differential geometry, for a perfect circle of any radius, the closed-path integral value of the square of its local curvature is inversely proportional to the radius, while its physical perimeter value is directly proportional to the radius. The product of the two is mathematically constant to four times the square of pi. In order to establish the morphological benchmark of benign nodules and normalize the morphological index of a perfect circle to a unit value of 1, the system introduces a topological normalization constant. In this embodiment, the topological normalization constant is preferably set to the reciprocal of four times the square of pi. The system multiplies the aforementioned scale-independent raw energy data by the topological normalization constant to finally obtain scale-normalized bending energy data.

[0040] By introducing this constant, the data becomes a pure, dimensionless shape descriptor, completely decoupled from the absolute size of the nodule. Any malignant nodule with infiltrative growth, spiculation, or lobulation features will have a scale-normalized bending energy data that is significantly greater than 1 due to the drastic fluctuations in its boundary curvature. This achieves scale-independent quantification of malignant morphological features and is defined as a geometric morphological parameter.

[0041] Obtain preset demographic distribution parameter data, perform standard score standardization on texture feature parameters using demographic distribution parameter data, perform standard score standardization on geometric morphology parameters using demographic distribution parameter data, and perform orthogonal vector concatenation operation on the standardized texture feature parameters and standardized geometric morphology parameters to generate morphological feature vector.

[0042] In this embodiment, the texture feature parameters belong to the acoustic statistical manifold, while the geometric morphology parameters belong to the Euclidean geometric manifold. The physical dimensions and numerical ranges of the two are completely different. Direct splicing will cause features with larger values ​​to mask features with smaller values. The multimodal image feature quantization module obtains preset demographic distribution parameter data, which comes from the statistical mean and standard deviation of a large-scale thyroid case database. Standardized processing (Z-Score) is performed to map the texture feature parameters and geometric morphology parameters to dimensionless relative deviation values. Subsequently, the multimodal image feature quantization module performs orthogonal vector splicing operation to generate a high-dimensional morphological feature vector.

[0043] This embodiment addresses the shortcomings of existing technologies, such as speckle noise reduction anisotropic diffusion, depth gain inversion, multi-scale Rayleigh entropy calculation, and scale-normalized bending energy analysis, by employing a series of techniques including speckle noise reduction anisotropic diffusion, depth gain inversion, multi-scale Rayleigh entropy calculation, and scale-normalized bending energy analysis. The physically inverted acoustic field data restores the true acoustic characteristics of the tissue, freeing feature extraction from the subjective adjustments of the operating physician. The scale-normalized bending energy data achieves absolute quantification of nodule morphological complexity, effectively distinguishing between the expansive growth of benign nodules and the invasive growth of malignant tumors. The generated morphological feature vector provides high-fidelity, physically meaningful input data for the subsequent physiological dynamics temporal completion module and causal tensor construction module, laying the foundation for accurate prediction by the entire system.

[0044] The physiological dynamics time-series completion module receives thyroid function biochemical test data and drug prescription record data. Using a kinetic model that includes drug metabolism and endogenous secretion terms, it calculates drug intake parameters based on drug prescription record data, calibrates metabolic rate parameters based on thyroid function biochemical test data, and performs interpolation operations between discrete time points of thyroid function biochemical test data to generate continuous hormone concentration curves.

[0045] This embodiment details the execution logic of the physiological dynamics temporal completion module in resolving the contradiction between the sparsity of hormone detection data and the dynamic changes in the patient's physiological state. By introducing stochastic process simulation and nonlinear dynamic inversion, the physiological dynamics temporal completion module reconstructs discrete medical data into a continuous physiological panorama that conforms to the hypothalamic-pituitary-thyroid axis (HPT axis) feedback mechanism. Compared with the linear interpolation or spline interpolation methods commonly used in the prior art, the solution in this embodiment is based on the law of conservation of mass and the principles of fluid mechanics, avoiding false peaks or physiological violations generated by mathematical interpolation in non-sampling intervals.

[0046] Specifically, in the physiological dynamics time-series completion module, the drug intake parameters are calculated based on drug prescription record data. The specific operation using a kinetic model that includes drug metabolism and endogenous secretion terms is as follows: A gastrointestinal absorption model based on the hydrodynamic hysteresis transport mechanism was constructed, generating a random compliance variable following a Bernoulli distribution. This random compliance variable was used to simulate patients' missed doses and dose deviations during medication administration. An absorption delay kernel function was constructed using the gamma distribution probability density function, and this function was used to simulate the drug's dissolution and transmembrane transport processes in the gastrointestinal tract. Drug prescription record data was convolved with the random compliance variable and the absorption delay kernel function to generate exogenous drug entry flux data, which was then defined as a drug intake parameter.

[0047] In existing technologies, the prescribed dose is usually directly regarded as the patient's actual intake, and the transit time of the drug in the digestive tract is ignored, which leads to significant errors at the model input.

[0048] In this embodiment, the physiological dynamics time sequence completion module first constructs a gastrointestinal absorption model and generates a random compliance variable that follows a Bernoulli distribution. The random compliance variable physically simulates the randomness of the patient's daily medication decision (i.e., the binary state of "taking medication" or "missing a dose").

[0049] In this embodiment, the probability of the random compliance variable is set based on the patient's historical follow-up records or self-reported data. For example, the probability of missing a dose is set to 10%. The physiological dynamics time-series completion module uses the random compliance variable to perform gating correction on the drug prescription record data.

[0050] Subsequently, a gamma distribution probability density function is introduced to construct an absorption delay kernel function. The absorption delay kernel function aims to simulate the hydrodynamic hysteresis effect of drug dissolution, transport, and transmembrane absorption in the gastrointestinal tract. Two key parameters are set: gamma distribution shape parameter and scale parameter. The product of the gamma distribution shape parameter and scale parameter physically corresponds to the average gastric emptying time. In this embodiment, based on human digestive physiology data, the product of the gamma distribution shape parameter and scale parameter is set to 2-4 hours to reflect the time delay from oral administration to entry into the blood circulation of the drug.

[0051] A convolution operation is performed, which convolves the corrected drug prescription record data with the absorption delay kernel function in the time domain. This operation is mathematically equivalent to calculating the weighted average arrival time of the drug molecule population in the digestive tract, and finally generates exogenous drug entry flux rate data. The exogenous drug entry flux rate data quantifies the mass flow rate of drug molecules entering the central blood circulation per unit time (in pmol / h), and the exogenous drug entry flux rate data is defined as the drug intake parameter.

[0052] A coupled differential equation system incorporating the hypothalamic-pituitary-thyroid axis feedback mechanism was constructed. The endogenous secretion term was divided into a Michaelis-Menten kinetic channel controlled by thyroid-stimulating hormone (TSH) concentration and a linear gain channel controlled by autoantibody titers. Thyroid anatomical region volume data from a multimodal image feature quantification module was received. Using the thyroid anatomical region volume data as the basic energy production base, the nonlinear driving effect of TSH concentration data on the Michaelis-Menten kinetic channel and the anomalous driving effect of autoantibody titer data on the linear gain channel were calculated. The outputs of the Michaelis-Menten kinetic channel and the linear gain channel were superimposed to generate total endogenous secretion rate data.

[0053] To accurately simulate the secretory behavior of the thyroid gland under different pathological conditions, especially to cover the full spectrum of Graves' disease (hyperthyroidism) and Hashimoto's disease (hypothyroidism), a coupled differential equation system including the HPT axis feedback mechanism was constructed. First, the endogenous secretion term was processed. Since thyroid secretion is not only controlled by pituitary signals but also affected by autoimmune antibodies, the physiological dynamics time sequence completion module divided it into two independent channels.

[0054] The first is the Michaelis kinetics channel, which describes the normal driving effect of thyroid-stimulating hormone (TSH) on the gland. The physiological kinetics timing completion module introduces the Michaelis constant, which represents the substrate concentration at which the enzyme-catalyzed reaction rate reaches half of its maximum value. In this embodiment, the Michaelis constant is set to 1-5 mIU / L based on enzyme kinetics experimental data to characterize the affinity of the TSH receptor.

[0055] The nonlinear driving effect of thyroid-stimulating hormone (TSH) concentration data on the Michaelis-Menten kinetic channel was calculated, showing that as TSH concentration increases, the secretion rate first rises rapidly and then tends to saturate.

[0056] The second is the linear gain channel, used to describe the abnormal stimulation of glands by antibodies in autoimmune diseases such as Graves' disease. The physiological dynamics timing completion module introduces the antibody stimulation sensitivity coefficient. In this embodiment, the antibody stimulation sensitivity coefficient is preferably set to 0.2-0.8 pmol / (h·IU / L). This value range is based on the linear regression analysis of clinical thyroid-stimulating receptor antibody (TRAb) titers and free thyroxine (FT4) levels. Its physical meaning represents the additional thyroxine secretion rate stimulated by each 1 IU / L increase in antibody concentration. Its function is to quantify the pathological gain intensity under autoimmune attack, so that the model can accurately simulate the hormone surge phenomenon in hyperthyroidism.

[0057] The system receives thyroid anatomical region volume data from the preceding multimodal image feature quantification module, uses the thyroid anatomical region volume data as the baseline energy production base (i.e., reactor volume, in mL), superimposes the output of the Michaelis kinetic channel (normal secretion) and the output of the linear gain channel (abnormal secretion), and multiplies it by the baseline energy production base to generate total endogenous secretion rate data.

[0058] Furthermore, in the physiological dynamics time-series completion module, metabolic rate parameters are calibrated based on thyroid function biochemical test data. Interpolation is performed between discrete time points of the thyroid function biochemical test data to generate a continuous hormone concentration curve. The specific operation is as follows: We acquire patients' liver and kidney function biochemical data, construct a physiological elimination mapping function, and use the liver function biochemical data to correct the weights of the liver deiodination metabolic pathway in the physiological elimination mapping function, and use the kidney function biochemical data to correct the weights of the kidney excretion metabolic pathway in the physiological elimination mapping function, generating a dynamically changing individualized elimination rate constant. The individualized elimination rate constant is defined as a metabolic rate parameter, thereby achieving heterogeneous adaptation for different liver and kidney function states.

[0059] The elimination rate of drugs in the body is not a constant, but is limited by the metabolic capacity of the patient's liver and kidneys. Existing technologies usually assume that the elimination rate is a fixed value, which leads to prediction bias in patients with liver and kidney dysfunction. In order to solve this heterogeneity matching problem, the physiological dynamics time sequence completion module constructs a physiological elimination mapping function.

[0060] The physiological elimination mapping function establishes a mapping relationship between biochemical indicators and elimination rates, and obtains patients' liver function biochemical indicator data (such as alanine aminotransferase ALT, unit U / L) and kidney function biochemical indicator data (such as creatinine Cr, unit μmol / L).

[0061] The physiological dynamics time-series completion module introduces metabolic pathway weight parameters. Since thyroid hormones are mainly deiodinated in the liver and excreted in small amounts through the kidneys, in this embodiment, the weight of the liver deiodination metabolic pathway is set to 0.8, and the weight of the kidney excretion metabolic pathway is set to 0.2. This weight allocation is based on physiological statistical data of thyroid hormone metabolic pathways. The physiological dynamics time-series completion module uses normalized values ​​of liver function biochemical index data to correct the weight of the liver deiodination metabolic pathway and uses normalized values ​​of kidney function biochemical index data to correct the weight of the kidney excretion metabolic pathway.

[0062] For example, when a patient's ALT is elevated, indicating impaired liver function, the elimination rate calculated by the physiological elimination mapping function will automatically decrease. Ultimately, the physiological dynamics time-series completion module generates a dynamically changing individualized elimination rate constant (in units of 1 / h) and defines the individualized elimination rate constant as a metabolic rate parameter.

[0063] This study utilizes the unscented Kalman filter algorithm to process discrete thyroid function biochemical test data. Based on the law of conservation of mass, a state-space model is constructed. The exogenous drug influx rate data and total endogenous secretion rate data are used as the material inputs of the state-space model, while the product of the individualized elimination rate constant and the current hormone concentration is used as the material output. In the prediction step, coupled differential equations are used to deduce the free thyroxine concentration and thyroid-stimulating hormone concentration at unsampled times. In the update step, the Kalman gain is calculated using real thyroid function biochemical test data. The state vector and covariance matrix are corrected using the Kalman gain, and a continuous hormone concentration curve optimized for the entire time period is output.

[0064] To achieve high-precision state estimation with sparse observation data, the physiological dynamics time-series completion module adopts the unscented Kalman filter (UKF) algorithm. First, a state-space model is constructed based on the law of conservation of mass to clarify the source and destination of substances: the exogenous drug entry flux rate data and the total endogenous secretion rate data constitute the substance input (source) of the state-space model, while the product of the individualized elimination rate constant and the current hormone concentration constitutes the substance output (sink) of the state-space model.

[0065] In the prediction step of the unscented Kalman filter algorithm, numerical integration is performed using the aforementioned coupled differential equation system. The fourth-order Runge-Kutta method is introduced, and the integration step size is set to 1h. The prior estimates of free thyroxine concentration and thyroid-stimulating hormone concentration are calculated by extrapolating from the previous observation time to the current time.

[0066] In the update step of the unscented Kalman filter algorithm, when real thyroid function biochemical test data is received, the Kalman gain matrix is ​​calculated. The Kalman gain matrix measures the relative weight between model prediction error and measurement noise. The physiological dynamics time-series completion module uses the Kalman gain to correct the prior estimated state vector (hormone concentration) and covariance matrix (uncertainty range), so that the state estimate converges to the real observation value. Through this prediction-correction loop mechanism, the physiological dynamics time-series completion module fills the time gap between discrete detection points and finally outputs a continuous hormone concentration curve optimized for the entire time period.

[0067] This embodiment constructs a complete physiological dynamic model that includes gastrointestinal hysteresis absorption, dual-channel endogenous secretion, and liver and kidney function coupling elimination. Combined with the unscented Kalman filter algorithm, it solves the information blind spot problem caused by low sampling frequency in traditional medical data analysis. The exogenous drug entry flux data accurately restores the real exposure of the drug in the body. The dynamically changing individualized elimination rate constant achieves adaptive matching to the individual liver and kidney function of the patient. The resulting continuous hormone concentration curve not only accurately matches the test values ​​at the detection point, but also strictly conforms to the physical laws of biological metabolism and the law of conservation of mass in the unsampled interval. It provides a time-series benchmark with high temporal resolution (1h) and clear physical meaning for the next-level time-delay causal tensor construction module to analyze the long-term impact of drug cumulative exposure on nodule morphological evolution.

[0068] The time-delay causal tensor construction module receives morphological feature vectors and continuous hormone concentration curves. It uses a preset non-uniform convolution kernel function to perform weighted convolution operations on the continuous hormone concentration curves to obtain hormone cumulative effect features. It then performs tensor concatenation operations with the morphological feature vectors to generate a joint feature tensor.

[0069] This embodiment details the execution logic of the time-delay causal tensor construction module in handling the time lag effect between the morphological evolution of thyroid nodules and the fluctuation of hormone levels. By designing an adaptive causal influence window and a physically protected convolution kernel, the time-delay causal tensor construction module solves the problem of misjudgment of causal association caused by ignoring the lag of biological response in the prior art, and provides the system with a discriminative basis with spatiotemporal depth.

[0070] Specifically, in the time-delay causal tensor construction module, the specific operation of using a preset non-uniform convolution kernel function to perform weighted convolution operation on the continuous hormone concentration curve to obtain the hormone cumulative effect features is as follows: extract the morphological feature vectors corresponding to two adjacent image examination times, calculate the geometric change rate data of the morphological feature vectors between two adjacent image examination times, determine the length value of the adaptive causal influence window based on the geometric change rate data, and use the length value of the adaptive causal influence window to extract the historical time-series data before the corresponding examination time from the continuous hormone concentration curve.

[0071] In existing technologies, only the hormone levels at the moment of examination are usually analyzed, ignoring the cumulative response characteristics of tissue deformation to physiological environmental stress. In this embodiment, the time-delay causal tensor construction module first extracts the morphological feature vectors of two adjacent examination moments generated by the multimodal image feature quantization module, and performs vector subtraction and scalar division operations to calculate the geometric change rate data of the morphological feature vectors.

[0072] Geometric rate of change data represents the degree of evolutionary activity of thyroid nodules in the spatial dimension. In this embodiment, the geometric rate of change data is represented by the percentage of volume change, with a rate threshold set at 5%. The geometric rate of change data is derived from clinical medical classification standards for nodule growth dynamics, and its role is to dynamically adjust the temporal sampling breadth of causal analysis.

[0073] The time-delay causal tensor construction module determines the length of the adaptive causal influence window based on the geometric rate of change data. The length of the adaptive causal influence window represents the time span of the system's retrospective analysis of historical hormone data. In this embodiment, the length of the adaptive causal influence window is set within the range of 30-365 days. When the geometric rate of change data is greater than 5%, the length of the adaptive causal influence window is set to 30-90 days to capture high-frequency physiological fluctuations; when the geometric rate of change data is less than or equal to 5%, the length of the adaptive causal influence window is set to 180-365 days to capture long-term chronic stress. The length of the adaptive causal influence window is derived from the physical lag law of deformation of biological tissues under hormone stimulation. The purpose of the length of the adaptive causal influence window is to ensure the real observability of the input variables in the current clinical scenario.

[0074] The preset skewed gamma distribution function is invoked, and a physical singularity protection constant term is introduced into the skewed gamma distribution function. The physical singularity protection constant term is configured to make the denominator of the skewed gamma distribution function non-zero when the backtracking time step is zero. A definite integral operation is performed on the skewed gamma distribution function after introducing the physical singularity protection constant term to obtain the total influence weight value. The total influence weight value is used to perform a division operation on the skewed gamma distribution function to generate a normalized biological response kernel function.

[0075] Specifically, the time-delay causal tensor construction module uses the skewed gamma distribution function to simulate receptor saturation and biological signal transduction delay. A physical singularity protection constant term is introduced into the skewed gamma distribution function. The physical singularity protection constant term represents the displacement added to the denominator to ensure computational stability.

[0076] In this embodiment, the physical singularity protection constant term is set to d (days), the physical singularity protection constant term comes from the overflow protection mechanism in numerical calculation. The function of the physical singularity protection constant term is to ensure that the system will not produce division operation errors when the backtracking time step is 0 days.

[0077] The time-delay causal tensor construction module sets the tissue response sensitivity parameter and the biological effect half-life parameter. The tissue response sensitivity parameter represents the sensitive shape characteristics of pathological tissue to hormone stimulation. In this embodiment, the tissue response sensitivity parameter is set to 2.0. The tissue response sensitivity parameter is based on the differences in cell dynamics between different pathological types, such as papillary carcinoma or medullary carcinoma. The biological effect half-life parameter represents the time constant of physiological effect decay. In this embodiment, the biological effect half-life parameter is set in the range of 7-30 days. The biological effect half-life parameter is based on the physical process of hormone binding and dissociation from receptor in pharmacodynamics. The time-delay causal tensor construction module performs definite integral operation on the skewed gamma distribution function to obtain the total influence weight value, and performs normalized division operation using the total influence weight value to generate a normalized biological response kernel function. The normalized biological response kernel function ensures the conservation of the total influence in the physical system.

[0078] Subsequently, the normalized biological response kernel function is applied to the historical time series data and a discrete weighted convolution operation is performed. The output of the discrete weighted convolution operation is defined as the hormone cumulative effect feature.

[0079] Specifically, the time-delay causal tensor construction module uses the normalized biological response kernel function as a weighting operator to perform discretized weighted accumulation operations on the continuous hormone concentration curves from the physiological dynamics time-series completion module.

[0080] Discretized weighted accumulation physically simulates the process of tissue cells continuously enduring the integral of biochemical stress over a period of time. This step eliminates the measurement random error of a single sampling point and transforms continuous hormone fluctuations into scalar values ​​representing the intensity of long-term physiological environmental exposure, namely the hormone cumulative effect characteristic. The physical unit of the hormone cumulative effect characteristic is pmol / L.

[0081] Furthermore, in the time-delay causal tensor construction module, the specific operation of generating a joint feature tensor by performing tensor concatenation on the hormone cumulative effect features and morphological feature vectors is as follows: mapping the morphological feature vectors to a preset spatial anatomical state manifold, mapping the hormone cumulative effect features to a preset temporal physiological stress manifold, calling the Riemannian manifold normalization algorithm, and using the Riemannian manifold normalization algorithm to perform centering and covariance normalization on the morphological feature vectors in the spatial anatomical state manifold and the hormone cumulative effect features in the temporal physiological stress manifold, respectively. Through centering and covariance normalization, the dimensional differences and numerical dynamic range differences between the morphological feature vectors and the hormone cumulative effect features are eliminated, so that the morphological feature vectors and the hormone cumulative effect features have consistent statistical weight contributions in subsequent calculations.

[0082] Specifically, the morphological feature vector and the hormone cumulative effect feature originate from completely different physical dimensions. In order to achieve deep alignment of heterogeneous data, the time delay is caused by the tensor construction module calling the Riemann manifold normalization algorithm. The Riemann manifold normalization algorithm involves calculating the mean vector and covariance matrix of the feature distribution and performing centering and scaling processing.

[0083] In this embodiment, the standardized numerical range is mapped to the 0-1 interval. The Riemannian manifold normalization algorithm is derived from high-dimensional geometric statistics theory. The role of the Riemannian manifold normalization algorithm is to prevent the phenomenon of larger numerical dimensions masking smaller numerical dimensions in subsequent deep learning training, thereby ensuring that spatial anatomical information and temporal physiological information contribute equally.

[0084] Subsequently, the morphological feature vectors after performing centering and covariance normalization are used as spatial dimension features, and the hormone cumulative effect features after performing centering and covariance normalization are used as temporal dimension features. Tensor dimension cascade operation is performed on the spatial dimension features and the temporal dimension features, and the coupled feature carrier generated by the tensor dimension cascade operation is defined as the joint feature tensor.

[0085] Specifically, the time-delay causal tensor construction module stacks the processed spatial and temporal features into vectors along a preset axis to construct a high-dimensional joint feature tensor. The joint feature tensor not only contains the physical characteristics of the nodules at the current moment, but also the historical pathological causes that led to the morphological characteristics. This spatiotemporal coupled feature expression method solves the technical bottleneck of the isolation between image analysis and biochemical analysis, and realizes the technical evolution from static image monitoring to dynamic causal prediction.

[0086] This embodiment captures the nonlinear time-delay effect of hormone fluctuations on nodule morphological evolution by designing an adaptive causal influence window and a normalized biological response kernel function. By coupling discrete morphological feature vectors with continuous hormone concentration curves through convolution and tensor concatenation operations, the time-delay causal tensor construction module physically restores the complete causal chain from the accumulation of inducing factors to the morphological result. This method avoids data distortion caused by uneven follow-up intervals and improves the accuracy of disease progression prediction. The generated joint feature tensor provides a physically deterministic high-dimensional input for the subsequent entropy-driven dynamic scheduling module, ensuring that the generation of follow-up recommendations is based on objective causal evolution laws.

[0087] The entropy-driven dynamic scheduling module inputs the joint feature tensor into the time series prediction model to generate probability distribution data of the disease status at future time steps, calculates the information entropy value of the probability distribution data of the disease status, and outputs the optimal re-examination time point when the cumulative value of the information entropy value exceeds the preset safety confidence threshold.

[0088] This embodiment details the specific execution logic of the entropy-driven dynamic scheduling module in addressing the drawbacks of a one-size-fits-all approach in medical follow-up time decision-making. By introducing neural stochastic differential equations and the principle of entropy increase in information thermodynamics, the entropy-driven dynamic scheduling module upgrades time-based follow-up to follow-up based on cognitive uncertainty, achieving precise and dynamic allocation of medical resources. Compared with the static solutions in the prior art that follow up only based on fixed time intervals (such as every 3 months) or only based on the current severity of the condition, the technical solution adopted in this embodiment can calculate the Lyapunov prediction time boundary of the disease prediction path in the time dimension in real time, thereby determining the critical duration for the prediction information to remain within a safe signal-to-noise ratio range at the physical level, avoiding the medical safety risks caused by prediction failure due to information dissipation.

[0089] Specifically, in the entropy-driven dynamic scheduling module, the specific operation of inputting the joint feature tensor into the time series prediction model to generate the probability distribution data of the disease status at future time steps is as follows: A variational encoder is constructed, and the encoding network of the variational encoder is used to map the joint feature tensor to an initial Gaussian distribution in the latent state space. The initial pathological state micro-vector is obtained by sampling from the initial Gaussian distribution, thereby introducing the initial cognitive uncertainty of observation error and biological noise.

[0090] In existing technologies, most prediction models assume that the input data is absolutely accurate, ignoring the inherent signal-to-noise ratio limitations of medical imaging equipment. In order to introduce physically real initial cognitive uncertainty at the prediction starting point, the entropy-driven dynamic scheduling module constructs a variational encoder, which receives a joint feature tensor from the time-delay causal tensor construction module. The joint feature tensor contains deterministic information of spatiotemporal coupling.

[0091] The joint feature tensor is mapped to an initial Gaussian distribution in the latent state space using a variational encoder network. The initial Gaussian distribution is uniquely determined by the mean vector and the variance vector. In this embodiment, the dimension of the latent state space is set to 128 dimensions. The basis for choosing 128 dimensions for the latent state space is to balance the richness of feature representation with the consumption of computational resources. Too low a dimension cannot bear complex pathological features, while too high a dimension will lead to sparsity problems.

[0092] Performing Monte Carlo sampling from the initial Gaussian distribution to obtain the initial pathological state microvector ensures that the starting point of the model prediction is not a definite geometric point, but a wave packet with a probabilistic volume, which is consistent with the physical fact that the observed value itself has uncertainty in the quantum measurement principle.

[0093] A neural stochastic differential equation is constructed as a time series prediction model. In the neural stochastic differential equation, a pathological drift field and a biological diffusion tensor fitted by a neural network are defined. The pathological drift field is used to simulate the deterministic growth trend of lesions under the constraints of anatomical structure, and the biological diffusion tensor is used to simulate the intensity of random fluctuations of lesions under the influence of metabolic microenvironment.

[0094] The pathological evolution of organisms is the result of the combined effects of deterministic trends and random fluctuations. The entropy-driven dynamic scheduling module constructs neural stochastic differential equations as a time series prediction model. The neural stochastic differential equations contain two core physical terms: the pathological drift field and the biological diffusion tensor.

[0095] The pathological drift field is fitted by a multilayer perceptron neural network. In this embodiment, the output value of the pathological drift field is typically between 0.001 and 0.05 per day, physically representing the deterministic evolution rate of the lesion per unit time. This value range is based on clinical statistics of the natural growth rate of thyroid nodules (i.e., the daily volume or diameter change rate is typically between 0.1% and 5%), and its physical meaning represents the deterministic growth rate of the lesion under anatomical constraints (such as growth along lymphatic vessels). Its function is to drive the predicted trajectory towards a specific direction of deterioration or improvement.

[0096] The bio-diffusion tensor is fitted by a convolutional neural network. In this embodiment, the output value of the bio-diffusion tensor typically falls within the range of... The square root of each day physically quantifies the intensity of instantaneous random fluctuations in lesions under the influence of metabolic microenvironment disturbances. This numerical range is based on empirical estimates of the intensity of Brownian motion in biological systems, ensuring that random fluctuations do not overwhelm deterministic trends. Its physical meaning represents the intensity of random fluctuations in lesions under the influence of the metabolic microenvironment (such as gene mutations and the randomness of angiogenesis). Its function is to introduce noise that conforms to biological laws, so that the prediction results present a probability distribution rather than a single curve.

[0097] Starting with the initial pathological state micro-vector, the evolution trajectory of the pathological state micro-vector at future time steps is deduced using neural stochastic differential equations, generating probability distribution data of disease state that diffuses over time.

[0098] The entropy-driven dynamic scheduling module starts with the initial pathological state micro-vector and uses Itō's integral method to solve the neural stochastic differential equation. In this embodiment, the time step of the numerical integration is set to 1 day. The selection of this step step is based on balancing the computational accuracy and the efficiency of long-term prediction, and it matches the day as the smallest unit of clinical follow-up. The evolution trajectory of the pathological state micro-vector in the future time step is deduced. As the prediction time progresses, due to the existence of the biological diffusion tensor, the evolution trajectory gradually diverges, forming a probability distribution data of the disease state that diffuses over time. This distribution data completely describes all the micro-states and their probability densities that the patient's condition may be in at future moments.

[0099] Furthermore, in the entropy-driven dynamic scheduling module, the specific operation for calculating the information entropy value of the disease status probability distribution data is as follows: Based on the second-moment property of the Falk-Planck equation, the biodiffusion tensor in the neural stochastic differential equation is extracted, the autocorrelation matrix of the biodiffusion tensor is calculated, and the trace value of the autocorrelation matrix is ​​calculated.

[0100] To quantify the growth rate of uncertainty in prediction, the entropy-driven dynamic scheduling module, based on the second-moment property of the Falk-Planck equation, focuses on the impact of the diffusion term on the volume of the probability density function. It extracts the bio-diffusion tensor from the neural stochastic differential equation, performs matrix multiplication, calculates the product of the bio-diffusion tensor and its transpose, and generates an autocorrelation matrix. The autocorrelation matrix physically represents the instantaneous diffusion covariance matrix, describing the intensity coupling of random forces in various dimensions. The trace value of the autocorrelation matrix, i.e., the sum of the main diagonal elements, is calculated. The trace value physically represents the total mean square displacement rate of the multidimensional random walk.

[0101] Half of the trace value is defined as the instantaneous information entropy generation rate. The instantaneous information entropy generation rate is used to quantify the rate at which information dissipates from the probability distribution data of the disease status over time.

[0102] The entropy-driven dynamic scheduling module defines half of the trace value as the instantaneous information entropy generation rate, measured in Nat / d. In this embodiment, the instantaneous information entropy generation rate typically ranges from 0.01 to 0.5 Nat / d, and its value is derived from the trace value calculation results of the biological diffusion tensor in the neural stochastic differential equation. Its physical meaning represents the rate at which the disease prediction distribution diffuses over time, i.e., the speed of information loss. Its function is to provide slope support for subsequent calculation of the Lyapunov cognitive failure threshold by quantifying the intensity of daily information dissipation. A larger value indicates a stronger randomness in the lesion's metabolic environment, and a faster rate at which the predicted state enters the chaotic region.

[0103] Perform time integration on the instantaneous information entropy generation rate to generate a cumulative information entropy value that monotonically increases with the prediction time length, thereby physically representing the total cumulative uncertainty of the system's perception of the patient's future condition.

[0104] According to the second law of thermodynamics, information entropy increases monotonically with time. The entropy-driven dynamic scheduling module performs time integration on the instantaneous information entropy generation rate to generate a cumulative information entropy value. The cumulative information entropy value physically represents the total accumulated uncertainty of the system's understanding of the patient's condition from the current moment to a certain predicted moment in the future, that is, the sum of the degree of blindness.

[0105] Furthermore, in the entropy-driven dynamic scheduling module, the specific operation of outputting the optimal review time point when the accumulated value of information entropy exceeds the preset safety confidence threshold is as follows: The current pathological state micro-vector is mapped to a malignant risk potential energy value that characterizes the degree of malignancy risk. A hyperbolic risk decay function is constructed, and the basic signal-to-noise ratio tolerance is dynamically adjusted using the malignant risk potential energy value to generate a risk-adaptive safety confidence threshold. This results in a higher malignant risk potential energy value and a lower risk-adaptive safety confidence threshold.

[0106] Clinical tolerance for uncertainty is inversely proportional to the patient’s current risk level. The entropy-driven dynamic scheduling module first maps the current pathological state micro-vector into a malignant risk potential value (normalized to the 0-1 range) through a fully connected layer.

[0107] A hyperbolic risk decay function is constructed, with the basic signal-to-noise ratio tolerance as the numerator and the sum of the risk sensitivity coefficient and the malignant risk potential energy value as the denominator. In this embodiment, the basic signal-to-noise ratio tolerance is set to 0.5-2.0Nat, which is determined based on the minimum bit requirement for state estimation of complex systems in information theory; the risk sensitivity coefficient is set to 2.0, which is chosen to ensure that when the malignant risk potential energy value reaches 0.5 (medium risk), the tolerance can be reduced to half of the baseline value.

[0108] The entropy-driven dynamic scheduling module uses a hyperbolic risk decay function to generate a risk-adaptive safety confidence threshold. Physically, when the potential energy of malignant risk increases, the denominator increases, resulting in a significant decrease in the risk-adaptive safety confidence threshold. This means that for high-risk patients, the system only allows a very small amount of cognitive uncertainty to accumulate.

[0109] The changes in the cumulative information entropy value are monitored, and the moment when the cumulative information entropy value first exceeds the risk-adaptive safety confidence threshold is determined. This moment is defined as the Lyapunov cognitive failure critical point.

[0110] The entropy-driven dynamic scheduling module monitors the growth of the cumulative information entropy value over the prediction time in real time. When the cumulative information entropy value exceeds the risk-adaptive safety confidence threshold for the first time, it indicates that the system's prediction of the future is no longer reliable and has entered the cognitive failure region. The entropy-driven dynamic scheduling module defines this moment as the Lyapunov cognitive failure critical point.

[0111] The maximum physical horizon is set as a fallback constraint. The Lyapunov cognitive failure threshold and the maximum physical horizon are compared, and the smaller value between the Lyapunov cognitive failure threshold and the maximum physical horizon is selected as the optimal review time.

[0112] To prevent unrealistically long follow-up intervals due to excessively slow scoring caused by extremely stable conditions, the entropy-driven dynamic scheduling module sets a maximum physical horizon as a fallback constraint. In this embodiment, the maximum physical horizon is set at 365-730 days according to the clinical guidelines for thyroid nodule management. The calculated Lyapunov cognitive failure threshold is compared with the maximum physical horizon, and the smaller value is selected as the optimal follow-up time. This mechanism ensures that the follow-up plan conforms to both the optimality of information thermodynamics and meets the bottom-line requirements of medical safety.

[0113] This embodiment resolves the contradiction between fixed periods and individual differences in traditional follow-up by constructing neural stochastic differential equations and a risk-adaptive entropy control mechanism. It accurately quantifies the cognitive dissipation process using instantaneous information entropy generation rate and achieves adaptive scheduling of high-risk-low-tolerance-short-cycle and low-risk-high-tolerance-long-cycle using a hyperbolic risk decay function. The optimal follow-up time is no longer a product of the doctor's experience, but a result of physical calculation based on the Lyapunov cognitive failure threshold, thus maximizing both the efficiency of medical resource utilization and the level of patient safety.

[0114] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A thyroid disease management and health data statistical analysis system, characterized in that, include: The system includes a multimodal image feature quantization module, a physiological dynamics temporal completion module, a time-delay causal tensor construction module, and an entropy-driven dynamic scheduling module. The multimodal image feature quantization module receives raw thyroid ultrasound image data, performs region of interest extraction on the raw thyroid ultrasound image data to obtain thyroid anatomical region data, performs pixel grayscale normalization processing on the thyroid anatomical region data, calculates texture feature parameters based on pixel grayscale distribution, calculates geometric morphology parameters based on thyroid anatomical region data, and performs vectorization concatenation processing on the texture feature parameters and geometric morphology parameters to generate morphological feature vectors. The physiological dynamics time-series completion module receives thyroid function biochemical test data and drug prescription record data. Using a kinetic model that includes drug metabolism and endogenous secretion terms, it calculates drug intake parameters based on drug prescription record data, calibrates metabolic rate parameters based on thyroid function biochemical test data, and performs interpolation operations between discrete thyroid function biochemical test data time points to generate continuous hormone concentration curves. The time-delay causal tensor construction module receives morphological feature vectors and continuous hormone concentration curves. It uses a preset non-uniform convolution kernel function to perform weighted convolution operations on the continuous hormone concentration curves to obtain hormone cumulative effect features. It then performs tensor concatenation operations on the hormone cumulative effect features and morphological feature vectors to generate a joint feature tensor. The entropy-driven dynamic scheduling module inputs the joint feature tensor into the time series prediction model to generate probability distribution data of the disease status at future time steps, calculates the information entropy value of the probability distribution data of the disease status, and outputs the optimal re-examination time point when the cumulative value of the information entropy value exceeds the preset safety confidence threshold.

2. The thyroid disease management and health data statistical analysis system according to claim 1, characterized in that, In the multimodal image feature quantization module, the thyroid anatomical region data undergoes pixel grayscale normalization processing. The specific operation for calculating texture feature parameters based on pixel grayscale distribution is as follows: The anisotropic diffusion equation for speckle denoising is used to process the thyroid anatomical region data. The diffusion coefficient function is called to calculate the gradient magnitude of the pixels in the thyroid anatomical region data. Based on the gradient magnitude, the pixels in the flat region and the pixels in the edge region are distinguished. The diffusion smoothing operation is performed on the pixels in the flat region and the diffusion stop operation is performed on the pixels in the edge region to generate pure scattering field data. The pure scattered field data is processed by calling the deep gain inversion algorithm. The physical compensation field data is constructed based on the preset average attenuation coefficient of thyroid tissue. The pure scattered field data and the physical compensation field data are multiplied point by point to generate the physical inversion sound field data that restores the true acoustic impedance characteristics of the tissue. A multi-scale Rayleigh entropy sliding window scan operation is performed on the physically inverted sound field data. The local background mean data of the physically inverted sound field data within the sliding window is calculated. The physically inverted sound field data within the sliding window is divided by the local background mean data to obtain normalized ratio data. The cube value of the normalized ratio data is calculated. The logarithm of the sum of the cube values ​​is calculated. The logarithm is used as the local acoustic entropy yield data and defined as the texture feature parameter.

3. The thyroid disease management and health data statistical analysis system according to claim 2, characterized in that, In the multimodal image feature quantization module, the geometric morphological parameters are calculated based on the thyroid anatomical region data. The specific operation of vectorizing and concatenating the texture feature parameters with the geometric morphological parameters to generate the morphological feature vector is as follows: The edge contour data of the thyroid anatomical region is extracted, and the edge contour data is smoothed and fitted using a cubic spline curve function to generate continuous and differentiable parameterized curve data. The scale-normalized bending energy data of continuously differentiable parameterized curve data is calculated based on the principle of minimum surface energy in continuum mechanics. The local curvature values ​​of the sampling points on the continuously differentiable parameterized curve data are calculated, the square values ​​of the local curvature values ​​are calculated, the integral operation is performed on the square values ​​along the closed path of the continuously differentiable parameterized curve data to obtain the integral result value, the physical perimeter value of the continuously differentiable parameterized curve data is calculated, the integral result value is multiplied by the physical perimeter value to obtain the scale-normalized bending energy data, and the scale-normalized bending energy data is defined as a geometric morphology parameter. Obtain preset demographic distribution parameter data, perform standard score standardization on texture feature parameters using demographic distribution parameter data, perform standard score standardization on geometric morphology parameters using demographic distribution parameter data, and perform orthogonal vector concatenation operation on the standardized texture feature parameters and standardized geometric morphology parameters to generate morphological feature vector.

4. The thyroid disease management and health data statistical analysis system according to claim 3, characterized in that, In the physiological dynamics time-series completion module, drug intake parameters are calculated based on drug prescription record data. The specific operation using a kinetic model that includes drug metabolism and endogenous secretion terms is as follows: A gastrointestinal absorption model based on the hydrodynamic hysteresis transport mechanism was constructed, and a random compliance variable following a Bernoulli distribution was generated. The random compliance variable was used to simulate the missed dose behavior and dose deviation behavior of patients during medication. An absorption delay kernel function was constructed using the gamma distribution probability density function. The absorption delay kernel function was used to simulate the drug dissolution process and transmembrane transport process in the gastrointestinal tract. The drug prescription record data, random compliance variable and absorption delay kernel function were convolved to generate exogenous drug entry flux rate data. The exogenous drug entry flux rate data was defined as the drug intake parameter. A coupled differential equation system incorporating the hypothalamic-pituitary-thyroid axis feedback mechanism was constructed. The endogenous secretion term was divided into a Michaelis-Menten kinetic channel controlled by thyroid-stimulating hormone (TSH) concentration and a linear gain channel controlled by autoantibody titers. Thyroid anatomical region volume data from a multimodal image feature quantification module was received. Using the thyroid anatomical region volume data as the basic energy production base, the nonlinear driving effect of TSH concentration data on the Michaelis-Menten kinetic channel and the anomalous driving effect of autoantibody titer data on the linear gain channel were calculated. The outputs of the Michaelis-Menten kinetic channel and the linear gain channel were superimposed to generate total endogenous secretion rate data.

5. The thyroid disease management and health data statistical analysis system according to claim 4, characterized in that, In the physiological dynamics time-series completion module, metabolic rate parameters are calibrated based on thyroid function biochemical test data. Interpolation is performed between discrete time points of thyroid function biochemical test data to generate a continuous hormone concentration curve. The specific operation is as follows: We acquire patients' liver and kidney function biochemical index data, construct a physiological elimination mapping function, use liver function biochemical index data to correct the weight of the liver deiodination metabolic pathway in the physiological elimination mapping function, use kidney function biochemical index data to correct the weight of the kidney excretion metabolic pathway in the physiological elimination mapping function, generate a dynamically changing individualized elimination rate constant, and define the individualized elimination rate constant as a metabolic rate parameter, thereby achieving heterogeneous adaptation for different liver and kidney function states. This study utilizes the unscented Kalman filter algorithm to process discrete thyroid function biochemical test data. Based on the law of conservation of mass, a state-space model is constructed. The exogenous drug influx rate data and total endogenous secretion rate data are used as the material inputs of the state-space model, while the product of the individualized elimination rate constant and the current hormone concentration is used as the material output. In the prediction step, coupled differential equations are used to deduce the free thyroxine concentration and thyroid-stimulating hormone concentration at unsampled times. In the update step, the Kalman gain is calculated using real thyroid function biochemical test data. The state vector and covariance matrix are corrected using the Kalman gain, and a continuous hormone concentration curve optimized for the entire time period is output.

6. The thyroid disease management and health data statistical analysis system according to claim 5, characterized in that, In the time-delay causal tensor construction module, the specific operation of performing weighted convolution operations on continuous hormone concentration curves using a preset non-uniform convolution kernel function to obtain the hormone cumulative effect characteristics is as follows: Extract the morphological feature vectors corresponding to two adjacent imaging examination times, calculate the geometric change rate data of the morphological feature vectors between two adjacent imaging examination times, determine the length value of the adaptive causal influence window based on the geometric change rate data, and use the length value of the adaptive causal influence window to extract the historical time series data before the corresponding examination time from the continuous hormone concentration curve. The preset skewed gamma distribution function is invoked, and a physical singularity protection constant term is introduced into the skewed gamma distribution function. The physical singularity protection constant term is configured to make the denominator of the skewed gamma distribution function non-zero when the backtracking time step is zero. A definite integral operation is performed on the skewed gamma distribution function after introducing the physical singularity protection constant term to obtain the total influence weight value. The total influence weight value is used to perform a division operation on the skewed gamma distribution function to generate a normalized biological response kernel function. The normalized biological response kernel function is applied to historical time series data and discrete weighted convolution is performed. The output of the discrete weighted convolution operation is defined as the hormone cumulative effect feature.

7. The thyroid disease management and health data statistical analysis system according to claim 6, characterized in that, In the time-delay causal tensor construction module, the specific operation of performing tensor concatenation between hormone cumulative effect features and morphological feature vectors to generate a joint feature tensor is as follows: The morphological feature vector is mapped to a preset spatial anatomical state manifold, and the hormone cumulative effect feature is mapped to a preset temporal physiological stress manifold. The Riemann manifold normalization algorithm is called, and the Riemann manifold normalization algorithm is used to perform centering and covariance normalization on the morphological feature vector in the spatial anatomical state manifold and the hormone cumulative effect feature in the temporal physiological stress manifold, respectively. The centering and covariance normalization processes eliminate the dimensional differences and numerical dynamic range differences between the morphological feature vector and the hormone cumulative effect feature, so that the morphological feature vector and the hormone cumulative effect feature have consistent statistical weight contributions in subsequent calculations. The morphological feature vectors after centering and covariance normalization are used as spatial dimension features, and the hormone cumulative effect features after centering and covariance normalization are used as temporal dimension features. Tensor dimension cascade operation is performed on the spatial dimension features and the temporal dimension features, and the coupled feature carrier generated by the tensor dimension cascade operation is defined as the joint feature tensor.

8. The thyroid disease management and health data statistical analysis system according to claim 7, characterized in that, In the entropy-driven dynamic scheduling module, the specific operation of inputting the joint feature tensor into the time series prediction model to generate the probability distribution data of the disease status at future time steps is as follows: A variational encoder is constructed, and the encoding network of the variational encoder is used to map the joint feature tensor to an initial Gaussian distribution in the latent state space. The initial pathological state micro-vector is obtained by sampling from the initial Gaussian distribution, thereby introducing the initial cognitive uncertainty of observation error and biological noise. A neural stochastic differential equation is constructed as a time series prediction model. In the neural stochastic differential equation, a pathological drift field and a biological diffusion tensor fitted by a neural network are defined. The pathological drift field is used to simulate the deterministic growth trend of lesions under the constraints of anatomical structure, and the biological diffusion tensor is used to simulate the intensity of random fluctuations of lesions under the influence of metabolic microenvironment. Starting with the initial pathological state micro-vector, the evolution trajectory of the pathological state micro-vector at future time steps is deduced using neural stochastic differential equations, generating probability distribution data of disease state that diffuses over time.

9. The thyroid disease management and health data statistical analysis system according to claim 8, characterized in that, In the entropy-driven dynamic scheduling module, the specific operation for calculating the information entropy value of the disease status probability distribution data is as follows: Based on the second-moment property of the Falk-Planck equation, the bio-diffusion tensor in the neural stochastic differential equation is extracted, the autocorrelation matrix of the bio-diffusion tensor is calculated, and the trace value of the autocorrelation matrix is ​​calculated. Half of the trace value is defined as the instantaneous information entropy generation rate. The instantaneous information entropy generation rate is used to quantify the information dissipation rate of the probability distribution data of the disease status over time. Perform time integration on the instantaneous information entropy generation rate to generate a cumulative information entropy value that monotonically increases with the prediction time length, thereby physically representing the total cumulative uncertainty of the system's perception of the patient's future condition.

10. The thyroid disease management and health data statistical analysis system according to claim 9, characterized in that, In the entropy-driven dynamic scheduling module, the specific operation of outputting the optimal review time point when the accumulated value of information entropy exceeds the preset security confidence threshold is as follows: The current pathological state micro-vector is mapped to a malignant risk potential energy value that characterizes the degree of malignancy risk. A hyperbolic risk decay function is constructed, and the basic signal-to-noise ratio tolerance is dynamically adjusted using the malignant risk potential energy value to generate a risk-adaptive safety confidence threshold. This results in a higher malignant risk potential energy value and a lower risk-adaptive safety confidence threshold. Monitor the changes in the cumulative information entropy value, determine the moment when the cumulative information entropy value first exceeds the risk-adaptive safety confidence threshold, and define this moment as the Lyapunov cognitive failure critical point. The maximum physical horizon is set as a fallback constraint. The Lyapunov cognitive failure threshold and the maximum physical horizon are compared, and the smaller value between the Lyapunov cognitive failure threshold and the maximum physical horizon is selected as the optimal review time.

Images