Intelligent management and risk warning analysis method and intelligent agent for the entire clinical trial cycle

Abstract

The present invention provides a clinical trial full-cycle intelligent management and risk warning analysis method and an intelligent agent, which relate to the technical field of clinical trial management, including collecting clinical trial data, extracting time-frequency features through nonlinear dynamic projection, spectral decomposition network, and wavelet Hilbert transform, and obtaining key feature signals using a dynamic compressed sensing model; generating a feature association network through manifold learning, constructing a nonlinear state migration field using stochastic differential equations, and capturing risk state mutation points in combination with an adaptive sampling strategy; obtaining a risk accumulation level through a dynamic prediction system, bifurcation theory, and singular spectrum analysis; based on the risk accumulation level, setting a dynamic warning threshold using critical state theory, and when the risk enters the threshold interval, starting a multi-objective optimization decision maker to calculate the optimal intervention strategy, performing risk regulation on the state migration field through a nonlinear response function, and feeding back the intervention results to iteratively optimize the risk warning.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical trial management, and in particular to a clinical trial full-cycle intelligent management and risk warning analysis method and an intelligent agent. Background Art

[0002] Clinical trials are a key link in the development of new drugs and the evaluation of medical devices. Their safety and effectiveness are directly related to the health of subjects and the success of research and development. As the scale and complexity of clinical trials increase, the multi-source heterogeneous data generated during the trial presents characteristics such as high dimension, nonlinearity and dynamic evolution.

[0003] However, there are some problems with the existing clinical trial risk management methods. Traditional risk assessment methods are difficult to effectively handle multi-source heterogeneous data in the clinical trial process and cannot fully explore the deep correlation between data; existing risk warning models are mostly based on static thresholds and simple statistical methods, and lack the ability to characterize the dynamic evolution of risks; the formulation of risk intervention strategies often relies on experience and judgment, and lacks a scientific optimization decision-making mechanism; the existing system generally has problems such as untimely information feedback and passive risk prevention and control, and cannot form a closed-loop risk management mechanism.

[0004] In summary, there is an urgent need for an intelligent risk warning method for the entire clinical trial cycle based on multi-objective optimization, which realizes the deep fusion of multi-source data through nonlinear dynamic projection and spectral decomposition network, extracts key feature information using time-frequency analysis and compressed sensing technology; constructs a state migration field based on stochastic differential equations, combines bifurcation theory and singular spectrum analysis to realize dynamic prediction of risk status; realizes intelligent early warning and active intervention of risks through multi-objective optimization decision-making and feedback compensation mechanism. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a clinical trial full-cycle intelligent management and risk warning analysis method and intelligent agent, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides a clinical trial full cycle intelligent management and risk warning analysis method, including:

[0008] Collect clinical trial data, map the clinical trial data to a high-dimensional feature space through nonlinear dynamic projection, extract time-frequency features based on a spectral decomposition network in the high-dimensional feature space through a composite algorithm based on wavelet transform and Hilbert transform, input the obtained time-frequency features into a dynamic compressed sensing model for sparse reconstruction, filter redundant information, obtain key feature signals, and generate a feature association network for the entire cycle of the clinical trial based on the key feature signals through a manifold learning algorithm;

[0009] Based on the characteristic association network, a nonlinear state migration field is constructed by stochastic differential equation calculation, and an adaptive sampling strategy is used to monitor the state migration field in real time to capture the mutation point information of the risk state. The mutation point information is input into the dynamic prediction system, and the risk evolution trend is calculated by combining bifurcation theory and singular spectrum analysis. Based on the risk evolution trend, the system state transition probability is calculated to obtain the risk accumulation level;

[0010] Based on the risk accumulation level, the dynamic warning threshold interval is set using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated through the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop, and the risk warning is iteratively optimized through adaptive filtering processing.

[0011] In an optional embodiment,

[0012] Mapping clinical trial data to a high-dimensional feature space through nonlinear dynamic projection includes:

[0013] Based on the kernel function, the feature mapping result of the kernel function is calculated, and the kernel function includes a radial basis kernel function, a polynomial kernel function and an exponential kernel function, an initial weight coefficient is assigned to each kernel function, the Mahalanobis distance between the sample points is calculated, a similarity matrix is ​​constructed, random matrix decomposition is performed, the kernel function parameters are updated and the weight coefficient is adjusted, and the weighted mapping results of the kernel function are combined to form an initial high-dimensional feature mapping matrix;

[0014] Calculate the conditional entropy value between each feature in the initial high-dimensional feature mapping matrix, determine the information redundancy, measure the mutual information coefficient between feature pairs, obtain the correlation strength, construct a feature importance scoring function based on the information redundancy and the correlation strength, calculate the importance score of each feature, construct an initial feature set, sort the initial feature set in descending order based on the importance score, select the feature with the highest importance score as the reference feature, calculate the similarity matrix between the reference feature and other features in the initial feature set, based on a preset similarity threshold, mark the feature whose similarity to the reference feature is greater than the similarity threshold as a redundant feature, remove the redundant feature from the initial feature set, obtain an updated feature set, recalculate the importance score of the updated feature set, select the feature with the highest importance score as the new reference feature, iterate and output a feature importance sequence;

[0015] Setting an importance threshold for the feature importance sequence, performing dimension screening, and executing a preset iterative projection algorithm to optimize and reconstruct the initial high-dimensional feature mapping matrix to obtain an optimal feature mapping matrix;

[0016] Inputting clinical trial data into the optimal feature mapping matrix for projection transformation, calculating the probability distribution characteristics of mapping errors, constructing an error compensation function based on a Gaussian process, performing gradient descent iteration to optimize compensation parameters, and generating an error compensation mapping matrix;

[0017] The error compensation mapping matrix and the optimal feature mapping matrix are combined in series to construct a nonlinear dynamic projection mapping model to achieve mapping conversion of the clinical trial data to a high-dimensional feature space.

[0018] In an optional embodiment,

[0019] Based on the key feature signals, a feature association network for the entire clinical trial cycle is generated through a manifold learning algorithm, including:

[0020] By calculating the statistical characteristics of the data sample, a distribution feature vector of each data point is obtained, based on the distribution feature vector, an adaptive function is constructed to determine the neighborhood range around each data point, a distance threshold is set based on the neighborhood range, and the Hadamard product metric distance between the data points is calculated. When the Hadamard product metric distance is less than the distance threshold, a connection relationship is established between the data points to generate a sparse adjacency matrix;

[0021] Based on the sparse adjacency matrix, the shortest path distance between data sample points is calculated through a sub-problem decomposition strategy to construct an initial geodesic distance matrix; a smoothing function is constructed using a piecewise cubic polynomial to perform smoothing on the initial geodesic distance matrix; abnormal distance values ​​are identified and corrected through an anti-interference regression method to obtain an optimized geodesic distance matrix;

[0022] Constructing a Laplacian operator based on the optimized distance matrix, performing eigenvalue decomposition on the Laplacian operator to obtain eigenvalues ​​and eigenvectors; selecting the eigenvector corresponding to the smallest non-zero eigenvalue as the basis of low-dimensional representation, and constructing a low-dimensional embedding space;

[0023] In the low-dimensional embedding space, the connection between each feature node and the corresponding other nodes is counted, the local connectivity is calculated, and the adaptive penalty function is determined according to the local connectivity. The adaptive penalty function is combined with the sparse constraint and the smooth constraint to construct the optimization target, and the network connection parameters are obtained. Based on the network connection parameters, the initial structure of the feature association network is constructed;

[0024] Calculate the probability overlap between feature distribution pairs to obtain distribution similarity, and convert the distribution similarity into a network update coefficient; iteratively adjust the connection strength and connection direction of the feature association network based on the network update coefficient; calculate the network structure difference between adjacent iterative steps, terminate the iteration when the network structure difference is less than a preset threshold, and output the final feature association network structure.

[0025] In an optional embodiment,

[0026] Based on the characteristic association network, constructing a nonlinear state migration field by stochastic differential equation calculation includes:

[0027] Performing variational mode decomposition on the topological structure of the feature association network, separating the dynamic features in the feature association network into multiple eigenmodes, and extracting a multi-frequency feature component matrix through spectral analysis;

[0028] Substituting the multi-frequency characteristic component matrix into the local linearization equation to calculate the state transfer gradient, constructing an adaptive drift term, constructing the multi-frequency characteristic component matrix as a random perturbation matrix to form a multi-scale diffusion term, and constructing a group of stochastic differential equations based on the adaptive drift term and the multi-scale diffusion term;

[0029] The stochastic differential equations are solved by using an adaptive variable step-size integration algorithm, a converged solution is obtained by dynamically adjusting the integration step-size and introducing a random disturbance term, and a system state trajectory sequence is generated;

[0030] According to the distribution of the system state trajectory sequence, a probability density model is constructed by using a hybrid kernel density estimation method, and the kernel function parameters of the probability density model are optimized by an iterative expectation algorithm to obtain the probability density distribution of the system state space;

[0031] Constructing a variational optimization objective function, wherein the variational optimization objective function includes a smoothness constraint term and a local structure constraint term, wherein the smoothness constraint term is measured using a second-order derivative norm, and the local structure constraint term is constructed based on a neighborhood relationship of the system state trajectory sequence;

[0032] The variational optimization objective function is solved by a splitting bidirectional optimization algorithm to form a nonlinear state migration field that maintains the local topological structure of the system state evolution.

[0033] In an optional embodiment,

[0034] Using the split bidirectional optimization algorithm to solve the variational optimization objective function includes:

[0035] Decomposing the variational optimization objective function into a smooth subproblem including a second-order derivative norm term of a probability density distribution and a non-smooth subproblem including a neighborhood structure constraint term of a system state trajectory sequence;

[0036] Calculating density gradients based on point density distribution of the system state trajectory sequence and constructing an adaptive weight matrix;

[0037] Calculate the second-order derivative of the probability density distribution, construct a momentum term based on the gradient direction of the previous iteration, substitute the momentum term into the gradient descent iterative formula, and solve the smooth subproblem;

[0038] Based on the system state trajectory sequence, a K nearest neighbor graph is constructed, the local topological feature vector of each state point is calculated, and the non-smooth sub-problem is iteratively solved through the nearest neighbor points;

[0039] According to the adaptive weight matrix, the step coefficients of the smooth subproblem and the non-smooth subproblem are calculated, the step coefficients are multiplied by the Lagrange multiplier, and the dual variables are updated;

[0040] Calculate the Euclidean distance between the solution of the smooth subproblem and the solution of the non-smooth subproblem to obtain the difference norm; calculate the difference between the current dual variable and the dual variable of the previous iteration to obtain the change of the dual variable; when the difference norm is less than a first preset threshold and the change of the dual variable is less than a second preset threshold, output the current solution as the optimal solution of the variational optimization objective function.

[0041] In an optional embodiment,

[0042] The mutation point information is input into the dynamics prediction system, and the risk evolution trend is calculated by combining bifurcation theory and singular spectrum analysis, including:

[0043] Constructing a feature matrix for the mutation point information, calculating the covariance of the feature matrix, obtaining a feature vector sequence, projecting the feature vector sequence into a dynamic space, constructing a bifurcation equation, and substituting the bifurcation equation into a prediction-correction iterator to generate a homotopic continuation path;

[0044] Identify the state mutation interval on the homotopy extension path, reconstruct the mutation point information into a trajectory matrix within the state mutation interval, perform singular value decomposition on the trajectory matrix, and extract left and right singular vectors whose singular values ​​are greater than a preset singular value threshold; use a kernel function to map the left and right singular vectors to a feature space, and construct a sparse optimization objective function with an L1 norm constraint based on the feature space; solve the sparse optimization objective function through an iterative soft threshold method to obtain a sparse representation coefficient of the state evolution mode;

[0045] According to the sparse representation coefficient, the homotopy extension path is divided into multiple time windows, and a state transition probability matrix is ​​constructed in each time window; based on the state transition probability matrix, a Gaussian correlation function is constructed to combine the marginal distribution of a single risk factor into a joint distribution probability;

[0046] Using the joint distribution probability, a random matrix sample set is generated, and the eigenvalue statistical distribution of the state transition probability matrix on the random matrix sample set is calculated;

[0047] According to the statistical distribution of the characteristic values, the state transition paths whose significance coefficients are greater than a preset significance coefficient threshold are screened, and the probability values ​​are accumulated along the state transition paths to obtain the risk evolution trend.

[0048] In an optional embodiment,

[0049] Based on the risk accumulation level, the dynamic warning threshold interval is set using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated by the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop. Through adaptive filtering processing, iterative optimization of risk warning includes:

[0050] Based on the cumulative level of system risk, the maximum eigenvalue of the state transition probability matrix is ​​calculated using the critical state theory, and the deviation range between the maximum eigenvalue and the preset characteristic threshold is determined as the upper and lower boundaries of the dynamic warning interval;

[0051] Constructing a multi-objective optimization model including a risk control objective function and an intervention cost objective function, wherein the risk control objective function and the intervention cost objective function are combined by a weight coefficient;

[0052] The constraints of the multi-objective optimization model are relaxed, Lagrangian multipliers are introduced to construct an augmented Lagrangian function, the augmented Lagrangian function is decomposed into quadratic penalty terms to obtain a sub-problem sequence, and based on the gradient information of the sub-problem sequence, the Lagrangian relaxation iteration method is used to alternately update the intervention strategy vector and the Lagrangian multiplier, and the penalty factor is adaptively adjusted to obtain the optimal intervention strategy;

[0053] Constructing a monotonically increasing nonlinear response function, mapping the optimal intervention strategy to a state transition field through the nonlinear response function, and obtaining a state transition probability matrix after intervention;

[0054] Based on the upper and lower boundaries of the dynamic warning interval and the system stability criterion, an expected state matrix is ​​constructed, a difference operation is performed between the state transition probability matrix after intervention and the expected state matrix to obtain an error matrix, and a feedback compensation loop is constructed based on the error matrix;

[0055] The feedback compensation loop is processed by adaptive Kalman filtering, and the error matrix is ​​iteratively corrected by adjusting the gain matrix until the error matrix converges to a preset accuracy range, thereby completing the dynamic optimization of risk warning.

[0056] According to a second aspect of the embodiments of the present invention,

[0057] Provides an intelligent management and risk warning analysis agent for the entire clinical trial cycle, including:

[0058] The first unit is used to collect clinical trial data, map the clinical trial data to a high-dimensional feature space through nonlinear dynamic projection, extract time-frequency features based on a spectral decomposition network in the high-dimensional feature space through a composite algorithm based on wavelet transform and Hilbert transform, input the obtained time-frequency features into a dynamic compressed sensing model for sparse reconstruction, filter redundant information, obtain key feature signals, and generate a feature association network for the entire clinical trial cycle based on the key feature signals through a manifold learning algorithm;

[0059] The second unit is used to construct a nonlinear state migration field based on the characteristic association network by stochastic differential equation calculation, monitor the state migration field in real time by using an adaptive sampling strategy, capture the mutation point information of the risk state, input the mutation point information into the dynamic prediction system, calculate the risk evolution trend by combining bifurcation theory and singular spectrum analysis, calculate the system state transition probability based on the risk evolution trend, and obtain the risk accumulation level;

[0060] The third unit is used to set the dynamic warning threshold interval according to the risk accumulation level by using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated by the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop, and the risk warning is iteratively optimized through adaptive filtering processing.

[0061] According to a third aspect of the embodiments of the present invention,

[0062] An electronic device is provided, comprising:

[0063] processor;

[0064] a memory for storing processor-executable instructions;

[0065] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0066] According to a fourth aspect of the embodiments of the present invention,

[0067] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0068] In the embodiment of the present invention, through nonlinear dynamic projection, spectral decomposition network, wavelet transform and Hilbert transform and other advanced signal processing technologies, combined with dynamic compressed sensing and manifold learning, it is possible to effectively extract key feature signals from complex clinical trial data and construct a full-cycle feature association network. This avoids the subjectivity and inefficiency of manual data analysis by traditional methods, realizes automated and intelligent data analysis, and improves the accuracy and efficiency of data analysis; based on the feature association network, using methods such as stochastic differential equations, bifurcation theory and singular spectrum analysis, the risk evolution trend of clinical trials can be dynamically monitored and predicted. The mutation point of the risk state is captured by an adaptive sampling strategy, and the dynamic warning threshold is set in combination with the critical state theory, so as to realize the early warning of the risk. Furthermore, through multi-objective optimization decision makers and nonlinear response functions, active intervention and regulation can be carried out for the risk situation, effectively reducing the risk of clinical trials; through early warning and active intervention of risks, potential problems can be avoided, unnecessary trial interruptions and repetitions can be reduced, thereby optimizing resource allocation, shortening the clinical trial cycle, and ultimately reducing the cost of clinical trials. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A schematic diagram of the process of the clinical trial full cycle intelligent management and risk early warning analysis method according to an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the structure of the clinical trial full-cycle intelligent management and risk warning analysis intelligent agent in an embodiment of the present invention. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0073] Figure 1 FIG. 1 is a flow chart of a method for intelligent management and risk early warning analysis of a clinical trial throughout a whole cycle according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0074] S101. Collect clinical trial data, map the clinical trial data to a high-dimensional feature space through nonlinear dynamic projection, extract time-frequency features based on a spectral decomposition network in the high-dimensional feature space through a composite algorithm based on wavelet transform and Hilbert transform, input the obtained time-frequency features into a dynamic compressed sensing model for sparse reconstruction, filter redundant information, obtain key feature signals, and generate a feature association network for the entire clinical trial cycle based on the key feature signals through a manifold learning algorithm;

[0075] First, the original data from the clinical trial process is collected, including time series data such as physiological signals and biochemical indicators. The collected data is preprocessed, including denoising and standardization, to ensure data quality.

[0076] The preprocessed data is mapped to a high-dimensional feature space through a kernel function. This process transforms the nonlinear relationship in the original space into a linear relationship in the high-dimensional feature space by constructing a radial basis kernel function. The mapping process maintains the topological structure of the data while enhancing the separability of the features.

[0077] In the high-dimensional feature space, an adjacency matrix is ​​constructed based on the similarity relationship between data points. By performing spectral decomposition on the adjacency matrix, eigenvectors and eigenvalues ​​are obtained. These eigenvectors constitute a spectral decomposition network that describes the intrinsic structure of the data.

[0078] First, the signal in the spectral decomposition network is subjected to wavelet transform to decompose the signal into wavelet coefficients of different scales, capturing the characteristics of the signal at different time and frequency scales. Then, the wavelet coefficients are subjected to Hilbert transform to obtain the instantaneous frequency and amplitude information of the signal. The results of wavelet transform and Hilbert transform are fused to form a complete description of time-frequency characteristics.

[0079] A dynamic compressed sensing model is constructed, which includes a measurement matrix and a reconstruction operator. The time-frequency features are input into the model and sparse reconstruction is performed through an iterative projection algorithm. During the reconstruction process, the model automatically identifies and retains important feature components while filtering out redundant and noise information, and finally obtains sparse key feature signals.

[0080] Based on the key feature signals obtained, the manifold learning algorithm is applied to reconstruct the intrinsic geometric structure of the data. First, the local adjacency relationship is constructed and the geodesic distance between data points is calculated. Then, the high-dimensional data is mapped to the low-dimensional manifold space by maintaining the constraints of the local structure. In the manifold space, the conditional mutual information between features is calculated, the association strength between feature nodes is established, and finally the feature association network of the entire clinical trial cycle is generated.

[0081] In this embodiment, through nonlinear dynamic projection and high-dimensional feature space mapping, the complex nonlinear relationships in clinical trial data can be more comprehensively captured, and the accuracy of feature extraction can be improved; the composite algorithm of wavelet transform and Hilbert transform can be combined to effectively extract time-frequency features, capture local and global feature information in the signal, and enhance the time-frequency expression ability of features; the dynamic compressed sensing model can remove redundant information, highlight key feature signals, and improve the computational efficiency and reliability of subsequent analysis; the manifold learning algorithm is used to generate a full-cycle feature association network, which can reveal the potential correlation between complex features and support in-depth analysis and insight into the entire cycle of clinical trials; through the above-mentioned technical combination, the key signal extraction capability is comprehensively improved, providing strong support for the comprehensive analysis of clinical trial data, and helping to discover new medical laws or potential mechanisms.

[0082] S102. Based on the characteristic association network, a nonlinear state migration field is constructed by stochastic differential equation calculation, and the state migration field is monitored in real time by using an adaptive sampling strategy to capture the mutation point information of the risk state, and the mutation point information is input into the dynamic prediction system, and the risk evolution trend is calculated by combining bifurcation theory and singular spectrum analysis. Based on the risk evolution trend, the system state transition probability is calculated to obtain the risk accumulation level;

[0083] Based on the obtained characteristic correlation network, a set of stochastic differential equations describing the dynamic evolution of the system is first established. This set of equations contains deterministic drift terms and random diffusion terms, where the drift term describes the main evolution trend of the system and the diffusion term describes the influence of random factors on the system. By numerically solving this set of equations, the trajectory distribution of the system in the characteristic space is obtained, thereby constructing a nonlinear state migration field.

[0084] The constructed nonlinear state migration field is monitored in real time, and an adaptive sampling strategy is adopted. This strategy dynamically adjusts the sampling frequency according to the rate of change of the system state: a lower sampling frequency is used when the system state is relatively stable, and a higher sampling frequency is used when the state changes rapidly. By calculating the state difference between adjacent sampling points and setting a dynamic threshold, a potential mutation point is identified when the difference exceeds the threshold. The identified potential mutation points are locally refined and sampled to confirm the real state mutation point.

[0085] The captured mutation point information is input into the dynamic prediction system. The system first establishes the state space reconstruction and determines the embedding dimension and time delay of the system. Based on the reconstructed state space, the dynamic behavior of the system under different parameters is analyzed, including stable points, limit cycles and chaotic attractors.

[0086] In the dynamic prediction system, bifurcation theory is used to analyze the stability changes of the system when the parameters change. By calculating the characteristic index of the system, the possible bifurcation types of the system are identified, including saddle-node bifurcation, Hopf bifurcation, etc. At the same time, singular spectrum analysis is performed to decompose the time series into several principal components and extract the dominant dynamic mode of the system. By analyzing the time evolution characteristics of these modes, the risk evolution trend of the system is predicted.

[0087] Based on the predicted risk evolution trend, a Markov chain model is constructed. The possible states of the system are divided into several regions, and the state transition probability matrix is ​​established by statistically analyzing the transition frequency between different states of the system. This matrix describes the possibility of the system transitioning from the current state to other states.

[0088] The state transition probability matrix is ​​used to calculate the steady-state distribution of the system in each state. Combined with the risk weights corresponding to each state, the system's risk accumulation level is calculated comprehensively. This risk accumulation level reflects the current overall risk status of the system and provides a basis for subsequent risk warnings and interventions.

[0089] In this embodiment, a nonlinear state migration field is constructed by stochastic differential equations, and an adaptive sampling strategy is used for real-time monitoring, so as to effectively identify the mutation point information of the risk state in the system and improve the sensitivity and accuracy of risk identification. By combining bifurcation theory and singular spectrum analysis, the risk evolution trend is dynamically calculated to provide deep insights into the future state changes of the system and enhance the foresight of risk management. By calculating the system state transition probability and the risk accumulation level, an accurate risk quantification tool is provided to help assess the severity of potential risks in the system. The assessment results based on the risk accumulation level can provide reliable data support for early warning and decision-making intervention, and reduce the threat of potential risks to system stability. The combination of nonlinear state migration field and adaptive sampling strategy can adapt to the risk change characteristics in complex dynamic environments and improve the applicability and robustness of risk analysis.

[0090] S103. Based on the risk accumulation level, the dynamic warning threshold interval is set using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated through the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop, and the risk warning is iteratively optimized through adaptive filtering processing.

[0091] In this embodiment, a dynamic warning threshold interval is set based on the critical state theory, which can timely identify abnormal changes in the risk accumulation level and realize accurate and real-time risk warning; the optimal intervention strategy is calculated through a multi-objective optimization decision maker and a Lagrangian relaxation algorithm to provide a scientific and reasonable intervention plan for risk control; the nonlinear response function is used to act on the state migration field to dynamically control the risk situation, significantly reduce the risk accumulation level, and ensure system stability; the intervention results are input through a feedback compensation loop, and combined with adaptive filtering processing to realize dynamic iterative optimization of the risk warning system and improve the intervention effect and system response efficiency; the entire process combines dynamic adjustment and feedback optimization mechanisms to improve the flexibility and robustness of risk control and adapt to the dynamic evolution of risks in complex environments.

[0092] In an optional embodiment, mapping the clinical trial data to a high-dimensional feature space by nonlinear dynamic projection includes:

[0093] Based on the kernel function, the feature mapping result of the kernel function is calculated, and the kernel function includes a radial basis kernel function, a polynomial kernel function and an exponential kernel function, an initial weight coefficient is assigned to each kernel function, the Mahalanobis distance between the sample points is calculated, a similarity matrix is ​​constructed, random matrix decomposition is performed, the kernel function parameters are updated and the weight coefficient is adjusted, and the weighted mapping results of the kernel function are combined to form an initial high-dimensional feature mapping matrix;

[0094] Calculate the conditional entropy value between each feature in the initial high-dimensional feature mapping matrix, determine the information redundancy, measure the mutual information coefficient between feature pairs, obtain the correlation strength, construct a feature importance scoring function based on the information redundancy and the correlation strength, calculate the importance score of each feature, construct an initial feature set, sort the initial feature set in descending order based on the importance score, select the feature with the highest importance score as the reference feature, calculate the similarity matrix between the reference feature and other features in the initial feature set, based on a preset similarity threshold, mark the feature whose similarity to the reference feature is greater than the similarity threshold as a redundant feature, remove the redundant feature from the initial feature set, obtain an updated feature set, recalculate the importance score of the updated feature set, select the feature with the highest importance score as the new reference feature, iterate and output a feature importance sequence;

[0095] Setting an importance threshold for the feature importance sequence, performing dimension screening, and executing a preset iterative projection algorithm to optimize and reconstruct the initial high-dimensional feature mapping matrix to obtain an optimal feature mapping matrix;

[0096] Inputting clinical trial data into the optimal feature mapping matrix for projection transformation, calculating the probability distribution characteristics of mapping errors, constructing an error compensation function based on a Gaussian process, performing gradient descent iteration to optimize compensation parameters, and generating an error compensation mapping matrix;

[0097] The error compensation mapping matrix and the optimal feature mapping matrix are combined in series to construct a nonlinear dynamic projection mapping model to achieve mapping conversion of the clinical trial data to a high-dimensional feature space.

[0098] The iterative projection algorithm is a type of algorithm used to optimize high-dimensional data mapping or feature reconstruction, which gradually approaches the optimal result by iteratively adjusting the mapping or projection operation. It is often used in scenarios such as dimensionality reduction, feature extraction, and data optimization; the usual goal is to project high-dimensional data into a low-dimensional space or an optimized feature space while retaining important information of the original data; the general steps are initial projection, generating an initial projection matrix according to the set mapping rules; error calculation, calculating the error of the current projection matrix or optimizing the target value (such as data reconstruction error); parameter update, adjusting the parameters of the projection matrix, reducing the error or optimizing the target value, and commonly using optimization methods such as gradient descent; convergence judgment, repeating the above process until the error is less than the preset threshold or the maximum number of iterations is reached.

[0099] The nonlinear dynamic projection mapping method for clinical trial data aims to map clinical trial data into a high-dimensional feature space in order to better mine the hidden information in the data, thereby improving the accuracy and efficiency of data analysis.

[0100] First, select a suitable kernel function, such as the radial basis kernel function, the polynomial kernel function, and the exponential kernel function. Assign an initial weight coefficient, such as an equal weight, to each kernel function. Then, calculate the Mahalanobis distance between the sample points to measure the similarity between the sample points. Based on the Mahalanobis distance, construct a similarity matrix. Next, perform random matrix decomposition on the similarity matrix to extract the main feature information. According to the decomposition results, update the parameters of the kernel function and adjust the weight coefficient, for example, increase the weight of the kernel function that contributes more to the classification. Finally, combine the weighted mapping results of different kernel functions to form an initial high-dimensional feature mapping matrix. For example, assuming there are three sample points, after the above steps, an initial high-dimensional feature mapping matrix of 3×10 is obtained.

[0101] Next, the conditional entropy values ​​between the features in the initial high-dimensional feature mapping matrix are calculated to determine the information redundancy. At the same time, the mutual information coefficients between the feature pairs are measured to obtain the correlation strength. For example, the mutual information coefficient between feature 1 and feature 2 is 0.8, indicating that there is a strong correlation between them. Based on the information redundancy and the correlation strength, a feature importance scoring function is constructed, for example, a linear combination of information redundancy and correlation strength is used as a scoring function. Then, the importance score of each feature is calculated, and an initial feature set is constructed. For example, the importance score of feature 1 is 0.9, and the importance score of feature 2 is 0.85. The initial feature set is sorted in descending order, and the feature with the highest importance score is selected as the reference feature. For example, feature 1 has the highest importance score, so it is selected as the reference feature. The similarity matrix between the reference feature and other features in the initial feature set is calculated. Based on a preset similarity threshold, such as 0.7, features whose similarity to the reference feature is greater than the threshold are marked as redundant features. Redundant features are removed from the initial feature set to obtain an updated feature set. For example, the similarity between feature 2 and feature 1 is 0.75, which is greater than the threshold of 0.7, so feature 2 is marked as a redundant feature and removed. The importance score of the updated feature set is recalculated, and the feature with the highest importance score is selected as the new reference feature. The above steps are iterated until the updated feature set no longer changes or the preset number of iterations is reached, and finally the feature importance sequence is output.

[0102] Set importance thresholds for feature importance sequences to perform dimension screening. For example, select the top 5 features with the highest importance scores. Then, execute a preset iterative projection algorithm, such as the t-SNE algorithm, to optimize and reconstruct the initial high-dimensional feature mapping matrix to obtain the optimal feature mapping matrix. For example, optimize and reconstruct the initial high-dimensional feature mapping matrix of 3×10 into the optimal feature mapping matrix of 3×5.

[0103] Input the clinical trial data into the optimal feature mapping matrix for projection transformation. Calculate the probability distribution characteristics of the mapping error, for example, calculate the mean and variance of the error. Construct an error compensation function based on the Gaussian process, and perform gradient descent iteration to optimize the compensation parameters. For example, optimize the hyperparameters of the Gaussian process by minimizing the mean square error of the error. Finally, generate the error compensation mapping matrix.

[0104] The error compensation mapping matrix is ​​combined in series with the optimal feature mapping matrix to construct a nonlinear dynamic projection mapping model. For example, the 3×5 optimal feature mapping matrix and the 3×2 error compensation mapping matrix are combined in series to form a 3×7 nonlinear dynamic projection mapping model. Ultimately, the mapping transformation of clinical trial data to a high-dimensional feature space is achieved.

[0105] In this embodiment, nonlinear dynamic projection can be used to better capture nonlinear relationships and dynamic changes in clinical trial data, thereby improving the accuracy of data analysis; feature importance sorting and dimension screening can be used to identify key features that have the greatest impact on clinical trial results, thereby enhancing data interpretability; and optimized reconstruction and error compensation can be used to reduce data dimensions and computational complexity, thereby improving data processing efficiency.

[0106] In an optional implementation, based on the key feature signal, generating a feature association network for the entire clinical trial cycle through a manifold learning algorithm includes:

[0107] By calculating the statistical characteristics of the data sample, a distribution feature vector of each data point is obtained, based on the distribution feature vector, an adaptive function is constructed to determine the neighborhood range around each data point, a distance threshold is set based on the neighborhood range, and the Hadamard product metric distance between the data points is calculated. When the Hadamard product metric distance is less than the distance threshold, a connection relationship is established between the data points to generate a sparse adjacency matrix;

[0108] Based on the sparse adjacency matrix, the shortest path distance between data sample points is calculated through a sub-problem decomposition strategy to construct an initial geodesic distance matrix; a smoothing function is constructed using a piecewise cubic polynomial to perform smoothing on the initial geodesic distance matrix; abnormal distance values ​​are identified and corrected through an anti-interference regression method to obtain an optimized geodesic distance matrix;

[0109] Constructing a Laplacian operator based on the optimized distance matrix, performing eigenvalue decomposition on the Laplacian operator to obtain eigenvalues ​​and eigenvectors; selecting the eigenvector corresponding to the smallest non-zero eigenvalue as the basis of low-dimensional representation, and constructing a low-dimensional embedding space;

[0110] In the low-dimensional embedding space, the connection between each feature node and the corresponding other nodes is counted, the local connectivity is calculated, and the adaptive penalty function is determined according to the local connectivity. The adaptive penalty function is combined with the sparse constraint and the smooth constraint to construct the optimization target, and the network connection parameters are obtained. Based on the network connection parameters, the initial structure of the feature association network is constructed;

[0111] Calculate the probability overlap between feature distribution pairs to obtain distribution similarity, and convert the distribution similarity into a network update coefficient; iteratively adjust the connection strength and connection direction of the feature association network based on the network update coefficient; calculate the network structure difference between adjacent iterative steps, terminate the iteration when the network structure difference is less than a preset threshold, and output the final feature association network structure.

[0112] First, the distribution characteristics of the data sample are obtained. For example, for each subject participating in a clinical trial, a series of indicator data are collected, such as blood pressure, heart rate, blood sugar level, etc. Then, the statistical characteristics of these indicator data, such as mean, variance, skewness, etc., are calculated, and these statistical characteristics are combined into a vector as the distribution characteristic vector of the subject.

[0113] Next, determine the neighborhood range around each data point. Use an adaptive function to dynamically adjust the neighborhood range based on the distribution feature vector of the data point. For example, for data points with large differences in distribution feature vectors, set a larger neighborhood range; conversely, set a smaller neighborhood range. Set a distance threshold, such as the average distance or median distance as the threshold.

[0114] Then, the distance between the data points is calculated. The Hadamard product metric is used to calculate the distance between the data points. When the Hadamard product metric distance between two data points is less than the preset distance threshold, it is considered that there is a connection relationship between the two data points, and the corresponding position in the sparse adjacency matrix is ​​marked as 1, otherwise it is marked as 0.

[0115] Based on the generated sparse adjacency matrix, the shortest path distance between the data sample points is calculated. The sub-problem decomposition strategy is adopted to decompose the complex full-sample shortest path calculation into multiple small-scale sub-problems, calculate the shortest path in each sub-problem separately, and then merge the solutions of the sub-problems to obtain the final full-sample shortest path distance matrix, that is, the initial geodesic distance matrix.

[0116] Smoothing the initial geodesic distance matrix. Using piecewise cubic polynomials to construct a smoothing function, each element in the initial geodesic distance matrix is ​​smoothed to reduce the impact of noise and outliers.

[0117] Identify and correct abnormal distance values. Use anti-interference regression method to identify abnormal distance values. By fitting the distance distribution of data points, identify abnormal values ​​that deviate from the overall trend and correct them to more reasonable values. Obtain the optimized geodesic distance matrix.

[0118] The Laplacian operator is constructed based on the optimized distance matrix, and the eigenvalue decomposition is performed to obtain the eigenvalues ​​and eigenvectors. The eigenvector corresponding to the smallest non-zero eigenvalue is selected as the basis of the low-dimensional representation to construct the low-dimensional embedding space.

[0119] In the low-dimensional embedding space, a feature association network is constructed. First, the connection between each feature node and other nodes is counted to calculate the local connectivity. According to the local connectivity, an adaptive penalty function is determined. The adaptive penalty function is combined with sparse constraints and smooth constraints to construct an optimization target and solve to obtain the network connection parameters. Based on the network connection parameters, the initial structure of the feature association network is constructed.

[0120] Iteratively adjust the network structure. Calculate the probability overlap between feature distribution pairs to obtain distribution similarity. Convert the distribution similarity into a network update coefficient. Iteratively adjust the connection strength and connection direction of the feature association network based on the network update coefficient. Calculate the network structure difference between adjacent iteration steps. When the network structure difference is less than the preset threshold, terminate the iteration and output the final feature association network structure.

[0121] For example, suppose there are blood pressure, heart rate and blood sugar data for 100 subjects. First, the mean and variance of these three indicators for each subject are calculated to obtain a 6-dimensional distribution eigenvector. Then, the distance between the data points is calculated based on the distribution eigenvector, and a sparse adjacency matrix is ​​constructed. Assuming that the distance threshold is set to 0.5, if the Hadamard product metric distance between two subjects is less than 0.5, the corresponding position in the adjacency matrix is ​​marked as 1. Next, the shortest path distance is calculated based on the adjacency matrix, and smoothing and outlier correction are performed. Then, the Laplace operator is constructed and eigenvalue decomposition is performed, and the eigenvector corresponding to the smallest non-zero eigenvalue is selected to construct a low-dimensional embedding space. Finally, a feature association network is constructed in the low-dimensional space, and the network structure is iteratively adjusted according to the feature distribution similarity until the network structure difference is less than the preset threshold (for example, 0.01).

[0122] In this embodiment, the manifold learning algorithm can effectively capture the nonlinear relationship in the clinical trial data, so as to more accurately construct the feature association network and reveal the real association between features; the generated feature association network shows the association between features in a graphical way, which is intuitive and easy to understand, and helps researchers to better understand the inherent laws of the test data; the use of adaptive functions, smoothing processing and outlier correction techniques enhances the robustness of the method, enabling it to effectively cope with the influence of noise and abnormal data, and improves the reliability of the results.

[0123] In an optional implementation, based on the feature association network, constructing a nonlinear state transition field by stochastic differential equation calculation includes:

[0124] Performing variational mode decomposition on the topological structure of the feature association network, separating the dynamic features in the feature association network into multiple eigenmodes, and extracting a multi-frequency feature component matrix through spectral analysis;

[0125] Substituting the multi-frequency characteristic component matrix into the local linearization equation to calculate the state transfer gradient, constructing an adaptive drift term, constructing the multi-frequency characteristic component matrix as a random perturbation matrix to form a multi-scale diffusion term, and constructing a group of stochastic differential equations based on the adaptive drift term and the multi-scale diffusion term;

[0126] The stochastic differential equations are solved by using an adaptive variable step-size integration algorithm, a converged solution is obtained by dynamically adjusting the integration step-size and introducing a random disturbance term, and a system state trajectory sequence is generated;

[0127] According to the distribution of the system state trajectory sequence, a probability density model is constructed by using a hybrid kernel density estimation method, and the kernel function parameters of the probability density model are optimized by an iterative expectation algorithm to obtain the probability density distribution of the system state space;

[0128] Constructing a variational optimization objective function, wherein the variational optimization objective function includes a smoothness constraint term and a local structure constraint term, wherein the smoothness constraint term is measured using a second-order derivative norm, and the local structure constraint term is constructed based on a neighborhood relationship of the system state trajectory sequence;

[0129] The variational optimization objective function is solved by a splitting bidirectional optimization algorithm to form a nonlinear state migration field that maintains the local topological structure of the system state evolution.

[0130] Feature-related networks, such as social networks, biological networks, or sensor networks, contain rich dynamic information. In order to deeply understand the dynamic information and predict the evolution trend of the network, it is necessary to construct a nonlinear state transition field that can capture the law of network state changes.

[0131] First, the topology of the feature association network is subjected to variational mode decomposition. This is similar to decomposing a complex sound into different notes. Through this decomposition, the dynamic characteristics of the network can be separated into multiple independent eigenmodes, each of which represents an aspect of the dynamic changes of the network. For example, in a social network, one mode may represent the change of user interests, and another mode may represent the speed of information propagation. Then, through spectral analysis, multi-frequency feature component matrices are extracted from each eigenmode. These matrices capture the changing characteristics of each mode at different time scales. Assuming a social network has 100 users, the feature association network can be represented as a 100×100 matrix. After variational mode decomposition and spectral analysis, multiple multi-frequency feature component matrices of, for example, 100×100 can be obtained, each of which represents a specific frequency component.

[0132] Next, these multi-frequency eigencomponent matrices are used to construct stochastic differential equations. The multi-frequency eigencomponent matrices are substituted into the local linearization equations to calculate the state transfer gradient, which can be understood as calculating the direction and speed of the network state change. Then, based on the calculated gradient, an adaptive drift term is constructed. The drift term determines the overall trend of the network state evolution. At the same time, the multi-frequency eigencomponent matrix is ​​constructed as a random perturbation matrix to form a multi-scale diffusion term. The diffusion term simulates the random fluctuations in the evolution of the network state. By combining the adaptive drift term and the multi-scale diffusion term, a system of stochastic differential equations can be established. For example, if 3 main modes are extracted, a system of stochastic differential equations containing 3 equations will be formed.

[0133] Then, the adaptive variable step-size integration algorithm is used to solve the stochastic differential equations. This is like calculating the future path step by step according to a map and a compass. The adaptive variable step-size integration algorithm dynamically adjusts the integration step size according to the speed of state changes, and introduces random perturbations at each step to simulate the uncertainty in the real world. By solving the stochastic differential equations, a series of system state trajectory sequences can be obtained, which represents the possible evolution path of the network state in the future. For example, 1,000 trajectory sequences can be obtained, each of which represents the possible changes of the network state in the next 10 time steps.

[0134] According to the distribution of the obtained system state trajectory sequence, the probability density model is constructed using the hybrid kernel density estimation method. This is similar to drawing a smooth probability distribution graph based on known sample points. The hybrid kernel density estimation method selects appropriate kernel functions and parameters based on the distribution of the trajectory sequence to better fit the distribution of the trajectory sequence. Then, the kernel function parameters of the probability density model are optimized through the iterative expectation algorithm, and finally the probability density distribution of the system state space is obtained. For example, a model describing the probability distribution of the network state in the next 10 time steps can be obtained.

[0135] Finally, in order to construct the nonlinear state migration field, a variational optimization objective function needs to be constructed. This objective function contains two items: smoothness constraint and local structure constraint. The smoothness constraint ensures the smoothness of the state migration field and avoids drastic jumps. It can be measured using the second-order derivative norm, similar to evaluating the smoothness of a surface. The local structure constraint ensures that the state migration field can maintain the local topological structure of the system state evolution, for example, maintaining the connection relationship between nodes in the network. It can be constructed based on the neighborhood relationship of the system state trajectory sequence. Then, the split bidirectional optimization algorithm is used to solve the variational optimization objective function. This is similar to repeatedly adjusting the shape of a sculpture to make it smooth and meet the design requirements. Finally, a nonlinear state migration field that maintains the local topological structure of the system state evolution can be obtained.

[0136] In this embodiment, by combining variational mode decomposition and stochastic differential equations, the complex nonlinear dynamic characteristics in the feature association network can be captured more accurately, thereby more accurately predicting the evolution trend of the network state. Compared with the traditional linear model-based method, this method can better handle the nonlinear relationships and random disturbances in the network; the application of the adaptive variable step-size integral algorithm and the hybrid kernel density estimation method makes this method more robust to noise and abnormal data. Even in the case of poor data quality, relatively stable prediction results can still be obtained; through variational mode decomposition, the network dynamics are decomposed into multiple independent eigenmodes, making the output of the model easier to interpret. The contribution of each mode to the evolution of the network can be analyzed, so as to deeply understand the internal mechanism of the dynamic changes of the network.

[0137] In an optional implementation, solving the variational optimization objective function using a split bidirectional optimization algorithm includes:

[0138] Decomposing the variational optimization objective function into a smooth subproblem including a second-order derivative norm term of a probability density distribution and a non-smooth subproblem including a neighborhood structure constraint term of a system state trajectory sequence;

[0139] Calculating density gradients based on point density distribution of the system state trajectory sequence and constructing an adaptive weight matrix;

[0140] Calculate the second-order derivative of the probability density distribution, construct a momentum term based on the gradient direction of the previous iteration, substitute the momentum term into the gradient descent iterative formula, and solve the smooth subproblem;

[0141] Based on the system state trajectory sequence, a K nearest neighbor graph is constructed, the local topological feature vector of each state point is calculated, and the non-smooth sub-problem is iteratively solved through the nearest neighbor points;

[0142] According to the adaptive weight matrix, the step coefficients of the smooth subproblem and the non-smooth subproblem are calculated, the step coefficients are multiplied by the Lagrange multiplier, and the dual variables are updated;

[0143] Calculate the Euclidean distance between the solution of the smooth subproblem and the solution of the non-smooth subproblem to obtain the difference norm; calculate the difference between the current dual variable and the dual variable of the previous iteration to obtain the change of the dual variable; when the difference norm is less than a first preset threshold and the change of the dual variable is less than a second preset threshold, output the current solution as the optimal solution of the variational optimization objective function.

[0144] First, obtain the variational objective function to be optimized. This objective function usually consists of two parts: one part is related to the smoothness of the probability density distribution, and the other part is related to the neighborhood structure constraint of the system state trajectory. For example, the objective function can be to minimize the length of the trajectory while ensuring the smoothness of the trajectory.

[0145] Next, the variational optimization objective function is decomposed into two sub-problems. The first sub-problem is a smooth optimization problem, which contains the norm of the second-order derivative of the probability density distribution, for example, the curvature of the trajectory. The second sub-problem is a non-smooth optimization problem, which contains the constraints on the neighborhood structure of the system state trajectory sequence, for example, the trajectory must pass through certain specified points.

[0146] Then, the density gradient is calculated based on the point density distribution of the system state trajectory sequence, and an adaptive weight matrix is ​​constructed. For example, if the point density in some areas of the trajectory is high, these areas are given a larger weight. Assume that there are 10 state points with coordinates of (1, 1), (1, 2), (2, 1), (2, 2), (3, 3), (4, 3), (3, 4), (4, 4), (5, 5), (6, 6). By calculating the distance from each point to other points, a 10×10 distance matrix can be obtained. Then, a distance threshold is set, such as 2, and points with a distance less than the threshold are regarded as neighbor points. Based on the number of neighbor points, the density of each point can be calculated. Finally, the adaptive weight matrix is ​​calculated based on the density, for example, the density value is normalized to between 0 and 1 as the weight.

[0147] After that, the second-order derivative of the probability density distribution is calculated, and the momentum term is constructed in combination with the gradient direction of the previous iteration. The momentum term is substituted into the gradient descent iteration formula to solve the smooth subproblem. For example, assuming that the gradient direction of the current iteration is (0.1, 0.2) and the gradient direction of the previous iteration is (0.05, 0.1), the momentum term can be set to 0.5*(0.1, 0.2)+0.5*(0.05, 0.1)=(0.075, 0.15).

[0148] At the same time, based on the system state trajectory sequence, a K-nearest neighbor graph is constructed. For example, if K=3 is set, each state point is connected to its three nearest neighbor points. Then, the local topological feature vector of each state point is calculated, such as the distance and angle between each state point and its neighbor points. Finally, the non-smooth subproblem is iteratively solved through the neighbor points. For example, the coordinates of each state point are iteratively updated to satisfy the neighborhood structure constraints.

[0149] According to the adaptive weight matrix, the step coefficients of the smooth subproblem and the non-smooth subproblem are calculated. The step coefficients are multiplied by the Lagrange multiplier to update the dual variables. For example, assuming that the step coefficient of the smooth subproblem is 0.1, the step coefficient of the non-smooth subproblem is 0.2, and the Lagrange multiplier is 1, then the dual variables are updated to 1+0.1*0.1+0.2*0.2=1.05.

[0150] Calculate the Euclidean distance between the solution of the smooth subproblem and the solution of the non-smooth subproblem to get the difference norm. Calculate the difference between the current dual variable and the dual variable of the previous iteration to get the change of the dual variable. For example, assuming that the solution of the smooth subproblem is (1, 2) and the solution of the non-smooth subproblem is (1.1, 1.9), then the difference norm is sqrt((1.1-1) 2 +(1.9-2) 2 )=0.141. Assuming that the current dual variable is 1.05 and the dual variable of the previous iteration is 1, the change in the dual variable is 0.05.

[0151] When the difference norm is less than the first preset threshold and the change in the dual variable is less than the second preset threshold, the current solution is output as the optimal solution of the variational optimization objective function. For example, the first preset threshold is set to 0.2 and the second preset threshold is set to 0.1. If the difference norm 0.141 is less than 0.2 and the change in the dual variable 0.05 is less than 0.1, the current solution (1.1, 1.9) is output as the optimal solution.

[0152] In this embodiment, by decomposing the complex variational optimization problem into two sub-problems, the difficulty of solving the problem can be reduced and the efficiency of solving the problem can be improved; the introduction of the adaptive weight matrix and the momentum term can enhance the stability of the algorithm and avoid falling into the local optimal solution; by introducing the neighborhood structure constraint, the quality of the solution can be guaranteed to meet the needs of practical problems.

[0153] In an optional embodiment, the mutation point information is input into a dynamics prediction system, and the risk evolution trend is calculated by combining bifurcation theory and singular spectrum analysis, including:

[0154] Constructing a feature matrix for the mutation point information, calculating the covariance of the feature matrix, obtaining a feature vector sequence, projecting the feature vector sequence into a dynamic space, constructing a bifurcation equation, and substituting the bifurcation equation into a prediction-correction iterator to generate a homotopic continuation path;

[0155] Identify the state mutation interval on the homotopy extension path, reconstruct the mutation point information into a trajectory matrix within the state mutation interval, perform singular value decomposition on the trajectory matrix, and extract left and right singular vectors whose singular values ​​are greater than a preset singular value threshold; use a kernel function to map the left and right singular vectors to a feature space, and construct a sparse optimization objective function with an L1 norm constraint based on the feature space; solve the sparse optimization objective function through an iterative soft threshold method to obtain a sparse representation coefficient of the state evolution mode;

[0156] According to the sparse representation coefficient, the homotopy extension path is divided into multiple time windows, and a state transition probability matrix is ​​constructed in each time window; based on the state transition probability matrix, a Gaussian correlation function is constructed to combine the marginal distribution of a single risk factor into a joint distribution probability;

[0157] Using the joint distribution probability, a random matrix sample set is generated, and the eigenvalue statistical distribution of the state transition probability matrix on the random matrix sample set is calculated;

[0158] According to the statistical distribution of the characteristic values, the state transition paths whose significance coefficients are greater than a preset significance coefficient threshold are screened, and the probability values ​​are accumulated along the state transition paths to obtain the risk evolution trend.

[0159] The homotopy continuation path specifically refers to the process of gradually "transforming" a relatively simple problem into a complex problem in a gradual way. Simply put, it provides a "path" that starts from a known simple solution and slowly transitions to the solution of a complex problem. It's like starting from a flat hillside and gradually walking along the set route to the rugged top of the mountain. In application, each step on this path represents a state of the problem, and the goal of continuation is to find the answer to the complex problem through gradual adjustments. It is usually used to solve nonlinear equations that are difficult to solve directly or to analyze changes in system states.

[0160] The left and right singular vectors specifically refer to a kind of "directional" information that describes the internal characteristics of the matrix. Among them, the left singular vector can be understood as describing the main direction of the "row" of the matrix, indicating the main change pattern of the original data in its space. The right singular vector describes the main direction of the "column" of the matrix, revealing the main characteristics of the matrix in another projection space. The left and right singular vectors are an in-depth interpretation of the structure of a matrix, just like when observing a three-dimensional object, you can look at its outline from the front (left singular vector) or look at its structure from the side (right singular vector). In practice, these vectors are often used to extract the main components of the data and remove irrelevant or redundant information.

[0161] The soft threshold method specifically refers to a technique used to "simplify" data, with the goal of highlighting the important parts and suppressing the unimportant parts. Similar to a "filter", the smaller values ​​in the data (considered to be noise or unimportant components) are directly reset to zero, and the larger values ​​are also reduced to a certain extent, making the data more concise and sparse. This method is very commonly used in data processing and feature selection, especially when it is necessary to reduce data redundancy or improve model efficiency, it can help remove "irrelevant" information.

[0162] First, obtain the mutation point information of the system. This information can be the time point, amplitude, and type of change when the system state changes significantly. Then, convert this mutation point information into a feature matrix. Each row of the feature matrix represents a mutation point, and each column represents a feature of the mutation point, such as time, amplitude, and type.

[0163] Next, the covariance matrix of the characteristic matrix is ​​calculated, and the eigenvalue decomposition of the covariance matrix is ​​performed to obtain the eigenvector sequence. These eigenvectors are projected into the dynamic space to construct the bifurcation equation. The bifurcation equation describes the law of qualitative change of the system state with the change of control parameters. Substitute the bifurcation equation into the prediction-correction iterator to generate the homotopy continuation path. The homotopy continuation path is a continuous path connecting the current state and the future state of the system.

[0164] Identify state mutation intervals on the homotopy extension path. These intervals are areas where the system state changes dramatically. In these intervals, reconstruct the mutation point information into a trajectory matrix. Perform singular value decomposition on the trajectory matrix and extract the left and right singular vectors whose singular values ​​are greater than the preset threshold. Assuming the preset threshold is 0.1, the extracted left and right singular vectors whose singular values ​​are greater than 0.1 form a new matrix.

[0165] The kernel function is used to map the left and right singular vectors to a high-dimensional feature space. In this feature space, a sparse optimization objective function with an L1 norm constraint is constructed. The objective function is solved by an iterative soft threshold method to obtain the sparse representation coefficients of the state evolution mode. These coefficients represent the contribution of different evolution modes to the system state.

[0166] According to the sparse representation coefficient, the homotopy extension path is divided into multiple time windows. In each time window, a state transition probability matrix is ​​constructed. This matrix describes the probability of the system transitioning from one state to another. For example, in a time window, the probability of the system transitioning from state A to state B is 0.8, and the probability of transitioning from state A to state C is 0.2.

[0167] Based on the state transition probability matrix, a Gaussian correlation function is constructed to combine the marginal distributions of individual risk factors into a joint distribution probability.

[0168] Generate a random matrix sample set using the joint distribution probability and calculate the eigenvalue statistical distribution of the state transition probability matrix on the random matrix sample set.

[0169] According to the statistical distribution of the eigenvalues, the state transition paths with significance coefficients greater than the preset threshold are screened. The probability values ​​are accumulated along these paths to obtain the risk evolution trend. For example, suppose the system has three possible states: stable, volatile, and collapsed. By analyzing the state transition paths and cumulative probability values, the probability of the system being in different states in the future can be predicted, thereby evaluating the evolution trend of system risk.

[0170] In this embodiment, by combining mutation point information, bifurcation theory and singular spectrum analysis, the law of system state change can be captured more accurately, thereby improving the accuracy of risk prediction; by using sparse optimization methods and random matrix theory, the impact of noise and abnormal data on the prediction results can be effectively reduced, and the robustness of the prediction can be enhanced; by analyzing the state transition path and the cumulative probability value, a more comprehensive risk assessment can be provided, providing decision makers with more effective risk management strategies.

[0171] In an optional implementation, based on the risk accumulation level, a dynamic warning threshold interval is set using critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, a multi-objective optimization decision maker is started, and an optimal intervention strategy is calculated using a Lagrangian relaxation algorithm. The optimal intervention strategy is applied to a nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop. Through adaptive filtering processing, iterative optimization of risk warning includes:

[0172] Based on the cumulative level of system risk, the maximum eigenvalue of the state transition probability matrix is ​​calculated using the critical state theory, and the deviation range between the maximum eigenvalue and the preset characteristic threshold is determined as the upper and lower boundaries of the dynamic warning interval;

[0173] Constructing a multi-objective optimization model including a risk control objective function and an intervention cost objective function, wherein the risk control objective function and the intervention cost objective function are combined by a weight coefficient;

[0174] The constraints of the multi-objective optimization model are relaxed, Lagrangian multipliers are introduced to construct an augmented Lagrangian function, the augmented Lagrangian function is decomposed into quadratic penalty terms to obtain a sub-problem sequence, and based on the gradient information of the sub-problem sequence, the Lagrangian relaxation iteration method is used to alternately update the intervention strategy vector and the Lagrangian multiplier, and the penalty factor is adaptively adjusted to obtain the optimal intervention strategy;

[0175] Constructing a monotonically increasing nonlinear response function, mapping the optimal intervention strategy to a state transition field through the nonlinear response function, and obtaining a state transition probability matrix after intervention;

[0176] Based on the upper and lower boundaries of the dynamic warning interval and the system stability criterion, an expected state matrix is ​​constructed, a difference operation is performed between the state transition probability matrix after intervention and the expected state matrix to obtain an error matrix, and a feedback compensation loop is constructed based on the error matrix;

[0177] The feedback compensation loop is processed by adaptive Kalman filtering, and the error matrix is ​​iteratively corrected by adjusting the gain matrix until the error matrix converges to a preset accuracy range, thereby completing the dynamic optimization of risk warning.

[0178] The relaxation process specifically refers to a method of making the problem easier to solve by relaxing or transforming strict constraints. The original problem may contain some constraints that must be strictly met. These constraints will limit the solution space and thus increase the difficulty of calculation. The core of relaxation is to allow these constraints to be "violated" within a certain range, but by introducing a penalty mechanism or adjustment method, gradually approaching the original strict conditions. For example, a hard equality constraint can be transformed into an additional penalty term in the objective function, or relaxed to an inequality constraint that allows a certain error. The purpose of this is to simplify the problem into a more flexible form that is easier to solve through numerical methods or iterative algorithms while retaining the essential characteristics of the problem. In the end, after multiple adjustments and optimizations, the relaxation process will allow the solution to gradually meet the constraints of the original problem while achieving the optimization goal.

[0179] The risk warning system uses critical state theory to set the dynamic warning threshold interval according to the current risk accumulation level of the system. First, the system calculates the maximum eigenvalue of the state transition probability matrix. This maximum eigenvalue represents the potential speed and range of system risk propagation. The system compares this maximum eigenvalue with a preset characteristic threshold, and the deviation range of the two determines the upper and lower boundaries of the dynamic warning interval. For example, assuming that the preset characteristic threshold is 0.8, the current calculated maximum eigenvalue is 0.7, and the deviation range is set to ±0.05, then the dynamic warning interval is [0.75, 0.85].

[0180] When the system risk accumulation level enters the dynamic warning interval, the multi-objective optimization decision maker will be activated. This decision maker contains two objective functions: the risk control objective function and the intervention cost objective function. The two objective functions are combined by weight coefficients. For example, the weight of the risk control objective function is 0.7, and the weight of the intervention cost objective function is 0.3. The goal of the risk control objective function is to minimize risk losses, and the goal of the intervention cost objective function is to minimize the cost of intervention measures. Assuming that risk losses are measured by risk value and intervention costs are measured by funds, the decision maker will try to find an intervention strategy that minimizes 0.7*risk value+0.3*funds.

[0181] In order to solve this multi-objective optimization problem, the system uses the Lagrangian relaxation algorithm. First, the constraints of the multi-objective optimization model are relaxed, and Lagrangian multipliers are introduced to construct the augmented Lagrangian function. Then, the augmented Lagrangian function is decomposed into quadratic penalty terms to obtain a series of sub-problems. Based on the gradient information of these sub-problems, the system uses the Lagrangian relaxation iteration method to alternately update the intervention strategy vector and the Lagrangian multiplier, and adaptively adjusts the penalty factor to finally obtain the optimal intervention strategy. For example, assuming that the intervention strategy vector contains two parameters: intervention intensity and intervention time, through iterative calculation, the optimal intervention intensity is 0.5 and the optimal intervention time is 10 days.

[0182] The system constructs a monotonically increasing nonlinear response function, maps the optimal intervention strategy to the state transition field, and obtains the state transition probability matrix after intervention. For example, assuming that the nonlinear response function is y = x 2 , the optimal intervention strategy is 0.5, then the mapped value is 0.25, which will be used to adjust the state transition probability matrix.

[0183] The system will construct an expected state matrix based on the upper and lower boundaries of the dynamic warning interval and the system stability criterion. Then, the state transition probability matrix after intervention is subtracted from the expected state matrix to obtain an error matrix. Based on this error matrix, the system will construct a feedback compensation loop. For example, assuming that the maximum eigenvalue of the expected state matrix is ​​0.8 and the maximum eigenvalue of the state transition probability matrix after intervention is 0.75, then the error is 0.05.

[0184] Finally, the system uses adaptive Kalman filtering for the feedback compensation loop. The error matrix is ​​iteratively corrected by adjusting the gain matrix until the error matrix converges to a preset accuracy range, such as 0.01, completing the dynamic optimization of risk warning.

[0185] In this embodiment, by dynamically adjusting the warning threshold interval, the changing trend of system risk can be captured more accurately, avoiding missed reports and false alarms, thereby improving the accuracy of the warning; the combination of the multi-objective optimization decision maker and the Lagrangian relaxation algorithm can effectively balance risk control and intervention costs, and find the optimal intervention strategy, thereby minimizing risk losses; the application of feedback compensation loops and adaptive Kalman filtering can monitor and adjust the intervention effect in real time, and continuously optimize the intervention strategy, thereby enhancing the stability of the system.

[0186] Figure 2 FIG. 1 is a schematic diagram of the structure of the intelligent management and risk warning analysis agent for the entire clinical trial cycle according to an embodiment of the present invention. Figure 2 As shown, the intelligent agent includes:

[0187] The first unit is used to collect clinical trial data, map the clinical trial data to a high-dimensional feature space through nonlinear dynamic projection, extract time-frequency features based on a spectral decomposition network in the high-dimensional feature space through a composite algorithm based on wavelet transform and Hilbert transform, input the obtained time-frequency features into a dynamic compressed sensing model for sparse reconstruction, filter redundant information, obtain key feature signals, and generate a feature association network for the entire clinical trial cycle based on the key feature signals through a manifold learning algorithm;

[0188] The second unit is used to construct a nonlinear state migration field based on the characteristic association network by stochastic differential equation calculation, monitor the state migration field in real time by using an adaptive sampling strategy, capture the mutation point information of the risk state, input the mutation point information into the dynamic prediction system, calculate the risk evolution trend by combining bifurcation theory and singular spectrum analysis, calculate the system state transition probability based on the risk evolution trend, and obtain the risk accumulation level;

[0189] The third unit is used to set the dynamic warning threshold interval according to the risk accumulation level by using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated by the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop, and the risk warning is iteratively optimized through adaptive filtering processing.

[0190] According to a third aspect of the embodiments of the present invention,

[0191] An electronic device is provided, comprising:

[0192] processor;

[0193] a memory for storing processor-executable instructions;

[0194] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0195] According to a fourth aspect of the embodiments of the present invention,

[0196] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0197] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The clinical trial full cycle intelligent management and risk early warning analysis method is characterized by: include: Collect clinical trial data, map the clinical trial data to a high-dimensional feature space through nonlinear dynamic projection, extract time-frequency features based on a spectral decomposition network in the high-dimensional feature space through a composite algorithm based on wavelet transform and Hilbert transform, input the obtained time-frequency features into a dynamic compressed sensing model for sparse reconstruction, filter redundant information, obtain key feature signals, and generate a feature association network for the entire cycle of the clinical trial based on the key feature signals through a manifold learning algorithm; Based on the characteristic association network, a nonlinear state migration field is constructed by stochastic differential equation calculation, and an adaptive sampling strategy is used to monitor the state migration field in real time to capture the mutation point information of the risk state. The mutation point information is input into the dynamic prediction system, and the risk evolution trend is calculated by combining bifurcation theory and singular spectrum analysis. Based on the risk evolution trend, the system state transition probability is calculated to obtain the risk accumulation level; Based on the risk accumulation level, the dynamic warning threshold interval is set using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated through the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop, and the risk warning is iteratively optimized through adaptive filtering processing.

2. The method according to claim 1, characterized in that Mapping clinical trial data to a high-dimensional feature space through nonlinear dynamic projection includes: Based on the kernel function, the feature mapping result of the kernel function is calculated, and the kernel function includes a radial basis kernel function, a polynomial kernel function and an exponential kernel function, an initial weight coefficient is assigned to each kernel function, the Mahalanobis distance between the sample points is calculated, a similarity matrix is ​​constructed, random matrix decomposition is performed, the kernel function parameters are updated and the weight coefficient is adjusted, and the weighted mapping results of the kernel function are combined to form an initial high-dimensional feature mapping matrix; Calculate the conditional entropy value between each feature in the initial high-dimensional feature mapping matrix, determine the information redundancy, measure the mutual information coefficient between feature pairs, obtain the correlation strength, construct a feature importance scoring function based on the information redundancy and the correlation strength, calculate the importance score of each feature, construct an initial feature set, sort the initial feature set in descending order based on the importance score, select the feature with the highest importance score as the reference feature, calculate the similarity matrix between the reference feature and other features in the initial feature set, based on a preset similarity threshold, mark the feature whose similarity to the reference feature is greater than the similarity threshold as a redundant feature, remove the redundant feature from the initial feature set, obtain an updated feature set, recalculate the importance score of the updated feature set, select the feature with the highest importance score as the new reference feature, iterate and output a feature importance sequence; Setting an importance threshold for the feature importance sequence, performing dimension screening, and executing a preset iterative projection algorithm to optimize and reconstruct the initial high-dimensional feature mapping matrix to obtain an optimal feature mapping matrix; Inputting clinical trial data into the optimal feature mapping matrix for projection transformation, calculating the probability distribution characteristics of mapping errors, constructing an error compensation function based on a Gaussian process, performing gradient descent iteration to optimize compensation parameters, and generating an error compensation mapping matrix; The error compensation mapping matrix and the optimal feature mapping matrix are combined in series to construct a nonlinear dynamic projection mapping model to achieve mapping conversion of the clinical trial data to a high-dimensional feature space.

3. The method according to claim 1, characterized in that Based on the key feature signals, a feature association network for the entire clinical trial cycle is generated through a manifold learning algorithm, including: By calculating the statistical characteristics of the data sample, a distribution feature vector of each data point is obtained, based on the distribution feature vector, an adaptive function is constructed to determine the neighborhood range around each data point, a distance threshold is set based on the neighborhood range, and the Hadamard product metric distance between the data points is calculated. When the Hadamard product metric distance is less than the distance threshold, a connection relationship is established between the data points to generate a sparse adjacency matrix; Based on the sparse adjacency matrix, the shortest path distance between data sample points is calculated through a sub-problem decomposition strategy to construct an initial geodesic distance matrix; a smoothing function is constructed using a piecewise cubic polynomial to perform smoothing on the initial geodesic distance matrix; abnormal distance values ​​are identified and corrected through an anti-interference regression method to obtain an optimized geodesic distance matrix; Constructing a Laplacian operator based on the optimized distance matrix, performing eigenvalue decomposition on the Laplacian operator to obtain eigenvalues ​​and eigenvectors; selecting the eigenvector corresponding to the smallest non-zero eigenvalue as the basis of low-dimensional representation, and constructing a low-dimensional embedding space; In the low-dimensional embedding space, the connection between each feature node and the corresponding other nodes is counted, the local connectivity is calculated, and the adaptive penalty function is determined according to the local connectivity. The adaptive penalty function is combined with the sparse constraint and the smooth constraint to construct the optimization target, and the network connection parameters are obtained. Based on the network connection parameters, the initial structure of the feature association network is constructed; Calculate the probability overlap between feature distribution pairs to obtain distribution similarity, and convert the distribution similarity into a network update coefficient; iteratively adjust the connection strength and connection direction of the feature association network based on the network update coefficient; calculate the network structure difference between adjacent iterative steps, terminate the iteration when the network structure difference is less than a preset threshold, and output the final feature association network structure.

4. The method according to claim 1, characterized in that: Based on the characteristic association network, constructing a nonlinear state migration field by stochastic differential equation calculation includes: Performing variational mode decomposition on the topological structure of the feature association network, separating the dynamic features in the feature association network into multiple eigenmodes, and extracting a multi-frequency feature component matrix through spectral analysis; Substituting the multi-frequency characteristic component matrix into the local linearization equation to calculate the state transfer gradient, constructing an adaptive drift term, constructing the multi-frequency characteristic component matrix as a random perturbation matrix to form a multi-scale diffusion term, and constructing a group of stochastic differential equations based on the adaptive drift term and the multi-scale diffusion term; The stochastic differential equations are solved by using an adaptive variable step-size integration algorithm, a converged solution is obtained by dynamically adjusting the integration step-size and introducing a random disturbance term, and a system state trajectory sequence is generated; According to the distribution of the system state trajectory sequence, a probability density model is constructed by using a hybrid kernel density estimation method, and the kernel function parameters of the probability density model are optimized by an iterative expectation algorithm to obtain the probability density distribution of the system state space; Constructing a variational optimization objective function, wherein the variational optimization objective function includes a smoothness constraint term and a local structure constraint term, wherein the smoothness constraint term is measured using a second-order derivative norm, and the local structure constraint term is constructed based on a neighborhood relationship of the system state trajectory sequence; The variational optimization objective function is solved by a splitting bidirectional optimization algorithm to form a nonlinear state migration field that maintains the local topological structure of the system state evolution.

5. The method according to claim 4, characterized in that Using the split bidirectional optimization algorithm to solve the variational optimization objective function includes: Decomposing the variational optimization objective function into a smooth subproblem including a second-order derivative norm term of a probability density distribution and a non-smooth subproblem including a neighborhood structure constraint term of a system state trajectory sequence; Calculating density gradients based on point density distribution of the system state trajectory sequence and constructing an adaptive weight matrix; Calculate the second-order derivative of the probability density distribution, construct a momentum term based on the gradient direction of the previous iteration, substitute the momentum term into the gradient descent iterative formula, and solve the smooth subproblem; Based on the system state trajectory sequence, a K nearest neighbor graph is constructed, the local topological feature vector of each state point is calculated, and the non-smooth sub-problem is iteratively solved through the nearest neighbor points; According to the adaptive weight matrix, the step coefficients of the smooth subproblem and the non-smooth subproblem are calculated, the step coefficients are multiplied by the Lagrange multiplier, and the dual variables are updated; Calculate the Euclidean distance between the solution of the smooth subproblem and the solution of the non-smooth subproblem to obtain the difference norm; calculate the difference between the current dual variable and the dual variable of the previous iteration to obtain the change of the dual variable; when the difference norm is less than a first preset threshold and the change of the dual variable is less than a second preset threshold, output the current solution as the optimal solution of the variational optimization objective function.

6. The method according to claim 1, characterized in that The mutation point information is input into the dynamics prediction system, and the risk evolution trend is calculated by combining bifurcation theory and singular spectrum analysis, including: Constructing a feature matrix for the mutation point information, calculating the covariance of the feature matrix, obtaining a feature vector sequence, projecting the feature vector sequence into a dynamic space, constructing a bifurcation equation, and substituting the bifurcation equation into a prediction-correction iterator to generate a homotopic continuation path; Identify the state mutation interval on the homotopy extension path, reconstruct the mutation point information into a trajectory matrix within the state mutation interval, perform singular value decomposition on the trajectory matrix, and extract left and right singular vectors whose singular values ​​are greater than a preset singular value threshold; use a kernel function to map the left and right singular vectors to a feature space, and construct a sparse optimization objective function with an L1 norm constraint based on the feature space; solve the sparse optimization objective function through an iterative soft threshold method to obtain a sparse representation coefficient of the state evolution mode; According to the sparse representation coefficient, the homotopy extension path is divided into multiple time windows, and a state transition probability matrix is ​​constructed in each time window; based on the state transition probability matrix, a Gaussian correlation function is constructed to combine the marginal distribution of a single risk factor into a joint distribution probability; Using the joint distribution probability, a random matrix sample set is generated, and the eigenvalue statistical distribution of the state transition probability matrix on the random matrix sample set is calculated; According to the statistical distribution of the characteristic values, the state transition paths whose significance coefficients are greater than a preset significance coefficient threshold are screened, and the probability values ​​are accumulated along the state transition paths to obtain the risk evolution trend.

7. The method according to claim 1, characterized in that Based on the risk accumulation level, the dynamic warning threshold interval is set using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated by the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop. Through adaptive filtering processing, iterative optimization of risk warning includes: Based on the cumulative level of system risk, the maximum eigenvalue of the state transition probability matrix is ​​calculated using the critical state theory, and the deviation range between the maximum eigenvalue and the preset characteristic threshold is determined as the upper and lower boundaries of the dynamic warning interval; Constructing a multi-objective optimization model including a risk control objective function and an intervention cost objective function, wherein the risk control objective function and the intervention cost objective function are combined by a weight coefficient; The constraints of the multi-objective optimization model are relaxed, Lagrangian multipliers are introduced to construct an augmented Lagrangian function, the augmented Lagrangian function is decomposed into quadratic penalty terms to obtain a sub-problem sequence, and based on the gradient information of the sub-problem sequence, the Lagrangian relaxation iteration method is used to alternately update the intervention strategy vector and the Lagrangian multiplier, and the penalty factor is adaptively adjusted to obtain the optimal intervention strategy; Constructing a monotonically increasing nonlinear response function, mapping the optimal intervention strategy to a state transition field through the nonlinear response function, and obtaining a state transition probability matrix after intervention; Based on the upper and lower boundaries of the dynamic warning interval and the system stability criterion, an expected state matrix is ​​constructed, a difference operation is performed between the state transition probability matrix after intervention and the expected state matrix to obtain an error matrix, and a feedback compensation loop is constructed based on the error matrix; The feedback compensation loop is processed by adaptive Kalman filtering, and the error matrix is ​​iteratively corrected by adjusting the gain matrix until the error matrix converges to a preset accuracy range, thereby completing the dynamic optimization of risk warning.

8. An intelligent agent for full-cycle intelligent management and risk early warning analysis of clinical trials, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect clinical trial data, map the clinical trial data to a high-dimensional feature space through nonlinear dynamic projection, extract time-frequency features based on a spectral decomposition network in the high-dimensional feature space through a composite algorithm based on wavelet transform and Hilbert transform, input the obtained time-frequency features into a dynamic compressed sensing model for sparse reconstruction, filter redundant information, obtain key feature signals, and generate a feature association network for the entire clinical trial cycle based on the key feature signals through a manifold learning algorithm; The second unit is used to construct a nonlinear state migration field based on the characteristic association network by stochastic differential equation calculation, monitor the state migration field in real time by using an adaptive sampling strategy, capture the mutation point information of the risk state, input the mutation point information into the dynamic prediction system, calculate the risk evolution trend by combining bifurcation theory and singular spectrum analysis, calculate the system state transition probability based on the risk evolution trend, and obtain the risk accumulation level; The third unit is used to set the dynamic warning threshold interval according to the risk accumulation level by using the critical state theory. When the risk accumulation level enters the dynamic warning threshold interval, the multi-objective optimization decision maker is started, and the optimal intervention strategy is calculated by the Lagrangian relaxation algorithm. The optimal intervention strategy is applied to the nonlinear state migration field through a nonlinear response function to regulate the risk situation. The intervention result is input into the nonlinear state migration field through a feedback compensation loop, and the risk warning is iteratively optimized through adaptive filtering processing.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images