A clinical trial risk identification and early warning method and system

Abstract

This application relates to a clinical trial risk identification and early warning method and system, which includes: integrating multiple sources through multi-view fusion with complementary specific constraints to obtain fused risk feature vectors for each center and each monitoring time window; constructing a multi-dimensional risk factor system and establishing dynamic association weights for each risk factor using a variable-level self-attention mechanism; performing three-level association learning prediction, using dynamic association weights as attention priors, outputting short-term risk level predictions using a convolutional-long short-term memory network, outputting medium-term risk transmission path predictions through center-level token cross-attention, and outputting long-term risk situation assessments based on covariance variational autoencoders and mutual information; generating comprehensive risk predictions and comprehensive risk scores through gated attention fusion, and embedding an inductive conformal predictor to distinguish between definite and uncertain early warnings, and providing hierarchical alerts for definite early warnings.

Description

Technical Field

[0001] This application relates to the field of risk warning technology, and in particular to a method and system for identifying and warning of risks in clinical trials. Background Technology

[0002] Clinical trials are a crucial step in drug development, verifying the safety and efficacy of candidate drugs and determining whether a new drug can be successfully marketed. A complete clinical trial involves multiple parties, including sponsors, contract research organizations, research institutions, researchers, subjects, and ethics committees. It covers multiple stages, such as trial design, center selection, ethical approval, subject recruitment and enrollment, clinical data collection and monitoring, data cleaning and statistical analysis, summary report writing, and center closure. It is characterized by a large number of participating entities, a long execution cycle, a wide range of data sources, and high regulatory requirements.

[0003] Currently, the industry primarily relies on a combination of rule-based manual monitoring and regular reporting to manage risks in clinical trials. During the project initiation phase, based on attendees' past project experience, company standard operating procedures, and limited preliminary project investigation data, potential risk events are identified through meeting discussions, and the results are recorded in a project risk management table. This table mainly includes: risk category, risk description, and risk level. A risk prediction model is then used to organize and predict the information in the risk management table, outputting the results. Simultaneously, the handling of each risk event is regularly checked and tracked until the risk is closed.

[0004] The aforementioned technologies primarily rely on personal experience for risk identification, and the sources of risk factors are limited to the subjective perceptions of the participants. This makes it difficult to comprehensively cover multiple information sources, resulting in a limited scope of risk identification and insufficient systematicity. Summary of the Invention

[0005] To address the issues of risk identification relying primarily on personal experience, where risk factors are limited to the subjective perceptions of attendees and fail to comprehensively cover multiple information sources, resulting in a narrow scope and insufficient systematicity in risk identification, this application provides a method and system for risk identification and early warning in clinical trials.

[0006] Firstly, this application provides a clinical trial risk identification and early warning method, which adopts the following technical solution: including: The system receives multi-source heterogeneous data from clinical trials, extracts complementary and specific risk features from each data view, and fuses these complementary and specific risk features through orthogonal constraints and adversarial training to obtain a fused risk feature vector for each trial center within each monitoring time window. A multi-dimensional risk factor system is constructed, and the risk score of each risk factor is calculated based on the fused risk feature vector. The dynamic correlation weight between each risk factor is constructed using a variable-level self-attention mechanism. The fused risk feature vectors of each center are constructed into a center risk state sequence in chronological order. The dynamic association weights are used as attention priors. Convolutional and long short-term memory networks are used to perform temporal modeling on the center risk state sequence to obtain short-term risk level predictions. The central risk state sequence is embedded as a central-level token, and cross-central risk transmission relationships are captured through cross-attention to obtain mid-term risk transmission path predictions; The central risk state sequence is reduced in dimensionality based on the covariance variational autoencoder, and a global risk association network is constructed using the dynamic association weights to obtain a long-term risk situation assessment. By using gated attention to fuse the short-term risk level prediction, the medium-term risk transmission path prediction, and the long-term risk situation assessment, a comprehensive risk prediction and a comprehensive risk score are obtained. An inductive conformal predictor is embedded to distinguish between early warning and uncertain early warning based on the comprehensive risk prediction and a preset confidence threshold; for the determined early warning result, a tiered alert is issued based on the comprehensive risk score and a preset risk threshold.

[0007] Preferably, the receiving of multi-source heterogeneous data from clinical trials involves extracting complementary and specific risk features from each data view, and fusing these complementary and specific risk features through orthogonal constraints and adversarial training to obtain a fused risk feature vector for each trial center within each monitoring time window, including: The experimental plan data, project operation data, safety data, quality data, and external environment data were coded and preprocessed to obtain the initial feature representations of the five types of data views; For each type of data view, a complementary encoder and a specific encoder are constructed respectively, and the initial feature representation of each view is mapped to the complementary feature subspace and the specific feature subspace respectively to obtain the complementary risk feature and the specific risk feature of each view; Apply orthogonal constraints to the complementary risk features and the specific risk features of the same view to make the two types of features orthogonal to each other, thereby eliminating information redundancy between different views; The complementary risk features of each view are subjected to adversarial training. The complementary risk features of different views are made to be consistent through game between the generator and the discriminator. The complementary risk features of the aligned views are averaged to obtain the shared complementary risk features. The specific risk features of each view are stitched together to obtain the fused specific risk features; The shared complementary risk features are concatenated with the fusion-specific risk features to obtain the fusion risk feature vector.

[0008] Preferably, the construction of a multi-dimensional risk factor system, calculating the risk score of each risk factor based on the fused risk feature vector, and constructing dynamic correlation weights between each risk factor using a variable-level self-attention mechanism, includes: Obtain the identifiers of each risk factor in the predefined risk factor system, which covers multiple risk factors under the dimensions of safety risk, quality risk, schedule risk, compliance risk, and management risk. Based on the fused risk feature vector, the probability of occurrence and the degree of risk hazard of each risk factor within the current monitoring time window are calculated by the first multilayer perceptron. The product of the probability of occurrence and the degree of risk hazard is used as the risk score of each risk factor. The identifiers of each risk factor are mapped to risk factor embedding vectors, and the risk factor embedding vectors of each risk factor are concatenated with the fused risk feature vectors to obtain risk factor tokens. A linear transformation is performed on all risk factor tokens to generate query vectors, key vectors, and value vectors. The attention weight matrix between each risk factor is calculated by the dot product of the query vector and the key vector. Each element in the attention weight matrix represents the correlation strength between two risk factors. Based on the attention weight matrix, the value vectors of each risk factor are weighted and aggregated to obtain the dynamic interactive representation of each risk factor. The dynamic interaction representation of the risk factors is used as a new query vector, and secondary attention is calculated with the key vector to obtain the dynamic association weight. The dynamic association weight represents the contribution weight of each risk factor to the overall risk of the clinical trial within the current monitoring time window.

[0009] Preferably, the step of constructing a center risk state sequence by arranging the fused risk feature vectors of each center in chronological order, using the dynamic association weights as attention priors, and performing temporal modeling of the center risk state sequence using convolutional and long short-term memory networks to obtain short-term risk level predictions includes: For each test center, the fusion risk feature vectors generated by the test center in each monitoring time window are arranged in chronological order of the visit cycle to obtain the center risk status sequence corresponding to the test center; Construct a temporal prediction network, which includes convolutional layers, temporal prior attention layers, and long short-term memory network layers; The central risk state sequence is input into the convolutional layer, and the local spatial correlation between multidimensional risk features within each time window is extracted through one-dimensional convolution operation to obtain the convolutional risk feature sequence. The convolutional risk feature sequence is input into the temporal prior attention layer, and the dynamic association weight is used as the initial attention bias of the temporal prior attention layer. The risk features of each time step in the convolutional risk feature sequence are weighted to generate a temporally weighted risk feature sequence. The time-weighted risk feature sequence is input into the long short-term memory network layer. The time-series dependency of risk state between adjacent monitoring time windows is captured through the gating mechanism of forget gate, input gate and output gate, and the hidden state sequence of each time step is obtained. The hidden state of the last time step is mapped to a preset risk level space through a fully connected classification layer to obtain the short-term risk level prediction; the short-term risk level prediction represents the risk level of each test center within a preset number of future visit cycles.

[0010] Preferably, the step of embedding the central risk state sequence as a central-level token, capturing cross-central risk transmission relationships through cross-attention, and obtaining a mid-term risk transmission path prediction includes: For each test center, the sequence of center risk states of the test center is mapped to a preset embedding dimension through a linear projection layer to obtain the center-level token vector corresponding to the test center; Each test center is assigned a learnable center location code, and the center location code is added element by element to the corresponding center-level token vector to obtain a location-aware center token. The location-aware center token of the current test center is used as the query vector, and the location-aware center tokens of the other test centers are used as the key vector and value vector, respectively. The inter-center attention weight between the current test center and each of the other test centers is calculated by the dot product of the query vector and the key vector. The value vectors of the remaining test centers are weighted and aggregated using the inter-center attention weights to obtain the risk transmission aggregate representation of the current test center. The risk transmission aggregate representation characterizes the comprehensive impact of the risk status of the remaining test centers on the current test center. Calculations are performed for each test center to obtain the risk transmission aggregate representation for all test centers. The risk transmission aggregate representations of all test centers are combined into a risk transmission strength matrix, where the elements of the risk transmission strength matrix represent the risk transmission strength between test centers. Based on the risk transmission intensity matrix, the risk transmission path between each test center is extracted through connectivity analysis, and the mid-term risk transmission path prediction is output; the mid-term risk transmission path prediction represents the path and intensity of risk transmission from one test center to other test centers within a preset time period in the future.

[0011] Preferably, the step of reducing the dimensionality of the central risk state sequence based on the covariance variational autoencoder and constructing a global risk association network using the dynamic association weights to obtain a long-term risk situation assessment includes: A covariance variational autoencoder is constructed, which includes an encoder and a decoder; the encoder performs nonlinear dimensionality reduction on the input central risk state sequence to obtain the mean vector and covariance matrix of each test center; the covariance matrix represents the correlation between multidimensional risk features within each test center; Gaussian distribution assumptions are applied to the mean vector and covariance matrix, and latent variables are generated from the Gaussian distribution through reparameterized sampling. The latent variables are input into the decoder for reconstruction to obtain the reconstructed risk state sequence. The covariance variational autoencoder is trained by minimizing the reconstruction error between the central risk state sequence and the reconstructed risk state sequence. A variable-level attention module is embedded in the encoder to model the time series of each risk factor as an independent semantic unit in order to capture the interaction dependencies of each risk factor in the global time series. The trained covariance variational autoencoder is used to reduce the dimensionality of the central risk state sequence of each test center to obtain the dimensionality-reduced mean vector and dimensionality-reduced covariance matrix of each test center; based on the dimensionality-reduced covariance matrix, the joint entropy and mutual information value between each test center are calculated under Gaussian distribution to obtain the risk correlation matrix between centers. The inter-center risk correlation matrix is ​​weighted and fused with the dynamic correlation weight to generate a fused correlation matrix. Based on the aforementioned fusion correlation matrix, a global risk correlation network is constructed with each test center as a node and the fusion correlation as the edge weight. Network analysis is performed on the global risk association network to identify the central nodes connected by edges whose association degree exceeds a preset first association threshold as key risk nodes, and to identify the edges connected to nodes whose association degree exceeds a preset second association threshold and whose structural importance exceeds a preset structural threshold as vulnerable links. Based on the key risk nodes and the vulnerable links, the long-term risk situation assessment is obtained; the long-term risk situation assessment represents the risk clustering areas and vulnerable risk transmission paths within the overall scope of the project.

[0012] Preferably, the step of fusing the short-term risk level prediction, the medium-term risk transmission path prediction, and the long-term risk situation assessment using gated attention to obtain a comprehensive risk prediction and a comprehensive risk score includes: The short-term risk level prediction is mapped to a preset fusion dimension through a first linear projection layer to obtain a short-term risk embedding vector; the medium-term risk transmission path prediction is mapped to a fusion dimension through a second linear projection layer to obtain a medium-term risk embedding vector; and the long-term risk situation assessment is mapped to a fusion dimension through a third linear projection layer to obtain a long-term risk embedding vector. The short-term risk embedding vector, the medium-term risk embedding vector, and the long-term risk embedding vector are stacked to form a three-layer risk representation matrix; A gated attention fusion module is constructed, which performs a dot product between the risk embedding vector of each layer in the three-layer risk representation matrix and the shared context vector, and calculates the corresponding gating coefficients of each layer through an activation function; at the same time, self-attention is calculated on the three-layer risk representation matrix to obtain the attention weights of each layer. The gating coefficients of each layer are multiplied element-wise with the attention weights to obtain the fusion modulation weights of each layer; The risk embedding vectors of each layer are weighted and summed using the fusion modulation weights to obtain the comprehensive risk fusion vector; The comprehensive risk fusion vector is input into the classification output head and the scoring output head respectively; the classification output head maps the comprehensive risk fusion vector to a preset risk level space and outputs the comprehensive risk prediction; the scoring output head maps the comprehensive risk fusion vector into a continuous value and outputs the comprehensive risk score. Using the comprehensive risk fusion vector, the marginal contribution value of each risk factor to the comprehensive risk prediction is calculated using the Shapley additivity interpretation method. The marginal contribution values ​​are then arranged in descending order to generate a risk factor contribution ranking.

[0013] Secondly, this application discloses a clinical trial risk identification and early warning device, which adopts the following technical solution, including: The feature fusion module is used to receive multi-source heterogeneous data from clinical trials, extract complementary risk features and specific risk features from each data view, and fuse the complementary risk features and specific risk features through orthogonal constraints and adversarial training to obtain the fused risk feature vector of each trial center within each monitoring time window. The dynamic weighting module is used to construct a multi-dimensional risk factor system, calculate the risk score of each risk factor based on the fused risk feature vector, and construct the dynamic correlation weights between each risk factor using a variable-level self-attention mechanism. The short-term risk module is used to construct a center risk state sequence by arranging the fused risk feature vectors of each center in chronological order, use the dynamic association weight as an attention prior, and use convolutional and long short-term memory networks to perform temporal modeling on the center risk state sequence to obtain a short-term risk level prediction. The intermediate risk module is used to embed the central risk state sequence into a central-level token, capture cross-central risk transmission relationships through cross-attention, and obtain intermediate risk transmission path predictions. The long-term risk module is used to reduce the dimensionality of the central risk state sequence based on the covariance variational autoencoder, construct a global risk association network using the dynamic association weights, and obtain a long-term risk situation assessment. The comprehensive risk module is used to fuse the short-term risk level prediction, the medium-term risk transmission path prediction, and the long-term risk situation assessment using gated attention to obtain a comprehensive risk prediction and a comprehensive risk score. The hierarchical alarm module is used to embed an inductive conformal predictor to distinguish between early warning and uncertain early warning based on the comprehensive risk prediction and a preset confidence threshold; for the determined early warning result, hierarchical alarms are issued based on the comprehensive risk score and a preset risk threshold.

[0014] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed, such as the clinical trial risk identification and early warning method described above.

[0015] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in the clinical trial risk identification and early warning method.

[0016] In summary, this application employs complementary-specific constraint multi-view fusion of multi-source heterogeneous information, including test plan data, operational data, safety data, quality data, and external environment data, to obtain fused risk feature vectors for each test center within each monitoring time window. Based on this, a risk factor system covering safety, quality, schedule, compliance, and management dimensions is constructed, and a variable-level self-attention mechanism is used to establish dynamic correlation weights between various risk factors. A three-layer progressive association learning model is used to achieve multi-scale risk prediction: at the short-term scale, using dynamic correlation weights as attention priors, convolutional and long short-term memory networks are used to perform temporal modeling of the risk state sequences of each center, predicting the short-term risk level of each center in future visit cycles; at the medium-term scale, the risk state sequences of each center are embedded as center-level tokens, and cross-center risk transmission relationships are captured through cross-attention to predict risk transmission paths; at the long-term scale, a covariance variational autoencoder is used to reduce the dimensionality of the risk state sequences of each center, and mutual information is used to construct a global risk association network to assess the overall risk situation of the project. This method utilizes gated attention to fuse three-layer prediction results to obtain a comprehensive risk prediction and comprehensive risk score, and generates a ranking of risk factor contributions based on the Shapley additivity interpretation method. An inductive conformal predictor is embedded to measure the uncertainty of the prediction results, distinguishing between certain and uncertain early warnings. For certain early warnings, a three-level hierarchical alarm is triggered based on the comprehensive risk score, and after triggering the alarm, a safety reinforcement learning model incorporating conservative Q-learning constraints is used to recommend candidate countermeasures. This achieves multi-source data fusion, multi-scale prediction, and interpretable early warning for risks throughout the entire clinical trial lifecycle, significantly improving the systematic nature of risk identification and the reliability of risk early warning. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a clinical trial risk identification and early warning method.

[0018] Figure 2 This is a structural block diagram of a clinical trial risk identification and early warning device. Detailed Implementation

[0019] The following combination Figures 1-2 This application will be described in further detail.

[0020] Reference Figure 1 The embodiments of this application include at least steps S10 to S50.

[0021] S10 receives multi-source heterogeneous data from clinical trials, extracts complementary and specific risk features from each data view, and fuses the complementary and specific risk features through orthogonal constraints and adversarial training to obtain the fused risk feature vector of each trial center within each monitoring time window. S20, construct a multi-dimensional risk factor system, calculate the risk score of each risk factor based on the fused risk feature vector, and use a variable-level self-attention mechanism to construct the dynamic correlation weights between each risk factor; S30, based on association learning, constructs a three-level risk prediction model to predict clinical trial risks at multiple scales, and obtains short-term risk level prediction, medium-term risk transmission path prediction and long-term risk situation assessment. S40 utilizes gated attention to integrate short-term risk level prediction, medium-term risk transmission path prediction, and long-term risk situation assessment to obtain a comprehensive risk prediction and comprehensive risk score. S50, with its embedded inductive conformal predictor, distinguishes between early warning and uncertain early warning based on comprehensive risk prediction and preset confidence thresholds; for confirmed early warning results, it issues tiered alerts based on comprehensive risk scores and preset risk thresholds.

[0022] Specifically, complementary and specific multi-view fusion is performed on multi-source heterogeneous data from clinical trials. Shared common risk features and data-specific risk features are extracted synchronously from the original data. After orthogonal constraint redundancy removal and adversarial training alignment, fused risk feature vectors for each center and time window are obtained. A multi-dimensional risk factor system is constructed, and the risk score of each factor is calculated. A variable-level self-attention mechanism is used to automatically learn the dynamic correlation weights between risk factors. A three-layer progressive association learning model is used to achieve multi-scale risk prediction. Gated attention is used to adaptively fuse the three-layer prediction results to obtain a comprehensive risk prediction and a comprehensive risk score. An inductive conformal predictor is embedded to measure the uncertainty of the prediction results. Only high-confidence, definitive warning results are triggered based on the comprehensive risk score to establish a three-level stratified alarm. This realizes a feasible path from multi-source data fusion to multi-scale prediction and then to reliable early warning, improving the systematic nature and reliability of clinical trial risk identification and early warning.

[0023] In some embodiments, step S10 specifically includes the following steps: Data generated during clinical trials exhibits typical multi-source heterogeneity, including trial protocol data, project operation data, safety data, quality data, and external environment data. These five types of data differ significantly in format, semantics, and collection frequency. However, the risk information they carry is both complementary—reflecting the operational status of the clinical trial from different perspectives—and specific, containing risk signals that cannot be replaced by other data sources. Therefore, this embodiment establishes a multi-view fusion mechanism with complementary and specific constraints to effectively integrate these five types of data.

[0024] Specifically, the five types of data are pre-processed by encoding to obtain initial feature representations for the five types of data views. A complementary encoder and a specific encoder are constructed for each type of data view: the complementary encoder is used to extract common risk features that can be shared with other views in the view, and the specific encoder is used to extract risk features unique to the view (e.g., the SAE event frequency increase signal unique to security data).

[0025] To ensure that complementary and specific features carry information of different natures, orthogonal constraints are applied to complementary and specific risk features in the same view. These orthogonal constraints are achieved by minimizing the Frobenius norm inner product of the two feature matrices, making the complementary and specific feature subspaces orthogonal to each other, thus avoiding information redundancy between the two types of features.

[0026] Simultaneously, to effectively fuse complementary risk features from different data views, adversarial training is performed on the complementary risk features of each view. The complementary encoder of each view is treated as a generator, and a multi-class discriminator is introduced to distinguish the source view of the generated complementary risk features. During training, the generator aims to prevent the discriminator from recognizing the source of the complementary risk features, while the discriminator aims to maximize its recognition accuracy. Through this minimax game, the complementary risk features of each view are aligned to the same semantic space, making the shared risk patterns between different views tend to be consistent. The average of the aligned complementary risk features of each view is then taken to obtain the shared complementary risk features.

[0027] For the specific risk features of each view, since they carry risk information unique to each data source and there are no common requirements across views, they can be directly fused through vector concatenation to obtain the fused specific risk features.

[0028] By concatenating shared complementary risk features with fusion-specific risk features, a fusion risk feature vector for each test center within the current monitoring time window is obtained. This fusion risk feature vector retains the unique risk information of each data source while integrating common risk patterns across data sources.

[0029] In some embodiments, step S20 specifically includes the following steps: Based on Good Clinical Practice (GCP) guidelines and industry standards, a predefined risk factor system encompasses five primary risk dimensions and multiple secondary risk factors: safety risk (including abnormal incidence of serious adverse events in subjects, clustering of laboratory outliers, and delays in center safety reports), quality risk, schedule risk, compliance risk, and management risk. Each risk factor has a unique identifier. Based on the fused risk feature vector, the probability of occurrence and the degree of risk hazard for each risk factor within the current monitoring time window are calculated using a first-level multilayer perceptron. Multiplying these two values ​​yields a risk score for each risk factor, which comprehensively measures the threat level of each risk factor.

[0030] Building upon this foundation, a variable-level self-attention mechanism is employed to construct dynamic association weights among various risk factors. The core principle is to treat each risk factor as an independent semantic unit, capturing the global interaction relationships between different risk factors through self-attention. Specifically, the identifiers of each risk factor are mapped to risk factor embedding vectors, which are then concatenated with the fused risk feature vectors to form risk factor tokens. Linear transformations are performed on all risk factor tokens to generate query vectors, key vectors, and value vectors. The attention weight matrix between each risk factor is calculated using the dot product of the query vector and the key vector. Each element in the matrix represents the association strength between two risk factors; a higher association strength indicates that a change in the state of one risk factor is more likely to cause a change in the state of another risk factor.

[0031] To further enhance the representation of risk factor interactions, the value vectors of each risk factor are weighted and aggregated based on an attention weight matrix to obtain a dynamic interaction representation of risk factors. This dynamic interaction representation is then used as a new query vector, and a second attention calculation is performed with the key vector to finally output the dynamic association weight. This weight represents the relative contribution of each risk factor to the overall risk of the clinical trial within the current monitoring time window; the higher the weight, the greater the impact of the factor on the current risk status.

[0032] The dynamic association weights change dynamically as the experiment progresses: from the initiation phase to the enrollment phase, and then to the data cleaning phase, the importance of different risk factors will be adjusted accordingly (for example, the progress risk factor has a higher weight in the enrollment phase, while the quality risk factor has a higher weight in the data cleaning phase). Thus, through end-to-end learning of the self-attention mechanism, the dynamic association weights can adaptively evolve with the experimental phase without the need for manually setting fixed weights.

[0033] In some embodiments, step S30 specifically includes the following steps: Short-term risk level prediction: For each test center, the fused risk feature vectors generated within each monitoring time window are arranged in chronological order of the visit cycle to obtain a center risk state sequence, which records the evolution trajectory of the center's risk state. Then, a temporal prediction network consisting of convolutional layers, temporal prior attention layers, and Long Short-Term Memory (LSTM) layers is constructed. The center risk state sequence is input into the convolutional layer, and one-dimensional convolution operations are used to extract the local spatial correlations between multi-dimensional risk features within each time window. For example, whether there is a co-occurrence pattern between the two risk features of decreasing enrollment rate and increasing protocol deviation rate within a certain time window. After convolution processing, a convolutional risk feature sequence is generated.

[0034] The convolutional risk feature sequence is input into the temporal prior attention layer. This layer uses the dynamic association weights generated in S20 as the initial attention bias to weight the convolutional risk features at each time step. The advantage of this approach is that it explicitly introduces the differences in importance of each risk factor at the current experimental stage into the temporal modeling. For example, risk features related to enrollment progress, which are given higher weights by the dynamic association weights during the enrollment stage, receive more attention in the temporal modeling, thereby generating a more discriminative temporal weighted risk feature sequence.

[0035] The time-weighted risk feature sequence is input into an LSTM layer. Through gating mechanisms including forget gates, input gates, and output gates, the temporal dependencies of risk states between adjacent monitoring time windows are captured. This long-range dependency is transmitted through the LSTM's cell state encoding, ultimately outputting the hidden state sequence for each time step. The hidden state of the last time step is mapped to a preset risk level space through a fully connected classification layer to obtain a short-term risk level prediction, representing the expected risk level of each center within a preset number of future visit cycles.

[0036] Mid-term Risk Transmission Path Prediction: In clinical trials, risks not only evolve within each center over time but also propagate between different centers. Multiple centers in the same region may be affected by the same regulatory changes, and the failure of a participant recruitment strategy at one center may be adopted by neighboring centers. Sponsor resource adjustments may simultaneously impact multiple related centers. To capture this spatial risk transmission, for each trial center, its central risk state sequence is mapped to a predefined embedding dimension through a linear projection layer, resulting in a center-level token vector. This token vector compresses the entire center's complete risk state sequence into a dense vector representation, carrying all risk state information from trial initiation to the current moment. To preserve the identity information of each center, a learnable center location code is assigned to each center and added element-wise to the center-level token vector to obtain a location-aware center token.

[0037] Cross-center risk transmission is captured using a cross-attention mechanism: The current center's location-aware center token is used as the query vector, and the location-aware center tokens of the other centers are used as the key and value vectors, respectively. The inter-center attention weight is calculated by the dot product of the query vector and the key vector. This weight reflects the degree of influence of the risk status of the other centers on the current center; the larger the weight, the stronger the reference or transmission signal of the corresponding center's risk status to the current center. The value vectors are then weighted and aggregated using the inter-center attention weights to obtain a weighted aggregation representation of the current center's risk transmission. This representation integrates the overall influence of the risk status of all relevant centers on the current center.

[0038] The above operations are performed on each center individually, aggregating and combining the risk transmission from all centers into a risk transmission intensity matrix. Rows in this matrix represent affected centers, and columns represent source centers, with each element quantifying the risk transmission intensity in a specific direction. Connectivity analysis is then performed on this matrix to filter center pairs whose risk transmission intensity exceeds a preset threshold. Directed graph analysis is used to extract risk transmission paths between centers, outputting a mid-term risk transmission path prediction. This prediction reveals where the risk might travel from and to, providing data support for risk prevention and early intervention.

[0039] Long-term risk profile assessment: From a holistic project perspective, a macro-level assessment of the long-term risk profile of the entire trial project is required. A covariance variational autoencoder (CovVAE) is constructed, consisting of an encoder and a decoder. Traditional variational autoencoders typically assume that variables in each dimension of the latent space are independent, but in clinical trials, risk characteristics between different centers often exhibit complex correlations. Therefore, CovVAE relaxes the assumption of variable independence: when the encoder performs nonlinear dimensionality reduction on the input center risk state sequence, in addition to outputting a mean vector representing the mean of risk states at each center, it also outputs a covariance matrix representing the correlations between multidimensional risk characteristics within each center.

[0040] Gaussian distribution assumptions are applied to the mean vector and covariance matrix. Latent variables are generated from the corresponding Gaussian distribution through reparameterized sampling and input into the decoder to reconstruct the reconstructed risk state sequence. CovVAE is trained by minimizing the error between the original sequence and the reconstructed sequence. To enhance the perception of the dependency structure of multiple risk factors, a variable-level attention module is embedded in the encoder, modeling the time series of each risk factor as an independent semantic unit to capture the interaction dependencies of each risk factor in the global time series.

[0041] After training, CovVAE is used to reduce the dimensionality of the risk state sequences of each center, obtaining the dimensionality-reduced mean vector and dimensionality-reduced covariance matrix. Based on the dimensionality-reduced covariance matrix, under the Gaussian distribution assumption, the joint entropy and mutual information values ​​between centers are calculated using the covariance determinant, resulting in the inter-center risk correlation matrix. The mutual information value measures the statistical dependence between the risk states of two centers; a higher value indicates a greater tendency for co-change. The inter-center risk correlation matrix is ​​then weighted and fused with dynamic correlation weights to generate a fused correlation matrix. With each center as a node and the fused correlation as the edge weight, a global risk correlation network is constructed.

[0042] The network is analyzed as follows: Central nodes connected by edges whose correlation exceeds a first correlation threshold are identified as critical risk nodes – representing high-risk clusters that will rapidly propagate through high-weighted edges in the event of a risk incident. Edges connected to nodes whose correlation exceeds a second correlation threshold and whose structural importance (e.g., betweenness centrality or degree centrality) exceeds a preset importance threshold are identified as vulnerable links – representing critical pathways for risk propagation; their breakage or anomaly will significantly impact the overall risk propagation of the project. Based on critical risk nodes and vulnerable links, a long-term risk situation assessment is output, describing global risk clusters and vulnerable risk propagation paths from a macro-level perspective, providing support for strategic decision-making.

[0043] In some embodiments, step S40 specifically includes the following steps: short-term forecasting focuses on recent risk level changes in a single center; medium-term forecasting focuses on the spatial transmission of risk between centers; and long-term assessment focuses on the overall structural risk characteristics of the project. To enable project managers to obtain a unified risk assessment, these three methods need to be integrated, while providing traceable forecasting basis.

[0044] Short-term risk level prediction, medium-term risk transmission path prediction, and long-term risk situation assessment are mapped to the same preset fusion dimension through the first, second, and third linear projection layers, respectively, to obtain short-term risk embedding vector, medium-term risk embedding vector, and long-term risk embedding vector. The three are then stacked to form a three-layer risk representation matrix.

[0045] A gated attention fusion module is constructed to modulate the three-layer risk representation from two dimensions: first, the gating coefficients, which are obtained by performing a dot product between the risk embedding vector of each layer and the shared context vector, and mapping them to the [0,1] interval through an activation function. These coefficients reflect which scale of prediction information should be given higher importance in the current experimental stage. Second, the attention weights, which are obtained by performing self-attention calculation on the three-layer risk representation matrix and performing a dot product operation on the query vector, key vector, and value vector. These weights reflect the semantic relevance between the prediction outputs of each layer. The gating coefficients of each layer are multiplied element-wise with the attention weights to obtain the fusion modulation weights. These weights simultaneously consider the cross-scale information importance and the inter-layer semantic relevance, enabling the fusion process to adaptively adjust according to the experimental stage and the relevance of the prediction content. The fusion modulation weights are then used to weight and sum the risk embedding vectors of each layer to obtain the comprehensive risk fusion vector.

[0046] The comprehensive risk fusion vector is input into the classification output head and the scoring output head, respectively. The classification output head is mapped to the preset risk level space through a fully connected layer and a softmax activation function, and outputs the comprehensive risk prediction (discrete risk level categories); the scoring output head is mapped to continuous values ​​in the interval [0,1] through a fully connected layer, and outputs the comprehensive risk score. The higher the score, the higher the overall risk level of the project.

[0047] In addition to outputting the prediction results, the interpretability of the model's decision-making logic needs to be analyzed. Using the Shapley additivity interpretation method, each dimension of the integrated risk fusion vector is considered as a participant in a cooperative game. The marginal contribution value of each risk factor to the integrated risk prediction is calculated—that is, from all possible combinations of risk factors, a risk factor is added or removed one by one, and the average impact of that factor on the prediction result is calculated. The marginal contribution values ​​of all risk factors are sorted in descending order to generate a risk factor contribution ranking, quantitatively answering the core question of which risk factors have the greatest impact on the overall project risk at the current moment, providing a quantitative basis for subsequent tiered alerts and auxiliary decision-making.

[0048] In some embodiments, step S50 specifically includes the following steps: The comprehensive risk prediction is a point estimate, but in actual risk management scenarios, the model may be unable to make a high-confidence judgment due to insufficient data or marginal cases. Forcing the output of potentially erroneous prediction results will lead to the triggering of false alarms or unnecessary neglect. Therefore, an uncertainty measurement mechanism needs to be introduced.

[0049] An inductive conformal predictor is embedded in the S40 prediction model. This predictor is a model-independent uncertainty estimation framework: using a calibration set reserved from the training data, a non-conformity measure score (NCM) is calculated for each prediction result. Based on the ordinal distribution of the NCM scores, a p-value is calculated for new test samples. The p-value represents the degree of consistency between the new sample and all known samples in the calibration set. For each possible risk level category in the comprehensive risk prediction, the corresponding p-value is calculated using the inductive conformal predictor. The certainty of the comprehensive risk prediction is determined based on a preset confidence threshold: if the p-value for a certain risk level category exceeds the confidence threshold and that category has the highest prediction probability, it is classified as a certainty warning; if the p-values ​​for multiple risk level categories all exceed the confidence threshold, and the risk level cannot be uniquely determined, it is classified as an uncertainty warning. Uncertainty warning samples are marked and pushed to project managers, prompting them to conduct manual judgment in conjunction with more clinical information to avoid the model directly outputting potentially erroneous judgments under low confidence conditions.

[0050] For determining the early warning results, a tiered alert management system is implemented based on a comprehensive risk score and three risk thresholds. When the comprehensive risk score exceeds the first risk threshold, a red alert is triggered—indicating high risk, requiring immediate notification of project management and activation of the emergency plan. When the comprehensive risk score exceeds the second risk threshold but does not exceed the first threshold, a yellow alert is triggered—indicating moderate risk, requiring the project team to develop and implement a response plan within a preset timeframe. When the comprehensive risk score exceeds the third risk threshold but does not exceed the second threshold, a blue alert is triggered—indicating low risk, and this is included in routine project risk monitoring and reported monthly. This tiered alert management system avoids the alert fatigue or missed key signals problems associated with traditional single-threshold alerts, enabling managers to allocate resources and respond accordingly based on the severity of the risk.

[0051] Once an alert is triggered, the contribution ranking of risk factors corresponding to the comprehensive risk prediction is obtained. The top K risk factors with contributions exceeding a preset contribution threshold are extracted as key driving factors. The risk score of each key driving factor within the current monitoring time window is obtained, and the contribution score is multiplied by the risk score to obtain the driving strength score. This score comprehensively measures the factor's contribution to the current alert and the factor's inherent risk level. The key driving factors are sorted in descending order of driving strength score to generate a ranked list. This ranked list of key driving factors, along with the tiered alert level and the comprehensive risk score, is packaged into an interpretable alert report and pushed to the project management terminal.

[0052] In some embodiments, step S60 specifically includes the following steps: When a stratified alert is triggered, a safety reinforcement learning model incorporating conservative Q-learning constraints is used to automatically recommend candidate response measures from a historical clinical trial risk response case library. The role of the conservative Q-learning constraint is to suppress the model's overestimation of high-risk response strategies—that is, the model will not blindly recommend the most aggressive response, but tends to recommend robust strategies that have been historically validated as safe and effective. The recommendation results, along with an interpretable alert report, are pushed to the project management personnel's terminal for their reference in decision-making. After the management personnel select and implement the response measures, the implementation results are fed back to the case library for continuous model iteration and optimization.

[0053] The implementation principle of a clinical trial risk identification and early warning method in this application is as follows: Multi-source heterogeneous information, including trial protocol data, operational data, safety data, quality data, and external environment data, is fused using complementary-specific constraints to obtain a fused risk feature vector for each trial center within each monitoring time window. Based on this, a risk factor system covering safety, quality, schedule, compliance, and management dimensions is constructed, and a variable-level self-attention mechanism is used to establish dynamic correlation weights between various risk factors. Multi-scale risk prediction is achieved through a three-layer progressive association learning model: In the short term, using dynamic correlation weights as attention priors, convolutional and long short-term memory networks are used to perform time-series modeling of the risk state sequences of each center, predicting the short-term risk level of each center's future visit cycle; in the medium term, the risk state sequences of each center are embedded as center-level tokens, and cross-center risk transmission relationships are captured through cross-attention to predict risk transmission paths; in the long term, the risk state sequences of each center are dimensionality-reduced based on a covariance variational autoencoder, and a global risk association network is constructed using mutual information to assess the overall risk situation of the project. This method utilizes gated attention to fuse three-layer prediction results to obtain a comprehensive risk prediction and comprehensive risk score, and generates a ranking of risk factor contributions based on the Shapley additivity interpretation method. An inductive conformal predictor is embedded to measure the uncertainty of the prediction results, distinguishing between certain and uncertain early warnings. For certain early warnings, a three-level hierarchical alarm is triggered based on the comprehensive risk score, and after triggering the alarm, a safety reinforcement learning model incorporating conservative Q-learning constraints is used to recommend candidate countermeasures. This achieves multi-source data fusion, multi-scale prediction, and interpretable early warning for risks throughout the entire clinical trial lifecycle, significantly improving the systematic nature of risk identification and the reliability of risk early warning.

[0054] Figure 1 This is a flowchart illustrating a clinical trial risk identification and early warning method in one embodiment. It should be understood that, although... Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0055] Based on the same technical concept, referring to Figure 2 This application also provides a clinical trial risk identification and early warning device, which adopts the following technical solution: the device includes: The feature fusion module is used to receive multi-source heterogeneous data from clinical trials, extract complementary and specific risk features from each data view, and fuse the complementary and specific risk features through orthogonal constraints and adversarial training to obtain the fused risk feature vector of each trial center within each monitoring time window. The dynamic weight module is used to construct a multi-dimensional risk factor system. It calculates the risk score of each risk factor based on the fused risk feature vector and uses a variable-level self-attention mechanism to construct the dynamic correlation weights between each risk factor. The short-term risk module is used to construct a center risk state sequence by merging the risk feature vectors of each center in chronological order, using dynamic association weights as attention priors, and using convolutional and long short-term memory networks to perform temporal modeling of the center risk state sequence to obtain short-term risk level predictions. The intermediate risk module is used to embed the central risk state sequence as a central-level token, capture cross-central risk transmission relationships through cross-attention, and obtain predictions of intermediate risk transmission paths. The long-term risk module is used to reduce the dimensionality of the central risk state sequence based on the covariance variational autoencoder, construct a global risk association network using dynamic association weights, and obtain a long-term risk situation assessment. The comprehensive risk module is used to integrate short-term risk level prediction, medium-term risk transmission path prediction and long-term risk situation assessment by using gated attention to obtain comprehensive risk prediction and comprehensive risk score. The hierarchical alert module is used to embed an inductive conformal predictor to distinguish between early warning and uncertain early warning based on comprehensive risk prediction and preset confidence thresholds; for the determined early warning results, hierarchical alerts are issued based on comprehensive risk scores and preset risk thresholds.

[0056] In some embodiments, the feature fusion module is specifically used to encode and preprocess the test plan data, project operation data, safety data, quality data and external environment data respectively to obtain the initial feature representations of the five types of data views; For each type of data view, a complementary encoder and a specific encoder are constructed respectively. The initial feature representation of each view is mapped to the complementary feature subspace and the specific feature subspace respectively to obtain the complementary risk features and specific risk features of each view. Apply orthogonal constraints to complementary and specific risk features in the same view to make the two types of features orthogonal to each other and eliminate information redundancy between different views; The complementary risk features of each view are subjected to adversarial training. The complementary risk features of different views are made to converge through game between the generator and the discriminator. The average of the complementary risk features of each aligned view is then taken to obtain the shared complementary risk features. The specific risk features of each view are stitched together to obtain the fused specific risk features; By concatenating the shared complementary risk characteristics with the fusion-specific risk characteristics, a fusion risk feature vector is obtained.

[0057] In some embodiments, the dynamic weighting module is specifically used to obtain the identifiers of each risk factor in a predefined risk factor system, which covers multiple risk factors under the dimensions of safety risk, quality risk, schedule risk, compliance risk, and management risk. Based on the fused risk feature vector, the probability of occurrence and the degree of risk hazard of each risk factor within the current monitoring time window are calculated by the first multilayer perceptron. The product of the probability of occurrence and the degree of risk hazard is used as the risk score of each risk factor. The identifiers of each risk factor are mapped to risk factor embedding vectors, and the risk factor embedding vectors of each risk factor are concatenated with the fused risk feature vectors to obtain risk factor tokens. Perform a linear transformation on all risk factor tokens to generate query vector, key vector and value vector. Calculate the attention weight matrix between each risk factor by the dot product of the query vector and the key vector. Each element in the attention weight matrix represents the correlation strength between two risk factors. The value vectors of each risk factor are weighted and aggregated based on the attention weight matrix to obtain the dynamic interactive representation of each risk factor. The dynamic interaction representation of risk factors is used as a new query vector, and secondary attention calculation is performed with the key vector to obtain dynamic association weights. The dynamic association weights represent the contribution weight of each risk factor to the overall risk of the clinical trial within the current monitoring time window.

[0058] In some embodiments, the short-term risk module is specifically used to arrange the fusion risk feature vectors generated by the test center in each monitoring time window according to the time sequence of the visit cycle for each test center, so as to obtain the center risk status sequence corresponding to the test center. Construct a temporal prediction network, which includes convolutional layers, temporal prior attention layers, and long short-term memory network layers; The central risk state sequence is input into the convolutional layer, and the local spatial correlation between multidimensional risk features within each time window is extracted through one-dimensional convolution operation to obtain the convolutional risk feature sequence. The convolutional risk feature sequence is input into the temporal prior attention layer. The dynamic association weight is used as the initial attention bias of the temporal prior attention layer. The risk features at each time step in the convolutional risk feature sequence are weighted to generate the temporal weighted risk feature sequence. The time-weighted risk feature sequence is input into the long short-term memory network layer. The time-series dependency of risk state between adjacent monitoring time windows is captured by the gating mechanism of forget gate, input gate and output gate, and the hidden state sequence of each time step is obtained. The hidden state of the last time step is mapped to the preset risk level space through a fully connected classification layer to obtain the short-term risk level prediction; the short-term risk level prediction represents the risk level of each experimental center within a preset number of future visit cycles.

[0059] In some embodiments, the intermediate risk module is specifically used to map the central risk state sequence of each test center to a preset embedding dimension through a linear projection layer to obtain the central-level token vector corresponding to the test center. Assign a learnable center location code to each experimental center, and add the center location code to the corresponding center-level token vector element by element to obtain the location-aware center token; The location-aware center token of the current test center is used as the query vector, and the location-aware center tokens of the other test centers are used as the key vector and value vector, respectively. The inter-center attention weight between the current test center and each of the other test centers is calculated by the dot product of the query vector and the key vector. By using the attention weights among centers, the value vectors of the remaining test centers are weighted and aggregated to obtain the risk transmission aggregate representation of the current test center. The risk transmission aggregate representation characterizes the comprehensive impact of the risk status of the remaining test centers on the current test center. For each test center, the risk transmission aggregation representation for all test centers is calculated separately. The risk transmission aggregation representations of all test centers are then combined into a risk transmission strength matrix, where the elements represent the risk transmission strength between test centers. Based on the risk transmission intensity matrix, the risk transmission path between each test center is extracted through connectivity analysis, and the mid-term risk transmission path prediction is output. The mid-term risk transmission path prediction represents the path and intensity of risk transmission from one test center to other test centers within a preset time period in the future.

[0060] In some embodiments, the long-term risk module is specifically used to construct a covariance variational autoencoder, which includes an encoder and a decoder; the encoder performs nonlinear dimensionality reduction on the input central risk state sequence to obtain the mean vector and covariance matrix of each test center; the covariance matrix represents the correlation between multidimensional risk features within each test center; Gaussian distribution assumptions are applied to the mean vector and covariance matrix, and latent variables are generated from the Gaussian distribution through reparameterized sampling. The latent variables are input into the decoder for reconstruction to obtain the reconstructed risk state sequence. The covariance variational autoencoder is trained by minimizing the reconstruction error between the central risk state sequence and the reconstructed risk state sequence. A variable-level attention module is embedded in the encoder to model the time series of each risk factor as an independent semantic unit in order to capture the interaction dependencies of each risk factor in the global time series. The trained covariance variational autoencoder is used to reduce the dimensionality of the central risk state sequence of each test center, resulting in the dimensionality-reduced mean vector and dimensionality-reduced covariance matrix of each test center. Based on the dimensionality-reduced covariance matrix, the joint entropy and mutual information values ​​between the test centers are calculated under a Gaussian distribution to obtain the risk correlation matrix between centers. The risk correlation matrix between centers is weighted and fused with the dynamic correlation weights to generate a fused correlation matrix; Based on the fusion correlation matrix, a global risk correlation network is constructed with each test center as a node and the fusion correlation as the edge weight. Network analysis is performed on the global risk correlation network. The central nodes connected by the edges whose correlation exceeds a preset first correlation threshold are identified as key risk nodes. The edges connected to the nodes whose correlation exceeds a preset second correlation threshold and whose structural importance exceeds a preset structural threshold are identified as vulnerable links. Based on key risk nodes and vulnerable links, a long-term risk situation assessment is obtained; the long-term risk situation assessment indicates the risk clustering areas and vulnerable risk transmission paths within the overall project scope.

[0061] In some embodiments, the integrated risk module is specifically used to map short-term risk level prediction to a preset fusion dimension through a first linear projection layer to obtain a short-term risk embedding vector; to map medium-term risk transmission path prediction to a fusion dimension through a second linear projection layer to obtain a medium-term risk embedding vector; and to map long-term risk situation assessment to a fusion dimension through a third linear projection layer to obtain a long-term risk embedding vector. The short-term risk embedding vector, the medium-term risk embedding vector, and the long-term risk embedding vector are stacked to form a three-layer risk representation matrix; A gated attention fusion module is constructed, which performs a dot product between the risk embedding vector of each layer in the three-layer risk representation matrix and the shared context vector, and calculates the corresponding gating coefficients of each layer through an activation function; at the same time, self-attention is calculated on the three-layer risk representation matrix to obtain the corresponding attention weights of each layer. The gating coefficients of each layer are multiplied element-wise with the attention weights to obtain the fusion modulation weights of each layer; By using the fusion modulation weights to sum the risk embedding vectors of each layer, a comprehensive risk fusion vector is obtained; The comprehensive risk fusion vector is input into the classification output head and the scoring output head respectively; the classification output head maps the comprehensive risk fusion vector to a preset risk level space and outputs the comprehensive risk prediction; the scoring output head maps the comprehensive risk fusion vector into a continuous value and outputs the comprehensive risk score. Using the comprehensive risk fusion vector, the marginal contribution value of each risk factor to the comprehensive risk prediction is calculated through the Shapley additivity interpretation method. The marginal contribution values ​​are then arranged in descending order to generate a ranking of risk factor contributions.

[0062] This application also discloses a control device.

[0063] Specifically, the control device includes a memory and a processor, with the memory storing a computer program that can be loaded by the processor and executed to implement the aforementioned clinical trial risk identification and early warning method.

[0064] This application also discloses a computer-readable storage medium.

[0065] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the clinical trial risk identification and early warning method described above. The computer-readable storage medium includes, for example, various media that can store program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for identifying and warning of risks in clinical trials, characterized in that, include: The system receives multi-source heterogeneous data from clinical trials, extracts complementary and specific risk features from each data view, and fuses these complementary and specific risk features through orthogonal constraints and adversarial training to obtain a fused risk feature vector for each trial center within each monitoring time window. A multi-dimensional risk factor system is constructed, and the risk score of each risk factor is calculated based on the fused risk feature vector. The dynamic correlation weight between each risk factor is constructed using a variable-level self-attention mechanism. The fused risk feature vectors of each center are constructed into a center risk state sequence in chronological order. The dynamic association weights are used as attention priors. Convolutional and long short-term memory networks are used to perform temporal modeling on the center risk state sequence to obtain short-term risk level predictions. The central risk state sequence is embedded as a central-level token, and cross-central risk transmission relationships are captured through cross-attention to obtain mid-term risk transmission path predictions; The central risk state sequence is reduced in dimensionality based on the covariance variational autoencoder, and a global risk association network is constructed using the dynamic association weights to obtain a long-term risk situation assessment. By using gated attention to fuse the short-term risk level prediction, the medium-term risk transmission path prediction, and the long-term risk situation assessment, a comprehensive risk prediction and a comprehensive risk score are obtained. An embedded inductive conformal predictor is used to distinguish between early warning and uncertain early warning based on the comprehensive risk prediction and a preset confidence threshold. For confirmed early warning results, tiered alerts are issued based on the comprehensive risk score and preset risk thresholds.

2. The clinical trial risk identification and early warning method according to claim 1, characterized in that, The multi-source heterogeneous data from received clinical trials are used to extract complementary and specific risk features from each data view. These complementary and specific risk features are then fused through orthogonal constraints and adversarial training to obtain a fused risk feature vector for each trial center within each monitoring time window, including: The experimental plan data, project operation data, safety data, quality data, and external environment data were coded and preprocessed to obtain the initial feature representations of the five types of data views; For each type of data view, a complementary encoder and a specific encoder are constructed respectively, and the initial feature representation of each view is mapped to the complementary feature subspace and the specific feature subspace respectively to obtain the complementary risk feature and the specific risk feature of each view; Apply orthogonal constraints to the complementary risk features and the specific risk features of the same view to make the two types of features orthogonal to each other, thereby eliminating information redundancy between different views; The complementary risk features of each view are subjected to adversarial training. The complementary risk features of different views are made to be consistent through game between the generator and the discriminator. The complementary risk features of the aligned views are averaged to obtain the shared complementary risk features. The specific risk features of each view are stitched together to obtain the fused specific risk features; The shared complementary risk features are concatenated with the fusion-specific risk features to obtain the fusion risk feature vector.

3. The clinical trial risk identification and early warning method according to claim 2, characterized in that, The construction of a multi-dimensional risk factor system, based on the fused risk feature vector, calculates the risk score of each risk factor, and utilizes a variable-level self-attention mechanism to construct dynamic correlation weights between risk factors, including: Obtain the identifiers of each risk factor in the predefined risk factor system, which covers multiple risk factors under the dimensions of safety risk, quality risk, schedule risk, compliance risk, and management risk. Based on the fused risk feature vector, the probability of occurrence and the degree of risk hazard of each risk factor within the current monitoring time window are calculated by the first multilayer perceptron. The product of the probability of occurrence and the degree of risk hazard is used as the risk score of each risk factor. The identifiers of each risk factor are mapped to risk factor embedding vectors, and the risk factor embedding vectors of each risk factor are concatenated with the fused risk feature vectors to obtain risk factor tokens. A linear transformation is performed on all risk factor tokens to generate query vectors, key vectors, and value vectors. The attention weight matrix between each risk factor is calculated by the dot product of the query vector and the key vector. Each element in the attention weight matrix represents the correlation strength between two risk factors. Based on the attention weight matrix, the value vectors of each risk factor are weighted and aggregated to obtain the dynamic interactive representation of each risk factor. The dynamic interaction representation of the risk factors is used as a new query vector, and secondary attention is calculated with the key vector to obtain the dynamic association weight. The dynamic association weight represents the contribution weight of each risk factor to the overall risk of the clinical trial within the current monitoring time window.

4. The clinical trial risk identification and early warning method according to claim 1, characterized in that, The process involves constructing a center risk state sequence from the fused risk feature vectors of each center in chronological order, using the dynamic association weights as attention priors, and employing convolutional and long short-term memory networks to perform temporal modeling on the center risk state sequence to obtain short-term risk level predictions. This includes: For each test center, the fusion risk feature vectors generated by the test center in each monitoring time window are arranged in chronological order of the visit cycle to obtain the center risk status sequence corresponding to the test center; Construct a temporal prediction network, which includes convolutional layers, temporal prior attention layers, and long short-term memory network layers; The central risk state sequence is input into the convolutional layer, and the local spatial correlation between multidimensional risk features within each time window is extracted through one-dimensional convolution operation to obtain the convolutional risk feature sequence. The convolutional risk feature sequence is input into the temporal prior attention layer, and the dynamic association weight is used as the initial attention bias of the temporal prior attention layer. The risk features of each time step in the convolutional risk feature sequence are weighted to generate a temporally weighted risk feature sequence. The time-weighted risk feature sequence is input into the long short-term memory network layer. The time-series dependency of risk state between adjacent monitoring time windows is captured through the gating mechanism of forget gate, input gate and output gate, and the hidden state sequence of each time step is obtained. The hidden state of the last time step is mapped to a preset risk level space through a fully connected classification layer to obtain the short-term risk level prediction; the short-term risk level prediction represents the risk level of each test center within a preset number of future visit cycles.

5. The clinical trial risk identification and early warning method according to claim 4, characterized in that, The step of embedding the central risk state sequence as a central-level token, capturing cross-central risk transmission relationships through cross-attention, and obtaining a prediction of the medium-term risk transmission path includes: For each test center, the sequence of center risk states of the test center is mapped to a preset embedding dimension through a linear projection layer to obtain the center-level token vector corresponding to the test center; Each test center is assigned a learnable center location code, and the center location code is added element by element to the corresponding center-level token vector to obtain a location-aware center token. The location-aware center token of the current test center is used as the query vector, and the location-aware center tokens of the other test centers are used as the key vector and value vector, respectively. The inter-center attention weight between the current test center and each of the other test centers is calculated by the dot product of the query vector and the key vector. The value vectors of the remaining test centers are weighted and aggregated using the inter-center attention weights to obtain the risk transmission aggregate representation of the current test center. The risk transmission aggregate representation characterizes the comprehensive impact of the risk status of the remaining test centers on the current test center. Calculations are performed for each test center to obtain the risk transmission aggregate representation for all test centers. The risk transmission aggregate representations of all test centers are combined into a risk transmission strength matrix, where the elements of the risk transmission strength matrix represent the risk transmission strength between test centers. Based on the risk transmission intensity matrix, the risk transmission path between each test center is extracted through connectivity analysis, and the mid-term risk transmission path prediction is output; the mid-term risk transmission path prediction represents the path and intensity of risk transmission from one test center to other test centers within a preset time period in the future.

6. The clinical trial risk identification and early warning method according to claim 5, characterized in that, The method of reducing the dimensionality of the central risk state sequence based on the covariance variational autoencoder and constructing a global risk association network using the dynamic association weights to obtain a long-term risk situation assessment includes: A covariance variational autoencoder is constructed, which includes an encoder and a decoder; the encoder performs nonlinear dimensionality reduction on the input central risk state sequence to obtain the mean vector and covariance matrix of each test center; the covariance matrix represents the correlation between multidimensional risk features within each test center; Gaussian distribution assumptions are applied to the mean vector and covariance matrix, and latent variables are generated from the Gaussian distribution through reparameterized sampling. The latent variables are input into the decoder for reconstruction to obtain the reconstructed risk state sequence. The covariance variational autoencoder is trained by minimizing the reconstruction error between the central risk state sequence and the reconstructed risk state sequence. A variable-level attention module is embedded in the encoder to model the time series of each risk factor as an independent semantic unit in order to capture the interaction dependencies of each risk factor in the global time series. The trained covariance variational autoencoder is used to reduce the dimensionality of the central risk state sequence of each test center to obtain the dimensionality-reduced mean vector and dimensionality-reduced covariance matrix of each test center; based on the dimensionality-reduced covariance matrix, the joint entropy and mutual information value between each test center are calculated under Gaussian distribution to obtain the risk correlation matrix between centers. The inter-center risk correlation matrix is ​​weighted and fused with the dynamic correlation weight to generate a fused correlation matrix. Based on the aforementioned fusion correlation matrix, a global risk correlation network is constructed with each test center as a node and the fusion correlation as the edge weight. Network analysis is performed on the global risk association network to identify the central nodes connected by edges whose association degree exceeds a preset first association threshold as key risk nodes, and to identify the edges connected to nodes whose association degree exceeds a preset second association threshold and whose structural importance exceeds a preset structural threshold as vulnerable links. Based on the key risk nodes and the vulnerable links, the long-term risk situation assessment is obtained; the long-term risk situation assessment represents the risk clustering areas and vulnerable risk transmission paths within the overall scope of the project.

7. The clinical trial risk identification and early warning method according to claim 1, characterized in that, The method of fusing the short-term risk level prediction, the medium-term risk transmission path prediction, and the long-term risk situation assessment using gated attention yields a comprehensive risk prediction and a comprehensive risk score, including: The short-term risk level prediction is mapped to a preset fusion dimension through a first linear projection layer to obtain a short-term risk embedding vector; the medium-term risk transmission path prediction is mapped to a fusion dimension through a second linear projection layer to obtain a medium-term risk embedding vector; and the long-term risk situation assessment is mapped to a fusion dimension through a third linear projection layer to obtain a long-term risk embedding vector. The short-term risk embedding vector, the medium-term risk embedding vector, and the long-term risk embedding vector are stacked to form a three-layer risk representation matrix; A gated attention fusion module is constructed, which performs a dot product between the risk embedding vector of each layer in the three-layer risk representation matrix and the shared context vector, and calculates the corresponding gating coefficients of each layer through an activation function; at the same time, self-attention is calculated on the three-layer risk representation matrix to obtain the attention weights of each layer. The gating coefficients of each layer are multiplied element-wise with the attention weights to obtain the fusion modulation weights of each layer; The risk embedding vectors of each layer are weighted and summed using the fusion modulation weights to obtain the comprehensive risk fusion vector; The comprehensive risk fusion vector is input into the classification output head and the scoring output head respectively; the classification output head maps the comprehensive risk fusion vector to a preset risk level space and outputs the comprehensive risk prediction; the scoring output head maps the comprehensive risk fusion vector into a continuous value and outputs the comprehensive risk score. Using the comprehensive risk fusion vector, the marginal contribution value of each risk factor to the comprehensive risk prediction is calculated using the Shapley additivity interpretation method. The marginal contribution values ​​are then arranged in descending order to generate a risk factor contribution ranking.

8. A clinical trial risk identification and early warning device, characterized in that, The device includes: The feature fusion module is used to receive multi-source heterogeneous data from clinical trials, extract complementary risk features and specific risk features from each data view, and fuse the complementary risk features and specific risk features through orthogonal constraints and adversarial training to obtain the fused risk feature vector of each trial center within each monitoring time window. The dynamic weighting module is used to construct a multi-dimensional risk factor system, calculate the risk score of each risk factor based on the fused risk feature vector, and construct the dynamic correlation weights between each risk factor using a variable-level self-attention mechanism. The short-term risk module is used to construct a center risk state sequence by arranging the fused risk feature vectors of each center in chronological order, use the dynamic association weight as an attention prior, and use convolutional and long short-term memory networks to perform temporal modeling on the center risk state sequence to obtain a short-term risk level prediction. The intermediate risk module is used to embed the central risk state sequence into a central-level token, capture cross-central risk transmission relationships through cross-attention, and obtain intermediate risk transmission path predictions. The long-term risk module is used to reduce the dimensionality of the central risk state sequence based on the covariance variational autoencoder, construct a global risk association network using the dynamic association weights, and obtain a long-term risk situation assessment. The comprehensive risk module is used to fuse the short-term risk level prediction, the medium-term risk transmission path prediction, and the long-term risk situation assessment using gated attention to obtain a comprehensive risk prediction and a comprehensive risk score. The hierarchical alarm module is used to embed an inductive conformal predictor to distinguish between early warning and uncertain early warning based on the comprehensive risk prediction and a preset confidence threshold; for the determined early warning result, hierarchical alarms are issued based on the comprehensive risk score and a preset risk threshold.

9. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images