Depth time sequence clustering enhancement-based disease deterioration risk identification method and system

Abstract

The invention relates to the technical field of clinical medical treatment, and discloses a disease deterioration risk identification method and system based on depth time sequence clustering enhancement, and the method comprises the steps: 1, obtaining multi-modal clinical sequence data of a patient, including physiological indexes of a time sequence, a laboratory detection result, historical diseases and medication data; 2, performing feature extraction and classification on the patient sequences by adopting a knowledge enhanced sequence clustering method, and grouping the patient sequences according to future outcome distribution of the patient sequences; 3, enabling the model to quickly adapt to a prediction task of a new patient subgroup through a meta-training process; and step 4, based on the trained meta-model, carrying out rapid adaptation on the new patient subtype, and predicting the possibility that the new patient subtype has a deterioration event in a certain time window in the future. The method and the system can effectively learn the disease change mode of the patient under the condition of limited clinical data, improve the prediction accuracy of the new patient subgroup, and are especially suitable for clinical prediction scenes under the condition of small samples.

Description

Technical Field

[0001] This invention relates to the field of clinical medical technology, specifically to a method and system for identifying the risk of disease deterioration based on deep temporal clustering enhancement. Background Technology

[0002] From a macro-technical perspective, monitoring the risk of deterioration in critically ill patients is a core challenge in clinical medicine. Traditional manual observation and threshold alarm schemes suffer from problems such as response lag and high false alarm rates. Although early warning scales such as MEWS and NEWS have been proposed in clinical practice, their ability to comprehensively process multi-source heterogeneous data is limited, making it difficult to support accurate prediction of disease deterioration. From a micro-technical perspective, existing machine learning-based early warning systems can integrate some dynamic physiological parameters (such as the time trend of blood pressure and blood oxygen), but when faced with heterogeneous multimodal data such as "electronic medical records + real-time monitoring" and small sample training scenarios, the generalization and robustness of the models are insufficient. At the same time, there is a lack of an effective stratification mechanism for individual patient heterogeneity, thus limiting the representation and prediction of complex dynamic risks. Existing research has developed three representative technical approaches in the field of intelligent medical risk prediction: Hyland et al., in *Nature Digital Medicine* (DOI: 10.1038 / s41746-020-00335-x), proposed a circulatory failure early warning system based on a high-resolution ICU database. This system selects 26 key physiological characteristics, including heart rate variability and blood pressure fluctuation coefficient, and uses an LSTM time-series network to capture dynamic trends to output the probability of circulatory failure risk. Compared to the traditional threshold method, this reduces the false alarm rate by 37%. Choi et al., in KDD 2017 (10.1145 / ...), proposed a similar system. The knowledge-enhanced disease prediction model proposed by 3097983.3098126 injects the hierarchical structure of ICD-10 disease coding into a graph attention network (GAT) and combines it with electronic medical record data to achieve joint prediction of multiple diseases, thereby strengthening the representation learning and cross-disease reasoning ability of rare diseases; while the MetaPred meta-learning risk prediction framework proposed by the Cornell University team in JAMIA (2020;27(3):431–440) is based on the model-independent meta-learning (MAML) algorithm to build a task-adaptive meta-learner, which achieves rapid adaptation and cross-task generalization on neurodegenerative disease data with small samples; While existing technologies have made some progress in predicting the risk of deterioration in critically ill patients, they still have several limitations, including the following: First, learning with small samples is challenging. Traditional deep learning models (such as LSTM-based temporal networks) rely on large-scale labeled data. When the sample size is limited, the model is prone to overfitting and struggles to learn stable temporal patterns and disease evolution characteristics. In extreme medical environments, critical illness data is scarce and noisy. Existing parameter sharing or transfer learning methods can only be effective on fixed tasks and lack a unified learning framework that can adapt to multiple tasks and small sample sizes simultaneously. Secondly, there is insufficient handling of patient heterogeneity. Critically ill patients exhibit significant differences, and the risk evolution patterns corresponding to different pathological mechanisms (such as cardiogenic shock and sepsis) differ fundamentally. While existing studies have attempted to subgroup patients using a two-stage approach of "clustering first, then prediction," the clustering process is separated from the prediction objective, failing to ensure consistency between the resulting subgroups and the risk identification task. This makes it difficult for the model to focus on the truly discriminative feature space. Third, the model's cross-domain generalization ability is weak. In different medical environments (such as inter-hospital, regional, battlefield, or disaster scenarios), the data distribution and monitoring conditions vary greatly, leading to a significant performance degradation of models trained in conventional scenarios in new domains. Existing technologies lack a systematic generalization enhancement mechanism and fail to actively reduce bias towards specific tasks during the training phase, resulting in insufficient adaptability of the model to new tasks and new populations. Summary of the Invention

[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention proposes a method and system for identifying the risk of disease deterioration based on deep temporal clustering enhancement, aiming to achieve early and accurate identification of disease deterioration in critically ill patients. Specific objectives include: (1) By introducing a meta-learning framework, rapid convergence and knowledge transfer can be achieved under conditions of limited data, thereby improving the learning efficiency and robustness of the model; (2) By using deep clustering mechanism to automatically identify homogeneous subgroups of patients, the model can adopt differentiated prediction strategies for different subtypes, thereby improving the overall risk identification accuracy. (3) Embed unbiased strategies (such as entropy regularization) during training to prevent the model from becoming overly dependent on the training task, thereby obtaining initial parameters that are balanced and generalizable for multiple tasks and enhancing cross-scenario adaptability. (4) Construct a knowledge graph of critical illness and embed prior knowledge such as changes in key vital signs and intervention events into the model through graph neural network so that it can still infer the evolution trend of the disease based on knowledge when the data is incomplete or missing.

[0004] In summary, this invention aims to establish a unified risk prediction framework for scenarios with limited data and high heterogeneity, achieving synergistic optimization of "dynamic clustering - knowledge enhancement - meta-learning generalization", and significantly improving the accuracy, robustness and interpretability of disease exacerbation risk identification.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method for identifying the risk of disease deterioration based on deep temporal clustering enhancement, comprising the following steps: Step 1: Obtain the patient's multimodal clinical sequence data, including time-series physiological indicators, laboratory test results, historical disease and medication data; Step 2: Use a knowledge-enhanced sequence clustering method to extract features and classify patient sequences, grouping them according to their future outcome distribution; Step 3: Through the meta-training process, the model can be quickly adapted to the prediction task of new patient subgroups; Step 4: Based on the trained meta-model, rapidly adapt the new patient subtype to predict the likelihood of disease exacerbation events occurring within a certain time window in the future.

[0006] A disease exacerbation risk identification system based on deep temporal clustering enhancement is applied to a disease exacerbation risk identification method based on deep temporal clustering enhancement, including: Data preprocessing module: Receives raw data from multiple sources, including hospital information systems and monitoring equipment, and performs cleaning, alignment, and standardization operations to generate structured time series and event sets; Knowledge enhancement module: It has a built-in critical care medicine knowledge graph database and includes an R-GCN graph network inference engine. It queries the events output by the preprocessing module, obtains the knowledge subgraph of the corresponding entity, and calculates the entity embedding vector. Clustering module: Implements deep clustering functionality, including a time series encoder and a clustering computation unit. It receives preprocessed data and event embeddings provided by knowledge enhancement, computes patient representations, and performs clustering operations. Prediction module: This is the disease deterioration risk prediction model module, which is connected to the clustering module. It takes the patient's fused representation or the representation combined with cluster information as input and outputs the disease risk prediction result. Model training module: used for offline training of the parameters of the clustering module and prediction module mentioned above; internally encapsulates meta-learning training logic, including task partitioning management, inner and outer loop optimizers, and task unbiased strategy; Risk identification module: An application-layer module for end users, which combines the output of the prediction module and rule strategies to provide intuitive risk identification results.

[0007] Preferably, the data preprocessing module acquires multi-source heterogeneous clinical data from multiple sources, including electronic medical records, monitor records, and laboratory information systems, including continuous monitoring time series data and discrete key event data, and uniformly completes missing value imputation, anomaly correction, time alignment, and feature normalization to generate standardized input tensors that can be directly used by the model, including; Multivariate time series Each of them For the first The patient was at a moment The multidimensional physiological feature vector; Key Events Collection Each event Includes event type and timestamp of occurrence. .

[0008] Preferably, the specific operation steps of the knowledge enhancement module are as follows: S2.1, will Encoding as a fixed-dimensional hidden representation Capture time-dependent dynamics; S2.2 Knowledge Graph Embedding: Relationship-aware embedding inference of entities is performed using a Relationship Graph Convolutional Network (R-GCN), and finally aggregated into a weighted average to form a holistic event knowledge vector for the patient. ; The node representation update rule of R-GCN is as follows: In R-GCN, the node embeddings of each layer are updated using the following formula to achieve information aggregation of relational dependencies: ; In the formula, Represents a node In the Layer embedding vector; Represents the set of all relation types in a knowledge graph; Indicates the relationship Connect to node The set of neighboring nodes; Indicates the number of elements in the neighbor node set; Representing relation type In the Layer weight matrix; The weight matrix represents the self-loop; Indicates the activation function; By running multiple layers of R-GCN on the knowledge graph, a low-dimensional vector representation is ultimately learned. ; S2.3 Representing the time series With knowledge graph embedding A comprehensive representation is obtained by fusion through splicing or attention mechanisms. .

[0009] Preferably, the clustering module operates as follows: Calculate soft cluster distribution Used for clustering loss Calculation; Optimization is performed using a clustering objective function based on Kullback-Leibler divergence: Model Specification Cluster centers The initial values ​​can be initialized either randomly or by the latent representation output by a pre-trained autoencoder, for each patient's embedding representation. Calculate its belonging to the first The soft assignment probability of each cluster Adopting students Distribution as a similarity measure: ; in, Indicates the first High-dimensional feature embedding vectors of each patient; For the first Cluster centers; The distance is Euclidean. To accelerate cluster convergence and improve intra-cluster consistency, a target distribution is introduced, where the student distribution is a free parameter. By adjusting the current soft allocation Reweight the samples to amplify the weights of high-confidence samples: ; The optimization goal of the clustering module is to enable soft allocation. As close as possible to the target distribution KL divergence is used as the loss function: ; By minimizing To iteratively update the encoder and cluster center parameters, after training convergence, each patient represents... Assign a cluster label After training convergence, the patients belong to the homogeneous subgroup.

[0010] Preferably, the prediction module is for each cluster. Build a dedicated risk prediction sub-model The specific steps are as follows: S4.1 Shared Encoder Weights: All sub-models share the parameters of the underlying encoder, ensuring a smooth transition from raw input to latent representation. The extraction process remains consistent: Encoder output Representing the High-dimensional health status representation of individual patients; All sub-models use the same As input, semantic biases caused by extraction from different paths are avoided; S4.2, Cluster-Specific Prediction Head: Each Sub-model It includes independent fully connected layers and a classifier to generate the final risk prediction result, as shown in the formula: ; in, For the first The weight matrix and bias terms of each sub-model For the number of disease categories, For the sample In the The probability distribution of disease risk prediction under each cluster is used; therefore, the final cross-entropy loss function is: ; in, The total number of training samples, No. The true label of each sample sample In its own cluster The logits value output in the sub-model. For patient samples Clustering labels.

[0011] Preferably, the prediction module jointly optimizes the prediction loss and clustering loss, and the total loss function of the model is a weighted sum of the clustering loss and the prediction loss: ; in, To balance the coefficients, the impact of clustering objectives on overall training is adjusted.

[0012] Preferably, the task partitioning management in the model training module includes: Cluster-driven task partitioning: based on the hard labels output by the clustering layer Divide the samples into A cluster subset Each subset Patient groups containing similar characteristics, each cluster subset It can be regarded as an independent task Its goal is to learn risk prediction patterns specific to this subgroup; External attribute extension: Based on the clustering task, the task is further divided according to external attributes; Supports set and query set partitioning: for each task Randomly divided into support sets and query set .

[0013] Preferably, the inner and outer loop optimizers in the model training module include: Inner loop phase: using globally shared parameters As the initial state, for each subtask ,in This represents the patient subgroups identified by clustering. External attributes are represented in the following form: ; in, Indicates in the task The task-specific parameters after one gradient update For the inner loop learning rate, For the task joint loss function For global parameters The gradient of is defined as: ; That is, both the accuracy of risk prediction and the consistency of patient temporal clustering are considered simultaneously; After the inner loop is completed, the outer loop phase begins, where each task is calculated. Use the updated parameters on its corresponding query set loss And minimize the average loss of all tasks through backpropagation: ; in, Indicates task Prediction performance on query sets.

[0014] Preferably, the task-unbiased strategy in the model training module includes: To balance task contributions through entropy regularization, a task entropy loss is added to the meta-loss. for: ; in, For the model in the task The output class probability distribution on the validation set; After adding task entropy regularization, the objective of meta-training becomes: ; in, These are the weighting coefficients.

[0015] Compared with existing technologies, this invention provides a method and system for identifying the risk of disease exacerbation based on deep temporal clustering enhancement, which has the following beneficial effects: Multi-step collaborative design for clustering and prediction: The patient subgroup clustering process and the risk prediction model training are optimized step-by-step in a collaborative manner. This design enables the model to automatically learn the clustering patterns most beneficial to the prediction task. The clustering results directly serve the prediction, making each identified patient subgroup show significant discrimination in terms of deterioration risk, thus improving the overall discriminative ability of the model; Task-unbiased meta-learning strategy: An improved meta-learning algorithm is employed, introducing task entropy regularization into the meta-loss to ensure that the initial model is not overly biased towards the training task set. The initial parameters of the model have greater uncertainty and generalization ability, allowing it to quickly adapt to new task environments in the future. This unbiased meta-learning strategy effectively improves the model's transfer performance in cross-hospital and cross-population scenarios; Task entropy regularization: An entropy regularization term added during meta-learning training encourages the model to achieve a more uniform output probability distribution across tasks. By maximizing prediction entropy, it prevents the model from becoming overconfident on certain tasks and getting trapped in local optima, thus achieving fair learning across all tasks. This technique ensures the neutrality of the meta-model, resulting in fewer iterations and better performance during fine-tuning. Key Event Graph Enhancement: A medical knowledge graph for critical care is constructed, covering relationships such as "abnormal indicators – changes in condition – treatment measures." R-GCN is used to encode the knowledge graph, and the resulting events are embedded to enhance patient representations. This technique enables the model to be driven by both "data + knowledge": it not only learns from patient data patterns but also infers risks based on medical knowledge. In particular, the ability to identify rare key events is significantly improved, enhancing the model's reliability even with incomplete data. Few-Shot Fast Transfer Mechanism: A Few-Shot task transfer scheme based on a meta-learning framework. The model simulates learning from different tasks during the training phase, enabling it to acquire internalized rapid adaptability. When deployed to new scenarios, only a minimal amount of new data samples are needed for rapid model fine-tuning, eliminating the need for retraining from scratch. This mechanism significantly improves the model's usability in extreme medical scenarios, achieving true "train once, deploy everywhere" capabilities. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the steps of the method of the present invention; Figure 2 This is a schematic diagram of the system modules of the present invention; Figure 3 This is a schematic diagram of a task-unbiased meta-learning framework based on deep temporal clustering enhancement. Figure 4 This is a schematic diagram of a disease prediction module based on knowledge-enhanced deep temporal clustering. Figure 5 This is a schematic diagram illustrating the embedding and aggregation of knowledge graphs for key events. Detailed Implementation

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

[0018] Please see Figures 1-4 A method for identifying the risk of disease deterioration based on deep temporal clustering includes the following steps: Step 1: Obtain the patient's multimodal clinical sequence data, including time-series physiological indicators, laboratory test results, historical disease and medication data; Step 2: Use a knowledge-enhanced sequence clustering method to extract features and classify patient sequences, grouping them according to their future outcome distribution; Step 3: Through the meta-training process, the model can be quickly adapted to the prediction task of new patient subgroups; Step 4: Based on the trained meta-model, rapidly adapt to the new patient subtype and predict the likelihood of disease exacerbation events occurring within a certain time window in the future. The above-mentioned method for identifying the risk of disease exacerbation based on deep temporal clustering enhancement has the following application systems: Data preprocessing module: Receives raw data from multiple sources including hospital information systems and monitoring equipment, performs cleaning, alignment, and normalization operations, and generates structured time series and event sets. Specific operation steps are as follows: Acquire multi-source heterogeneous clinical data from multiple sources, including electronic medical records, monitor records, and laboratory information systems, including continuous monitoring time series data and discrete key event data; uniformly complete missing value imputation, anomaly correction, time alignment, and feature normalization to generate standardized input tensors that can be directly used by the model; Multivariate time series Each of them For the first The patient was at a moment The multidimensional physiological feature vector; Key Events Collection Each event Includes event type and timestamp of occurrence. ; The above data is imputed for missing values, corrected for anomalies, time-aligned, and feature-normalized to generate a standardized input tensor. Knowledge Enhancement Module: This module includes a built-in critical care medicine knowledge graph database and an R-GCN graph network inference engine. It queries the events output by the preprocessing module to obtain the corresponding entity knowledge subgraphs and calculates entity embedding vectors. Specifically: S2.1, Using models such as Transformer or LSTM, ... Encoding as a fixed-dimensional hidden representation Capture time-dependent dynamics; S2.2, Embedding Knowledge Graph: Please see Figure 5 , Figure 5 This demonstrates the embedding learning and aggregation process of a key event knowledge graph, in which... Figure 5 The left part is an example subgraph of the critical care medicine knowledge graph, which takes "acute asthma" as the core and establishes its relationship with entities such as tiotropium bromide, prednisolone, symptomatic cough, high heart rate, and complication respiratory failure, forming a structured medical knowledge network. Figure 5 The right side presents the embedding and generation logic of key events—classifying knowledge graph entities such as drugs and symptoms, performing relation-aware embedding inference on entities through a relational graph convolutional network (R-GCN), and finally aggregating them into a holistic event knowledge vector for the patient through weighted averaging. The node representation update rule of R-GCN is as follows: In R-GCN, the node embeddings of each layer are updated using the following formula to achieve information aggregation of relational dependencies: ; In the formula, Represents a node In the Layer embedding vectors (e.g.) (This refers to the initial embedding of a node). Represents the set of all relation types in the knowledge graph (such as "treatment", "symptoms", "complications"); Indicates the relationship Connect to node The set of neighboring nodes (e.g., the neighbors of "acute asthma" under the "treatment" relationship include "tiotropium bromide" and "prednisolone"); This represents the number of elements in the neighbor node set, used to average neighbor information and avoid information dilution. Representing relation type In the The layer's weight matrix is ​​used to capture semantic dependencies under different relations (e.g., the contribution of the "treatment" relation and the "complications" relation to node embedding are different). The weight matrix represents the self-loop and is used to preserve the feature information of the node itself; Represents the activation function (e.g.) This introduces nonlinear modeling capabilities into the model; By running multiple layers of R-GCN on the knowledge graph, each medical entity node (such as drug, symptom, disease) can gradually fuse information from its neighboring nodes, ultimately learning a low-dimensional vector representation. This vector retains the characteristics of the entity itself while incorporating the semantic information of related entities (for example, the embedding of "acute asthma" will incorporate related information such as its treatment drugs, symptoms, and complications). S2.3 Representing the time series With knowledge graph embedding A comprehensive representation is obtained by fusion through splicing or attention mechanisms. ; Clustering module: Implements deep clustering functionality, including a time series encoder and a clustering computation unit. It receives preprocessed data and event embeddings provided by knowledge enhancement, computes patient representations, and performs clustering operations. To achieve automatic grouping and homogeneity modeling of patient populations, this invention employs an adaptive soft clustering module based on Kullback-Leibler divergence. This module dynamically discovers patient subgroups with similar physiological characteristics and disease evolution patterns by jointly optimizing the encoder and cluster centers, thereby providing structured prior support for subsequent personalized risk prediction. Calculate soft cluster distribution Used for clustering loss Calculation; Optimization is performed using a clustering objective function based on Kullback-Leibler divergence: Model Specification Cluster centers The initial values ​​can be initialized either randomly or by the latent representation output by a pre-trained autoencoder, for each patient's embedding representation. Calculate its belonging to the first The soft assignment probability of each cluster Adopting students Distribution as a similarity measure: ; in, Indicates the first The high-dimensional feature embedding vector of each patient (generated by the fusion of time series encoder and knowledge graph) reflects their overall health status; For the first Each cluster center represents a typical prototype of a certain patient subgroup (e.g., "elderly patients with COPD and heart failure"). The Euclidean distance measures the difference between a sample and its cluster center in the feature space. Free parameters for the student distribution (usually) To accelerate cluster convergence and improve intra-cluster consistency, a target distribution is introduced. By adjusting the current soft allocation Reweight the samples to amplify the weights of high-confidence samples: ; The objective distribution increases the weight of high-confidence samples, thereby "sharpening" the assignment. The optimization objective of the clustering module is to make the soft assignment more effective. As close as possible to the target distribution KL divergence is used as the loss function: ; By minimizing The encoder and cluster center parameters are iteratively updated. This adaptive clustering optimization process gradually improves intra-cluster compactness and inter-cluster separation. After training convergence, each patient represents... Assign a cluster label After training convergence, the homogeneous subgroup to which the patient belongs; Prediction module: This is the disease deterioration risk prediction model module, which is connected to the clustering module. It takes the patient's fused representation or the representation combined with cluster information as input and outputs the disease risk prediction result. After completing the soft clustering of patient embedding representations, this invention designs a cluster-independent prediction mechanism, that is, for each cluster... Build a dedicated risk prediction sub-model This design fully considers the heterogeneity of different patient subgroups in terms of pathological mechanisms, disease progression rate and treatment response, and realizes a medical paradigm of "co-treatment for the same group and different strategies for different groups". S4.1 Shared Encoder Weights: All sub-models share the parameters of the underlying encoder, ensuring that the raw input (such as time-series physiological signals, electronic medical record text) is transformed into a latent representation. The extraction process remains consistent: encoder output Representing the A high-dimensional health status representation of each patient; all sub-models use the same... As input, semantic biases caused by extraction from different paths are avoided; S4.2 Cluster-Specific Prediction Head: Although shared by the encoder, each cluster has a specific prediction head. Each sub-model corresponds to a separate fully connected classification head (i.e., prediction head). It includes independent fully connected layers and a classifier to generate the final risk prediction result, as shown in the formula: ; in, For the first The weight matrix and bias terms of each sub-model For the number of disease categories, For the sample In the The probability distribution of disease risk prediction under each cluster. Therefore, the final cross-entropy loss function is: ; in, The total number of training samples, No. The true labels of each sample (binary classification). sample In its own cluster The logits value output in the sub-model. For patient samples Clustering labels (determined by the output of the clustering layer); Ultimately, the model achieves multi-step synergy between clustering and prediction through joint optimization of prediction and clustering losses. The key to this invention lies in the fact that the model's total loss function is a weighted sum of the clustering and prediction losses: ; in, To balance the coefficients, the impact of clustering objectives on overall training is adjusted; Through joint training, the model simultaneously optimizes prediction accuracy and clustering effect, achieving a trade-off between the two performances. This joint framework ensures that the patient subgroups formed by clustering are highly correlated with risk prediction. The model is forced to learn a feature representation and grouping method that is most beneficial to the prediction task. It no longer clusters for the sake of clustering, but spontaneously divides patients into homogeneous subgroups according to "risk patterns". This ensures that patients in the same cluster not only have similar features, but also have highly consistent disease evolution paths and deterioration risks, thereby fundamentally improving the discriminative ability and interpretability of the prediction model and avoiding the error accumulation problem inherent in the step-by-step method. Model training module: used for offline training of the parameters of the clustering module and prediction module mentioned above; internally encapsulates meta-learning training logic, including task partitioning management, inner and outer loop optimizers, and task unbiased strategy; Task partitioning management specifically involves combining clustering results with external task attributes to form a more representative set of tasks; Cluster-driven task partitioning: based on the hard labels output by the clustering layer Divide the samples into A cluster subset Each subset Patient groups containing similar characteristics (such as elderly COPD patients, young trauma patients, etc.), each cluster subset It can be regarded as an independent task Its goal is to learn risk prediction patterns specific to this subgroup; External attribute extension: Based on the clustering task, the task is further divided according to external attributes (such as hospital, disease type). For example, for the same cluster... If the samples come from different hospitals, they can be broken down into multiple sub-tasks. (like Clustering (In the task at hospital A), therefore, the final task set ,in Represents a clustering index. Indicates an external attribute index; Supports set and query set partitioning: for each task Randomly divided into support sets (Used for updating inner loop parameters) and query sets (Used for calculating external circulation losses); Inner and outer loop optimizers: The inner loop phase aims to enable the model to rapidly adapt to tasks involving specific patient subgroups. This process simulates the "rapid learning" ability of clinicians when facing new departments, new hospitals, or new diseases: quickly adjusting treatment strategies based on a small number of typical cases. Specifically, using globally shared parameters As the initial state, for each subtask (in This indicates a patient subgroup defined by clustering, such as "elderly patients with chronic kidney disease"; Representing external attributes, such as "ICU of a top-tier hospital", a small number of gradient update operations are performed, in the following form: ; in, Indicates in the task The task-specific parameters after one gradient update The learning rate of the inner loop controls the step size of parameter updates. In medical scenarios, this value is usually set relatively small (e.g., ...). To avoid drastic fluctuations caused by data noise, For the task joint loss function For global parameters The gradient of is defined as: ; That is, both the accuracy of risk prediction and the consistency of patient temporal clustering are considered simultaneously; After completing the inner loop, the outer loop phase begins, with the goal of optimizing global parameters. This enables it to have stronger cross-task generalization ability. This process is equivalent to "teaching feedback": by evaluating the performance of each task on the query set, it guides the global model on how to initialize better so that it can adapt to new tasks more quickly in the future. Calculate each task Use the updated parameters on its corresponding query set loss And minimize the average loss of all tasks through backpropagation: ; in, Indicates task Prediction performance on the query set is typically achieved using cross-entropy loss or other classification losses.

[0019] In practical applications, it is assumed that data has been collected from multiple hospitals in several cities across the country, with each hospital representing a different task. The outer loop summarizes the "adaptation effects" of these tasks and trains a highly versatile initial model. When this model is deployed to a new hospital that has never been seen before, even with only a small amount of data, it can quickly converge to an effective solution thanks to its good initialization characteristics. This "learn first, then optimize" strategy enables the model to maintain stable performance in different regions, different populations, and different equipment environments, significantly improving the scalability and robustness of the system. Unbiased strategy for tasks: To avoid bias in model training for a specific disease prediction task, this invention introduces a task-unbiased meta-learning strategy. This strategy balances task contributions through entropy regularization, specifically by incorporating a task entropy loss into the meta-loss. for: ; in, For the model in the task The output class probability distribution on the validation set encourages the model to have greater uncertainty (higher entropy) in its prediction distribution across tasks, avoiding premature overfitting to certain tasks during the meta-training phase. Intuitively, this means making the initial model... By not being overconfident in any task, thus maintaining broader adaptability, this invention preferably adopts the entropy regularization form because it is simple to implement and applicable to different types of tasks such as classification and regression. After adding task entropy regularization, the objective of meta-training becomes: ; in, As the weight coefficients, the initial parameters of the model obtained by this improved meta-learning algorithm are more neutral and unbiased in the task space, with the greatest output uncertainty and potential shaping. This is the idea of ​​task-unbiased meta-learning, which avoids the meta-learning process from biasing certain training tasks, so as to improve the universality of new unseen tasks. Risk identification module: An application-layer module for end users, which combines the output of the prediction module and rule strategies to provide intuitive risk identification results; After the model is trained, two types of parameters are saved: basic network weights and knowledge graph entity embeddings. When deploying, the initial weights obtained from meta-learning are fine-tuned several times on a small number of samples in new scenarios to quickly adapt to local data. If new medical concepts that have not been seen before appear, the knowledge graph is expanded and R-GCN is trained again with a small number of relationships, so as to achieve "commonality preservation + individual fusion" Few-Shot transfer with very few samples. Alternatives to Deep Clustering Algorithms: While this invention employs LSTM as the time-series encoder combined with DEC-style KL divergence clustering, in other implementations, a Transformer encoder or a one-dimensional convolutional network can replace LSTM to extract time-series features. Similarly, the clustering objective function is not limited to KL divergence; soft clustering based on Gaussian mixture models (optimized using the expectation-maximization algorithm) or clustering discriminant loss based on contrastive learning can be considered. As long as the clustering results can be correlated with the prediction objective, the goal of identifying homogeneous patient subgroups and assisting in prediction can be achieved. Replacement of Meta-Learning Frameworks: This invention employs the gradient-optimized MAML paradigm for meta-learning, with improvements such as entropy regularization. However, in other implementations, metric-based meta-learning frameworks, such as Proto-Net or matching networks, can be considered. These methods learn a metric space, enabling similar samples to cluster into tight clusters and dissimilar samples to separate under limited sample conditions, thus achieving generalization without gradient fine-tuning. Furthermore, if the focus is on accelerating meta-training convergence, optimizer-based meta-learning (such as Latent-Optimizedor LEO) can be used, employing a small network specifically designed to learn the direction of model updates. Variations in model structure and loss: Other alternative structures exist for integrated clustering and prediction. For example, a Mixture-of-Experts model can be used, where different expert networks learn the characteristics of a particular patient class, essentially equivalent to manual clustering. Each expert network outputs a sub-prediction, which is then weighted by a gating mechanism. This architecture can adapt to heterogeneous populations. However, its drawbacks include the need to pre-assume the number of experts and high training complexity. Additionally, regarding loss design, instead of explicit clustering loss, multi-task learning can be used to allow the model to simultaneously predict patient class labels (if there are prior classes) and risks, indirectly encouraging the model to distinguish between different patient classes. This can be used as an alternative on some labeled clustered datasets. As for knowledge fusion, without a knowledge graph, attention mechanisms can be used to mine key events in the time series (treating events as special time points and assigning them greater weight) to highlight the impact of key events. Other implementations of unbiased strategies: Besides entropy regularization, other unbiased meta-learning strategies are available for different tasks. For example, methods that minimize the inequality of task losses can be used (e.g., making the initial losses of each task as similar as possible); or the gradient can be normalized at each meta-iteration to prevent any task gradient from dominating the update direction. FairMeta-Learning algorithms, which adjust meta-gradients to balance the contributions of all tasks, can also be used as alternative implementations. In our scenario, these methods share the same fundamental goal: improving the generalization of meta-initialization. For simplicity, this invention uses entropy as the means, but alternative methods are also feasible. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying a risk of exacerbation based on deep time-clustering enhancement, characterized in that, Comprise the following steps: Step one, obtain the patient's multi-modal clinical sequence data, including time series of physiological indicators, laboratory test results, historical disease and drug data; Step two, adopt knowledge enhanced sequence clustering method to extract features and classify patient sequences, and group patient sequences according to their future outcome distribution; Step three, through the meta training process, the model can quickly adapt to the prediction task of new patient subgroups; Step four, based on the trained meta model, quickly adapt to new patient subtypes, and predict the possibility of disease deterioration events in a certain time window in the future.

2. The disease exacerbation risk identification system based on deep temporal clustering enhancement, applied to the disease exacerbation risk identification method based on deep temporal clustering enhancement in claim 1, characterized in that, Comprise: Data preprocessing module: receiving multi-source raw data from hospital information system, monitoring equipment, cleaning, aligning, standardizing operation, generating structured time series and event set; Knowledge enhancement module: built-in critical illness medical knowledge graph database, containing R-GCN graph network reasoning engine, querying the events output by the preprocessing module to obtain the knowledge subgraph of the corresponding entity, and calculating the entity embedding vector; Clustering module: realize deep clustering function, including time series encoder and clustering calculation unit, receive preprocessed data and event embedding provided by knowledge enhancement, calculate patient representation and execute clustering operation; Prediction module: the disease deterioration risk prediction model module is connected with the clustering module, inputs the patient fusion representation or the representation combined with cluster information, and outputs the disease risk prediction result; Model training module: used for offline training parameters of the above clustering module and prediction module; Internal encapsulated meta learning training logic, including task division management, internal and external loop optimizer, and task unbiased strategy; Risk identification module: the application layer module for the end user, combining the prediction module output and rule strategy, giving intuitive risk identification result.

3. The deterioration risk identification system based on deep time-clustering enhancement of claim 2, wherein: The data preprocessing module acquires multi-source heterogeneous clinical data from electronic medical records, monitor records and laboratory information systems, including continuous monitoring time series data and discrete key event data, uniformly completes missing value filling, anomaly correction, time alignment and feature normalization, generates standardized input tensor that can be directly used by the model, including: Multivariate time series where each is the multidimensional physiological feature vector of the patient at time Key event set Each event With event type and occurrence timestamp .

4. The deterioration risk identification system based on deep time-clustering enhancement of claim 3, wherein: The specific operation steps of the knowledge enhancement module are: S2.1, to encoding into a fixed-dimension hidden representation , capturing temporal dynamics dependencies; S2.2, Knowledge graph embedding: Through the relation graph convolution network (R-GCN) to carry out the embedding derivation of the relation perception of the entity, and finally aggregate into the overall event knowledge vector of the patient through weighted average ; The node representation updating rule of R-GCN is: in R-GCN, the node embedding of each layer is updated through the following formula to realize the information aggregation of relationship dependence: ； wherein, representing a node In the first layer embedding vectors; representing a set of all relation types in the knowledge graph; representing the set of neighbor nodes connected to the node through the relation ; representing the number of elements in the set of neighbor nodes; representing the relation type In the first layer weight matrix; representing the weight matrix of the self-loop; representing an activation function; By running multi-layer R-GCNs on the knowledge graph, the final low-dimensional vector representations are learned ; S2.3, time series representation with knowledge graph embeddings by concatenation or attention mechanism fusion, to get comprehensive representation .

5. The deterioration risk identification system based on deep time-clustering enhancement of claim 1, wherein: The operation mode of the clustering module is: Computing soft cluster distribution , for computing clustering loss ; Optimization is performed using a clustering objective function based on the Kullback-Leibler divergence: Model setup with clusters The initial values can be initialized by random initialization or by the latent representation output by a pre-trained autoencoder for the embedding representation of each patient The soft assignment probability that it belongs to the Student t-distribution is used as a similarity measure: ； in, Indicates the first High-dimensional feature embedding vectors of each patient; For the first Cluster centers; The distance is Euclidean. To accelerate cluster convergence and improve intra-cluster consistency, a target distribution is introduced, where the student distribution is a free parameter. By adjusting the current soft allocation Reweight the samples to amplify the weights of high-confidence samples: ；； The optimization goal of the clustering module is to make the soft assignment as close as possible to the target distribution , using the KL divergence as the loss function: ； By minimizing the encoder and cluster center parameters are iteratively updated, after training convergence, each patient is represented assigned a cluster label , after training convergence, the homogeneous subgroups to which the patients belong.

6. The deep time-clustering enhancement-based illness deterioration risk identification system according to claim 1, wherein: The prediction module predicts the risk of each cluster constructing a dedicated risk prediction sub-model The specific steps are as follows: S4.1, Shared encoder weights: All sub-models share the parameters of the underlying encoder, ensuring that the extraction process remains consistent from the raw input to the latent representation of the input data: Encoder output representative of a high-dimensional health state representation of a first patient; All sub-models use the same As input, avoid semantic bias due to different path extraction; S4.2, Cluster-specific prediction head: Each sub-model contains an independent fully connected layer and classifier for generating the final risk prediction result, formulated as: ； wherein, is the weight matrix and bias term of the th sub-model, is the number of disease categories, is the number of samples is the probability distribution of disease risk prediction under the th cluster, thus, the final cross-entropy loss function is: ； in, The total number of training samples, No. The true label of each sample sample In its own cluster The logits value output in the sub-model. For patient samples Clustering tags.

7. The deep time-clustering enhanced deterioration risk identification system according to claim 6, wherein: The prediction module jointly optimizes the prediction loss and the clustering loss, and the total loss function of the model is the weighted sum of the clustering and prediction loss: ； wherein, is a trade-off coefficient that regulates the influence of the clustering objective on the overall training. 8.The deterioration risk identification system based on deep time-clustering enhancement of claim 1, wherein, The task division management in the model training module includes: Cluster-driven task partitioning: based on hard labels output by clustering layer The samples are partitioned into cluster subsets Each subset contains a patient population with similar characteristics, each cluster subset can be considered as an independent task with the goal to learn a risk prediction pattern specific to this sub-population; External attribute expansion: further divide the task according to external attributes on the basis of clustering task; Support set and query set division: for each task , randomly divided into support set and query set .

9. The deep time-clustering enhanced deterioration risk identification system according to claim 8, wherein: The internal and external loop optimizer in the model training module includes: Inner loop phase: with global shared parameters For initial state, for each subtask where denotes the patient subpopulation partitioned by the cluster, denotes the external attributes, in the form of: ； where, denotes the task-specific parameters after one gradient update on the task , is the inner loop learning rate, is the joint loss function for the tasks , is the gradient of the global parameters , which is defined as: ； That is, the risk prediction accuracy and patient time series clustering consistency are considered at the same time; After the inner loop is completed, the outer loop phase is entered, where each task is computed using the updated parameters on its corresponding query set The loss is computed for each task and the average loss over all tasks is minimized by backpropagation: ； wherein, representing a task Predictive performance on a query set.

10. The deep time-clustering enhancement-based illness deterioration risk identification system according to claim 9, characterized in that: The task unbiased strategy in the model training module includes: To balance the task contribution by entropy regularization, we add a task entropy loss in the meta-loss is: ； wherein, is the class probability distribution output by the model on the validation set of the task . After adding task entropy regularization, the goal of meta training becomes: ； wherein, are weight coefficients.

Images