Cross-hospital medical data adaptation system based on transfer learning

Abstract

The invention relates to the field of medical artificial intelligence, and discloses a cross-hospital medical data adaptation system based on transfer learning, and the system comprises a field adaptive layer which encodes data of a source hospital and a target hospital and achieves the distribution alignment; a dynamic feature weight adjustment mechanism for dynamically calculating the feature weight of the target hospital based on the coding feature; continuously adapting to the optimization framework, carrying out online incremental learning and retaining historical experience, and providing updated parameters for a cross-hospital cooperative training mechanism; a cross-hospital collaborative training mechanism is adopted, and joint training and global model fusion are performed under the condition of keeping data privacy; and the multi-modal data fusion module is used for aggregating different modal data of multiple hospitals and transmitting the data to the field adaptive layer. According to the method, the field adaptive layer design based on the generative adversarial network is adopted, automatic alignment of data distribution of the source hospital and the target hospital is achieved, and manual intervention is not needed when the model is migrated among different hospitals by introducing multi-granularity feature coding and an adversarial discrimination mechanism.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence, and specifically to a cross-hospital medical data adaptation system based on transfer learning. Background Art

[0002] In the current healthcare environment, with the explosive growth of medical data, cross-hospital data sharing and adaptation have become key to improving the efficiency and quality of medical services. However, data from different hospitals differ significantly in format, feature distribution, and other aspects. Traditional methods often face the dual challenges of data privacy protection and unsatisfactory transfer learning results when solving this problem.

[0003] In existing technologies, many systems use medical data adaptation methods based on transfer learning, using data from source hospitals to transfer knowledge to data from target hospitals. These methods usually rely on static weighting of data or pre-defined feature extraction methods, which can improve the data adaptability between source and target hospitals to a certain extent. In addition, some methods promote the collaborative sharing of medical data through incremental learning and cross-hospital model training, especially during model training, avoiding centralized data storage and ensuring the privacy of hospital data.

[0004] However, the existing technology still faces some key deficiencies. First, the existing cross-hospital data adaptation methods usually rely on manual feature selection or static weighting processing, which limits their adaptability to data feature heterogeneity; and with the continuous updating and changes of medical data, static methods often do not work well when processing new data, and it is difficult to dynamically adjust feature weights, resulting in a decrease in transfer learning performance; secondly, although the existing incremental learning methods can handle the gradual changes in data, when faced with heterogeneous data from multiple hospitals, it is difficult to effectively retain historical experience and avoid the occurrence of forgetting, thereby affecting the long-term stability of the model; finally, the traditional cross-hospital collaborative training method lacks fine processing of the differences in data features between different hospitals. It usually generates a global model through simple parameter synchronization, fails to take into account the complexity and differences in data distribution between hospitals, and results in unsatisfactory collaborative training results. For this reason, those skilled in the art have proposed a cross-hospital medical data adaptation system based on transfer learning to solve the above problems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a cross-hospital medical data adaptation system based on transfer learning, which solves the problems of data leakage risk and underutilization of distribution differences in existing centralized training methods.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a cross-hospital medical data adaptation system based on transfer learning, comprising: The domain adaptation layer is used to encode the source hospital data and the target hospital data and achieve distribution alignment, and pass the encoding results to the dynamic feature weight adjustment mechanism; A dynamic feature weight adjustment mechanism is used to dynamically calculate the weight distribution of target hospital features based on the feature representation output by the domain adaptation layer, and pass the adjusted features to the continuous adaptive optimization framework; A continuous adaptive optimization framework for performing online incremental learning of features output by the dynamic feature weight adjustment mechanism and retaining historical adaptation experience. The continuous adaptive optimization framework provides updated model parameters to the cross-hospital collaborative training mechanism; A cross-hospital collaborative training mechanism is used to jointly train local models of each hospital and fuse them into a global model while maintaining the privacy of each hospital's data. The global model is fed back into the continuous adaptive optimization framework; The multimodal data fusion module is used to aggregate features of medical data of different modalities from multiple hospitals and input the fused features into the domain adaptation layer.

[0007] Preferably, the domain adaptation layer includes: Multi-granularity feature encoder, used to extract statistical and temporal features of source and target hospitals; The adversarial domain alignment module is used to construct a domain discriminator through a generative adversarial network to minimize the feature distribution difference between the source hospital data and the target hospital data.

[0008] Preferably, the optimization objective of the adversarial domain alignment module includes the following loss function: Where: x s Enter data for the source hospital; x t is the target hospital input data; D(·) is the domain discriminant function; represents the expected operation; p s , p t are the characteristic distributions of the source hospital and target hospital data respectively.

[0009] Preferably, the dynamic feature weight adjustment mechanism includes: A meta-learning-driven feature weight prediction unit, which is used to adaptively adjust the target hospital's feature distribution based on the pre-trained meta-model; The uncertainty-aware regularization unit is used to model the confidence of the feature weights, thereby controlling the uncertainty of the model.

[0010] Preferably, the weight generation process of the meta-learning driven feature weight prediction unit is: w target =f meta (Xtarget ,w source ); Where: w target is the dynamic weight of the target hospital characteristics; X target Enter data for the target hospital; source is the prior weight of the source hospital; f meta It is a meta-learning model.

[0011] Preferably, the regularization objective of the uncertainty-aware regularization unit is: Where: w is the feature weight; logp(w) is the prior logarithmic probability density of the weight; λ is the regularization coefficient; ||w|| 2 is the L2 norm term.

[0012] Preferably, the continuous adaptation optimization framework includes: Online incremental learning unit, used to adapt to new data with parameter importance as constraint; A multi-source domain buffer pool unit is used to store historical hospital data and its feature representation to enhance model generalization capabilities.

[0013] Preferably, the online incremental learning unit adopts the following elastic weight consolidation loss function: Where: w i is the current model parameter; is the historical model parameter; i is the importance coefficient of the i-th parameter.

[0014] Preferably, the cross-hospital collaborative training mechanism includes: Decentralized training units, which enable each hospital to train models on medical data locally; The parameter aggregation module is used to periodically synchronize local model parameters among multiple hospitals and perform weighted averaging of local parameters to form a global shared model. The aggregation method is: Where: w (i) is the parameter of the local model of the i-th hospital; w global is the global model parameter; N is the number of hospitals participating in the training.

[0015] Preferably, the multimodal data fusion module includes a graph neural network fusion unit, and its node representation update method is: in: Represents the feature representation of node v at the k+1th layer; Represents the feature representation of neighbor node u at the kth layer; is the set of neighbor nodes of node v; W is the shared linear transformation weight matrix; σ(·) is the activation function.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention adopts a domain adaptation layer design based on a generative adversarial network to achieve automatic alignment of the data distribution of the source hospital and the target hospital. By introducing multi-granularity feature encoding and an adversarial discrimination mechanism, the model can be migrated between different hospitals without manual intervention. Compared with the existing technology that relies on manual feature selection and coarse-grained matching, this solves the problem of low adaptation accuracy caused by inconsistent feature distribution.

[0017] 2. The present invention introduces a dynamic feature weight adjustment mechanism, combined with meta-learning driven and uncertainty-aware regularization methods, to dynamically and adaptively adjust the importance of target hospital features. Compared with traditional transfer learning methods that only rely on static feature weighting processing, the present invention effectively solves the shortcomings of large cross-hospital data heterogeneity and feature redundancy affecting the adaptation effect, and improves the adaptability and stability of the system.

[0018] 3. This paper proposes a continuous adaptive optimization framework that combines online incremental learning with a multi-source domain buffer pool design, enabling the model to continuously learn and retain historical experience in the face of ever-changing hospital data. Compared with existing systems that only support single-shot static training or single-source fine-tuning, this significantly improves the model's proneness to forgetting and insufficient generalization capabilities in long-term cross-hospital applications.

[0019] 4. This invention designs a cross-hospital collaborative training mechanism. Through decentralized local training and parameter aggregation strategies, it ensures the privacy of data from each hospital while achieving global model sharing. This is different from existing methods that require centralized data collection or rely on third-party intermediary aggregation models. It effectively avoids the risk of data exposure and improves the security and practical deployability of the cross-hospital adaptation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the system architecture of the present invention; Figure 2 It is a schematic diagram of the domain adaptation layer architecture of the present invention; Figure 3 It is a schematic diagram of the dynamic feature weight adjustment mechanism framework of the present invention; Figure 4 It is a schematic diagram of the continuous adaptive optimization framework of the present invention; Figure 5 This is a schematic diagram of the cross-hospital collaborative training mechanism framework of the present invention; Figure 6It is a schematic diagram of the multimodal data fusion module architecture of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] Please see the attached Figure 1 -Attached Figure 6 The embodiment of the present invention provides a cross-hospital medical data adaptation system based on transfer learning, including: The domain adaptation layer is used to encode the source hospital data and the target hospital data and achieve distribution alignment, and pass the encoding results to the dynamic feature weight adjustment mechanism; Specifically, in this embodiment, the domain adaptation layer, a key module of the cross-hospital medical data adaptation system, is responsible for encoding source and target hospital data and aligning their feature distributions. Specifically, the domain adaptation layer includes a multi-granularity feature encoder and an adversarial domain alignment module, enabling data sharing and migration across different hospitals while ensuring data privacy and security.

[0023] Multi-granularity feature encoder: In this embodiment, the domain adaptation layer first encodes the data from the source and target hospitals using a multi-granularity feature encoder. The source and target hospitals may have different data formats, feature distributions, and dimensional structures. Therefore, the feature encoder is required to normalize the data for efficient processing in subsequent transfer learning and feature alignment.

[0024] The multi-granularity feature encoder extracts statistical and temporal features of the source and target hospitals based on different medical data types. For example, for time series data, the encoder can extract patient physiological signals and time-dependent features of the diagnosis and treatment process; for structured data, it extracts various hospital statistics such as the number of patients and disease types. This approach provides a unified representation of the characteristics of the source and target hospitals, facilitating subsequent model training and data fusion.

[0025] Adversarial Domain Alignment Module: In this embodiment, another core component of the domain adaptation layer is the adversarial domain alignment module. This module uses a generative adversarial network (GAN) approach to train a domain discriminator to minimize the difference in feature distributions between the source and target hospital data. The goal of this module is to minimize the distance between the feature distributions of the source and target hospital data, enabling the model to effectively transfer data between the two hospitals without losing the data feature information of the original hospital.

[0026] The adversarial domain alignment module uses a generative adversarial network training process to make the target hospital data features generated by the generator (domain adapter) indistinguishable from the source hospital data features. Specifically, the domain discriminator's task is to distinguish between source and target hospital data, while the generator's goal is to make it impossible for the discriminator to determine which hospital the input data belongs to. Through this adversarial training, the model automatically learns the shared features between the source and target hospitals, thereby achieving data alignment.

[0027] To train the adversarial domain alignment module, the following loss function is used to optimize the parameters of the generator and domain discriminator: Where: x s Enter data for the source hospital; x t is the target hospital input data; D(·) is the domain discriminant function; p s , p t are the characteristic distributions of the source hospital and target hospital data respectively; represents the expected operation; p s , p t are the characteristic distributions of the source hospital and target hospital data respectively; is the loss function used to train the generator and domain discriminator.

[0028] The loss function is designed to minimize the difference in feature distribution between the source and target hospital data, enabling the generator to produce source hospital features that are as close as possible to the target hospital's distribution. By optimizing this loss function, the generator continuously adjusts its output features, adapting the target hospital's data features during training and reducing the distribution difference between the source and target hospitals.

[0029] During training, the domain discriminator and generator are optimized alternately. The domain discriminator continuously identifies the characteristics of the source and target hospitals and provides feedback to the generator. Based on the discriminator's feedback, the generator gradually learns the common features between the source and target hospital data. This adversarial training method makes the target hospital features generated by the model more universal, thereby improving the effectiveness of transfer learning.

[0030] The specific training process includes the following steps: First, the domain discriminator is trained by the source hospital data to discriminate the distribution of the source hospital data and the target hospital data.

[0031] Then, the domain discriminator is trained using the target hospital data so that the discriminator can distinguish the target hospital data.

[0032] Then, the generator adjusts its output under the feedback of the domain discriminator to generate data that is more consistent with the characteristics of the target hospital.

[0033] Finally, through adversarial training, the generator and domain discriminator are repeatedly optimized until the model can achieve effective data adaptation between the source hospital and the target hospital.

[0034] The design of this domain adaptation layer effectively addresses the difficulties in transfer learning caused by data distribution differences between hospitals. By combining a multi-granularity feature encoder with an adversarial domain alignment module, this method enables seamless adaptation of source and target hospital data, avoiding the degradation of learning performance caused by data distribution differences. The use of a generative adversarial network ensures the authenticity and validity of data features during the transfer process, providing strong support for cross-hospital medical data sharing.

[0035] A dynamic feature weight adjustment mechanism is used to dynamically calculate the weight distribution of target hospital features based on the feature representation output by the domain adaptation layer, and pass the adjusted features to the continuous adaptive optimization framework; Specifically, in this embodiment, the dynamic feature weight adjustment mechanism dynamically calculates the weight distribution of the target hospital's features based on the feature representation output by the domain adaptation layer and passes the adjusted features to the continuous adaptive optimization framework. This mechanism, through meta-learning and uncertainty-aware regularization techniques, enables adaptive adjustment and optimization of features between the source and target hospitals, ensuring better adaptation to the target hospital's data.

[0036] Meta-learning-driven feature weight prediction unit: In this embodiment, the meta-learning-driven feature weight prediction unit is one of the core components of the dynamic feature weight adjustment mechanism. Its primary function is to adaptively adjust the target hospital's feature distribution based on a pre-trained meta-learning model. Specifically, the meta-learning model learns from the source hospital's prior knowledge and the target hospital's input data to generate dynamic feature weights that are used to adjust the importance of the target hospital's features.

[0037] The core idea of ​​meta-learning is to learn how to adjust learning strategies, especially when faced with new tasks (such as the target hospital's characteristic data). In this way, the target hospital's data features can be quickly optimized and adjusted based on the source hospital's knowledge during training, thereby achieving better transfer learning results.

[0038] The specific weight generation process can be expressed as the following formula: w target =f meta (X target ,w source ); Where: w target is the dynamic weight of the target hospital characteristics; X target Enter data for the target hospital; source is the prior weight of the source hospital; f meta It is a meta-learning model.

[0039] This formula describes the process of generating the target hospital's characteristic weights. The meta-learning model outputs dynamic weights for the target hospital's characteristics based on the source hospital's characteristics and the target hospital's input data. This allows the meta-learning model to quickly adapt to the target hospital's characteristics and, through weight adjustment, optimize the target hospital's characteristics throughout the transfer learning process.

[0040] Uncertainty-aware regularization unit: In this embodiment, the uncertainty-aware regularization unit is used to model the confidence level of feature weights, thereby controlling the uncertainty of the model. This unit is designed to constrain the confidence level of feature weights, thereby preventing the model from over-relying on unreliable features when data is insufficient or of poor quality. This makes the model more stable in feature selection and adjustment, thereby improving the robustness of cross-institutional data adaptation.

[0041] Specifically, the uncertainty-aware regularization unit uses the following regularization objective function: Where: w is the feature weight; logp(w) is the prior logarithmic probability density of the weight; λ is the regularization coefficient; ||w|| 2 is the L2 norm term; represents the Bayesian regularization objective function.

[0042] This regularization objective is designed to model the prior distribution of feature weights, ensuring that the weights do not deviate too much or too little during optimization. Furthermore, the L2 norm term helps control the size of the weights, preventing overfitting and model instability during training. This regularization method stabilizes the update process of feature weights, thereby enhancing the model's generalization ability.

[0043] The dynamic feature weight adjustment mechanism combines meta-learning with uncertainty-aware regularization to achieve more precise adaptive adjustment between target and source hospital features. The meta-learning-driven feature weight prediction unit enables rapid optimization of target hospital features based on prior information about the source hospital, while the uncertainty-aware regularization unit further improves the stability and robustness of the model by modeling the uncertainty of the weights.

[0044] This mechanism enables target hospitals to fully consider the differences between data from different hospitals when accepting knowledge transfer from source hospitals, avoiding over-reliance on unreliable features, thereby improving the effectiveness and adaptability of the cross-hospital medical data adaptation system. In this way, the target hospital's data can be more smoothly integrated into the global model, achieving more efficient data adaptation.

[0045] A continuous adaptive optimization framework is used to perform online incremental learning of features output by the dynamic feature weight adjustment mechanism and retain historical adaptation experience. The continuous adaptive optimization framework provides updated model parameters to the cross-hospital collaborative training mechanism; Specifically, in this embodiment, the continuous adaptive optimization framework primarily performs online incremental learning of features output by the dynamic feature weight adjustment mechanism and retains historical adaptation experience, thereby continuously optimizing model parameters to achieve cross-hospital data adaptation. This framework includes an online incremental learning unit and a multi-source domain buffer pool unit, which are used to dynamically adjust and store historical data, respectively, ensuring that the model can adapt to new hospital data and maintain good generalization capabilities.

[0046] Online Incremental Learning Unit: In this embodiment, the online incremental learning unit is a core component of the continuous adaptive optimization framework. Its primary function is to adapt to new data while constraining the importance of existing model parameters, thereby achieving gradual optimization of the new data. The goal of online incremental learning is to enable the model to adapt to changes in the characteristics of the target hospital as new data arrives, avoiding overfitting or excessive dependence on new data.

[0047] Specifically, the online incremental learning unit uses the elastic weight consolidation (EWC) method to adapt to new data. EWC assigns an importance coefficient to each model parameter, controlling the update amplitude of the parameter. This ensures that when new data arrives, the model does not forget important features learned previously. This method effectively prevents the model from forgetting or losing old data due to learning new data during the incremental learning process.

[0048] The expression of the EWC loss function is as follows: Where: w i is the i-th model parameter in the current task; is the i-th model parameter in the historical task; λ i is the importance coefficient of the i-th parameter; Represents the elastic weight solidification loss function.

[0049] The loss function is designed to ensure that important parameters are updated smaller by assigning different weights to each parameter, thereby protecting key features from being forgotten during the incremental learning process.

[0050] Through this EWC loss function, the online incremental learning unit can appropriately adjust the model parameters according to the arrival of new tasks to balance the learning effects between new and old tasks and prevent the model from losing previous data adaptation experience when adapting to new data.

[0051] Multi-source domain buffer unit: In this embodiment, the multi-source domain buffer unit preserves historical hospital data and its feature representations to enhance the model's generalization capabilities. By storing and managing data from multiple hospitals, this unit provides a long-term memory mechanism, enabling the model to draw on past data adaptation experience when new hospital data arrives, thereby improving its adaptability to different hospitals.

[0052] Specifically, the multi-source domain buffer pool unit stores data features and adaptation experience from multiple hospitals in a shared pool. During each training session, relevant data is extracted from the buffer pool for training based on the needs of the current task. This approach helps the model more efficiently leverage existing knowledge when encountering data from different hospitals, avoiding the need to start training from scratch every time new hospital data is encountered.

[0053] By storing the data characteristics and adaptation experience of historical hospitals, the multi-source domain buffer pool unit enables the model to have strong cross-hospital adaptability and can transfer knowledge between different hospitals, enhancing the generalization ability and stability of the model.

[0054] The continuous adaptive optimization framework combines online incremental learning with a multi-source domain buffer pool to ensure continuous optimization and stability during cross-hospital data adaptation. The online incremental learning unit uses elastic weight consolidation to control the model's adaptation to new data, preventing the inclusion of new data and the subsequent forgetting of old data. The multi-source domain buffer pool unit, by storing data and feature representations from multiple hospitals, provides the model with long-term memory and enhances its generalization capabilities.

[0055] This framework design enables the cross-hospital data adaptation system to maintain high adaptability and stability in the face of ever-changing hospital data, and ensures that the model effectively leverages existing knowledge and experience during sharing and migration between hospitals. This approach continuously optimizes model parameters, improving the effectiveness of cross-hospital collaborative training and achieving more accurate medical data adaptation.

[0056] A cross-hospital collaborative training mechanism is used to jointly train local models of each hospital and fuse them into a global model while maintaining the privacy of each hospital's data. The global model is then fed back into the continuous adaptation and optimization framework. Specifically, in this embodiment, the cross-hospital collaborative training mechanism primarily trains local models across multiple hospitals through decentralized training and periodic synchronization of local model parameters, while ensuring data privacy across hospitals. This model is then integrated into a global model. This global model is then fed back into the continuous adaptive optimization framework to further optimize the data adaptation process for the target hospital.

[0057] Decentralized Training Unit: In this embodiment, the decentralized training unit is designed to ensure that each hospital performs data training locally and enables collaborative learning through local model training without compromising hospital data privacy. Specifically, each hospital trains its model based on its own medical data, and no specific raw data is directly exchanged or shared during training. Decentralized training ensures the independence and privacy of each hospital's data while also improving cross-hospital data sharing through local model training.

[0058] The decentralized training unit independently trains each hospital's local data, allowing each hospital's model to be optimized based on its unique medical data characteristics. However, during training, each hospital is not completely isolated; instead, parameters are periodically synchronized with other hospitals to form a unified global model. This approach enables cross-hospital joint training while ensuring data privacy and security.

[0059] Parameter Aggregation Module: The function of the parameter aggregation module is to periodically synchronize the local model parameters of multiple hospitals and perform weighted averaging on these local parameters to form a global shared model. The generation of the global model is achieved by aggregating the weight parameters of the local models of each hospital. Specifically, the aggregation process uses the following formula: Where: w (i) is the parameter of the local model of the i-th hospital; w global is the global model parameter; N is the number of hospitals participating in the training.

[0060] This formula describes how the global model parameters are calculated. By taking a weighted average of each hospital's local model parameters, the parameter aggregation module generates a global shared model that aggregates the knowledge and experience of each hospital, has strong generalization capabilities, and is applicable to data from multiple hospitals.

[0061] Through periodic synchronization and weighted averaging, the parameter aggregation module ensures that the knowledge learned by each hospital during training is integrated into the global model, thereby improving the model's adaptability and cross-hospital performance. Because each hospital's model training is independent, the model parameter aggregation process effectively avoids data leakage and privacy issues, while also enabling cross-hospital collaborative learning.

[0062] The cross-hospital collaborative training mechanism combines decentralized training with parameter aggregation, enabling cross-hospital joint training while ensuring data privacy and security. The decentralized training unit enables each hospital to independently train its model, ensuring the privacy and security of its data. The parameter aggregation module, on the other hand, takes a weighted average of each hospital's local model parameters to form a globally shared model, ensuring that knowledge across multiple hospitals can be effectively shared.

[0063] This mechanism enables efficient knowledge sharing and joint training across hospitals, avoiding the privacy risks associated with directly exchanging raw data while improving the model's cross-hospital adaptability. Through global model aggregation and feedback, the system continuously optimizes each hospital's local model, continuously improving each hospital's feature adaptation capabilities and enabling more efficient data adaptation and collaborative learning.

[0064] The multimodal data fusion module is used to aggregate features of medical data of different modalities from multiple hospitals and input the fused features into the domain adaptation layer.

[0065] Specifically, in this embodiment, the multimodal data fusion module aggregates features from multimodal medical data from multiple hospitals and feeds these features into the domain adaptation layer. This module aggregates and updates data features from different modalities based on node adjacency via a graph neural network fusion unit, ensuring that the multimodal features of the data can be effectively integrated and applied to subsequent model training.

[0066] Graph Neural Network Fusion Unit: In this embodiment, a graph neural network (GNN) fusion unit is used to process multimodal data features from different hospitals. The GNN constructs a graph structure, represents data samples as nodes in the graph, and propagates and updates information through the adjacency relationship between nodes. Specifically, the GNN fusion unit can iteratively update data features based on the adjacency relationship of nodes, thereby achieving effective fusion of multimodal data features.

[0067] The update rules for each layer of the graph neural network are as follows: in: Represents the feature representation of node v at the k+1th layer; Represents the feature representation of neighbor node u at the kth layer; is the set of neighbor nodes of node v; W is the shared linear transformation weight matrix; σ(·) is the activation function.

[0068] This formula describes the node update process of the graph neural network. The feature representation of node v is is represented by the feature representation of its neighbor node u The weighted sum is obtained by the shared linear transformation weight matrix W, while the activation function σ(·) performs a nonlinear transformation on the weighted sum to generate the updated features of node v.

[0069] In this way, graph neural networks can effectively propagate information based on adjacency relationships within the graph, enabling nodes to share and aggregate features from multimodal data. Each layer update considers the relationship between the current node and its neighbors, allowing the final node representation to integrate data features from different hospitals.

[0070] In this embodiment, the graph neural network fusion unit of the multimodal data fusion module is capable of processing multimodal data from different hospitals, such as imaging data, electronic medical records, and medical records. Each data modality can be viewed as a distinct node in the graph, connected by certain adjacency relationships. This approach enables data features from different modalities to be fused and updated within the same graph structure, resulting in a unified representation of medical data features across different hospitals.

[0071] Through the aggregation mechanism of graph neural networks, multimodal data from multiple hospitals can be effectively integrated to generate a unified feature representation. This feature representation reflects both shared features and unique features across hospitals, facilitating subsequent feature alignment and optimization in the domain adaptation layer.

[0072] The graph neural network fusion unit ensures the effective fusion of features from different modalities while maintaining the structural relationships between data. By propagating adjacency relationships, the graph neural network can fully utilize the correlations between data from different modalities, avoiding the information loss that can occur when traditional methods process each data modality separately.

[0073] Furthermore, through the shared linear transformation weight matrix W, the graph neural network can uniformly process multimodal data and implement nonlinear feature transformation through the activation function σ(·), thereby enhancing the expressive power of feature representation. This approach enables the effective fusion of data from different hospitals at the feature level, providing high-quality input data for the subsequent domain adaptation layer.

[0074] The examples of this specific embodiment are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, any equivalent changes made based on the structure, shape, and principle of this application should be included in the scope of protection of this application.

Claims

1. A cross-hospital medical data adaptation system based on transfer learning, characterized by: include: The domain adaptation layer is used to encode the source hospital data and the target hospital data and achieve distribution alignment, and pass the encoding results to the dynamic feature weight adjustment mechanism; A dynamic feature weight adjustment mechanism is used to dynamically calculate the weight distribution of target hospital features based on the feature representation output by the domain adaptation layer, and pass the adjusted features to the continuous adaptive optimization framework; A continuous adaptive optimization framework for performing online incremental learning of features output by the dynamic feature weight adjustment mechanism and retaining historical adaptation experience. The continuous adaptive optimization framework provides updated model parameters to the cross-hospital collaborative training mechanism; A cross-hospital collaborative training mechanism is used to jointly train local models of each hospital and fuse them into a global model while maintaining the privacy of each hospital's data. The global model is fed back into the continuous adaptive optimization framework; The multimodal data fusion module is used to aggregate features of medical data of different modalities from multiple hospitals and input the fused features into the domain adaptation layer.

2. The cross-hospital medical data adaptation system based on transfer learning according to claim 1 is characterized in that: The domain adaptation layer includes: Multi-granularity feature encoder, used to extract statistical and temporal features of source and target hospitals; The adversarial domain alignment module is used to construct a domain discriminator through a generative adversarial network to minimize the feature distribution difference between the source hospital data and the target hospital data.

3. The cross-hospital medical data adaptation system based on transfer learning according to claim 2 is characterized in that: The optimization objectives of the adversarial domain alignment module include the following loss functions: Where: x s Enter data for the source hospital; x t is the target hospital input data; D(·) is the domain discriminant function; represents the expected operation; p s , p t are the characteristic distributions of the source hospital and target hospital data respectively.

4. The cross-hospital medical data adaptation system based on transfer learning according to claim 1 is characterized in that: The dynamic feature weight adjustment mechanism includes: A meta-learning-driven feature weight prediction unit, which is used to adaptively adjust the target hospital's feature distribution based on the pre-trained meta-model; The uncertainty-aware regularization unit is used to model the confidence of the feature weights, thereby controlling the uncertainty of the model.

5. The cross-hospital medical data adaptation system based on transfer learning according to claim 4 is characterized in that: The weight generation process of the meta-learning driven feature weight prediction unit is as follows: w target =f meta (X target ,w source ); Where: w target is the dynamic weight of the target hospital characteristics; X target Enter data for the target hospital; source is the prior weight of the source hospital; f meta It is a meta-learning model.

6. The cross-hospital medical data adaptation system based on transfer learning according to claim 4 is characterized in that: The regularization objective of the uncertainty-aware regularization unit is: Where: w is the feature weight; logp(w) is the prior logarithmic probability density of the weight; λ is the regularization coefficient; ||w|| 2 is the L2 norm term.

7. The cross-hospital medical data adaptation system based on transfer learning according to claim 1 is characterized in that: The continuous adaptation and optimization framework includes: Online incremental learning unit, used to adapt to new data with parameter importance as constraint; A multi-source domain buffer pool unit is used to store historical hospital data and its feature representation to enhance model generalization capabilities.

8. The cross-hospital medical data adaptation system based on transfer learning according to claim 7 is characterized in that: The online incremental learning unit adopts the following elastic weight consolidation loss function: Where: w i is the current model parameter; is the historical model parameter; i is the importance coefficient of the i-th parameter.

9. The cross-hospital medical data adaptation system based on transfer learning according to claim 1 is characterized in that: The cross-hospital collaborative training mechanism includes: Decentralized training units, which enable each hospital to train models on medical data locally; The parameter aggregation module is used to periodically synchronize local model parameters among multiple hospitals and perform weighted averaging of local parameters to form a global shared model. The aggregation method is: Where: w (i) is the parameter of the local model of the i-th hospital; w global is the global model parameter; N is the number of hospitals participating in the training.

10. The cross-hospital medical data adaptation system based on transfer learning according to claim 1, characterized in that: The multimodal data fusion module includes a graph neural network fusion unit, and its node representation update method is as follows: in: Represents the feature representation of node v at the k+1th layer; Represents the feature representation of neighbor node u at the kth layer; is the set of neighbor nodes of node v; W is the shared linear transformation weight matrix; σ(·) is the activation function.

Images