Improving Learning of New Diseases Through Knowledge-Enhanced Initialization for Federated Adapter Tuning.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Federated learning (FL) enables privacy-preserving collaboration in healthcare.
• Foundation models (FMs) are increasingly used with adapter tuning in FL for medical applications.
• Rapidly evolving healthcare demands quick adaptation to new tasks/diseases using past experiences.

Purpose of the Study:

• Introduce Federated Knowledge-Enhanced Initialization (FedKEI) for informed adapter tuning initializations.
• Leverage cross-client and cross-task transfer from past knowledge.
• Enable efficient adaptation to new medical tasks or diseases.

Main Methods:

• FedKEI employs global clustering for cross-task knowledge generalization.
• Optimizes aggregation weights (inter- and intra-cluster) for personalized knowledge transfer.
• Utilizes a bi-level optimization scheme for collaborative learning of weights.

Main Results:

• FedKEI demonstrates significant advantages in adapting to new diseases.
• Experiments conducted on dermatology, chest X-rays, and retinal OCT datasets.
• Outperforms state-of-the-art methods in new task adaptation.

Conclusions:

• FedKEI offers an effective approach for initializing adapters in federated learning for healthcare.
• The framework enhances adaptability to new medical challenges by utilizing prior knowledge.
• FedKEI improves the efficiency and performance of foundation models in diverse medical domains.