DFML: Dynamic Federated Meta-Learning for Rare Disease Prediction

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Rare diseases affect millions globally, but small sample sizes and data privacy concerns hinder AI-driven prediction.
• Traditional AI models struggle with feature extraction for rare diseases due to data limitations.

Purpose of the Study:

• To propose a Dynamic Federated Meta-Learning (DFML) approach for improved rare disease prediction.
• To address challenges of small sample sizes and data sensitivity in rare disease research.

Main Methods:

• Developed an Inaccuracy-Focused Meta-Learning (IFML) approach for dynamic task attention based on base learner accuracy.
• Implemented a dynamic weight-based fusion strategy for federated learning, selecting clients by local model accuracy.

Main Results:

• The DFML approach outperformed the original federated meta-learning algorithm in both accuracy and speed.
• Achieved significant improvements with as few as five shots on public datasets.
• Demonstrated a 13.28% average increase in prediction accuracy compared to individual hospital models.

Conclusions:

• The proposed DFML approach effectively enhances rare disease prediction accuracy and efficiency.
• Dynamic adjustments in meta-learning and federated learning are crucial for handling rare disease data challenges.