Few-Shot Class-Incremental Learning for Medical Time Series Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Informatics

Background:

• Continuous analysis of medical time series is crucial for health monitoring and decision-making.
• Few-shot class-incremental learning (FSCIL) addresses classifying new classes without forgetting old ones.
• Existing FSCIL research has limitations in medical time series due to high intra-class variability.

Purpose of the Study:

• To propose a novel framework, MAPIC, for few-shot class-incremental learning in medical time series.
• To enhance the classification accuracy and robustness of incremental learning models in the medical domain.
• To address the challenges of large intra-class variability and catastrophic forgetting in medical time series data.

Main Methods:

• MAPIC framework with three modules: embedding encoder, prototype enhancement, and distance-based classifier.
• Parameter protection strategy to freeze embedding encoder parameters and mitigate catastrophic forgetting.
• Self-attention mechanism in the prototype enhancement module to improve inter-class distinctions.
• Composite loss function including sample classification, prototype non-overlapping, and knowledge distillation losses.

Main Results:

• MAPIC significantly outperforms state-of-the-art approaches on three medical time series datasets.
• Achieved performance improvements of 27.99%, 18.4%, and 3.95% over existing methods.
• Demonstrated effectiveness in reducing intra-class variations and resisting catastrophic forgetting.

Conclusions:

• MAPIC provides a robust solution for few-shot class-incremental learning in medical time series.
• The proposed methods effectively enhance prototype expressiveness and reduce intra-class variations.
• MAPIC represents a significant advancement for continuous learning in healthcare applications.