A Scalable and Transferable Federated Learning System for Classifying Healthcare Sensor Data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Healthcare Informatics

Background:

• The Internet of Medical Things generates massive healthcare sensor data (HSD), raising significant privacy and security concerns.
• Edge computing offers opportunities for processing HSD but faces challenges in developing lightweight, dynamic classification systems.
• Existing systems struggle with evolving data distributions and the emergence of new, previously unseen classes in HSD.

Purpose of the Study:

• To propose a scalable and transferable classification system (SCALT) for edge computing environments handling dynamic HSD.
• To address the challenges of data distribution changes and unknown classes in HSD classification.
• To enhance patient privacy through automatic classification of sensitive healthcare data.

Main Methods:

• Developed SCALT, a one-classifier-per-class system leveraging federated learning.
• Employed a one-dimensional convolution-based network for feature extraction and individual mini-classifiers for each class.
• Implemented a parameter protection mechanism to prevent catastrophic forgetting in sequential HSD classification tasks.

Main Results:

• SCALT demonstrated high classification accuracies on physiological signal datasets: Electrocardiogram (98.65%), Electroencephalogram (91.10%), and Photoplethysmograph (89.93%).
• Achieved superior performance compared to existing state-of-the-art methods.
• Successfully applied SCALT to a patient privacy protection scenario.

Conclusions:

• SCALT provides a scalable and transferable solution for classifying dynamic HSD in edge computing.
• The system effectively handles evolving data characteristics and new classes while preserving patient privacy.
• SCALT represents a significant advancement in secure and efficient healthcare data analysis at the edge.