Energy-efficient dynamic sensor time series classification for edge health devices.

Area of Science:

• Internet of Things Medical (IoMT)
• Machine Learning (ML)
• Edge Computing

Background:

• Time series data is vital for real-time disease detection in IoMT.
• ML algorithms on edge devices offer reduced latency and enhanced privacy.
• Limited resources on edge devices necessitate energy-efficient algorithms that handle concept drift and catastrophic forgetting.

Purpose of the Study:

• Propose an energy-efficient online time series classification algorithm (OTCD) for edge health devices.
• Address concept drift (CD) and catastrophic forgetting (CF) in medical time series classification.
• Enhance the reliability and efficiency of real-time disease detection on IoMT devices.

Main Methods:

• OTCD detects concept drift and updates prototypes to mitigate its impact.
• Standardizes potential space distribution and selectively preserves training parameters to combat catastrophic forgetting.
• Evaluated using electrocardiogram (ECG) and photoplethysmogram (PPG) data, comparing feature extractors and state-of-the-art models.

Main Results:

• OTCD demonstrates superior accuracy (2.77%–14.74% higher on MIT-BIH arrhythmia dataset) compared to SOTA algorithms.
• Achieves low memory consumption (1 KB) and high computational efficiency (0.004 GFLOPs/sec).
• Exhibits significant energy savings and high time efficiency in runtime tests.

Conclusions:

• OTCD enables efficient real-time medical time series classification on edge health devices.
• The algorithm shows significant competitiveness and potential for secure, reliable healthcare applications.
• OTCD offers a promising solution for overcoming challenges in IoMT-based disease detection.