A Lightweight ML-Based ECG Classification System Using Self-Personalized Anomaly Detector.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Real-time arrhythmia diagnosis is crucial for patient care but challenging on resource-limited edge devices.
• Existing methods often require significant computational power, limiting their application in portable or embedded systems.

Purpose of the Study:

• To develop a lightweight, event-driven electrocardiogram (ECG) classification system for real-time arrhythmia diagnosis on edge devices.
• To improve diagnostic accuracy and efficiency by reducing computational load and energy consumption.

Main Methods:

• Developed a novel self-personalized anomaly detector using signal processing to dynamically update decision criteria based on patient ECG history.
• Implemented a Siamese neural network for detailed arrhythmia classification, comparing features from personalized normal data and abnormal inputs.
• Created a simplified Siamese model to reduce trainable parameters while maintaining classification accuracy.

Main Results:

• The event-driven system reduced machine learning model activations by 74% for normal beats.
• Achieved a high classification accuracy of 96.9%, comparable to leading solutions.
• Demonstrated a 3x reduction in energy consumption and a 3.6x faster processing latency on a mobile GPU platform compared to cost-aware methods.

Conclusions:

• The proposed event-driven system enables efficient and accurate real-time arrhythmia diagnosis on edge devices.
• The system's low power consumption and fast processing enhance battery life and facilitate continuous patient monitoring.
• This approach is suitable for resource-constrained environments, advancing portable cardiac diagnostic tools.