Severity-Based Hierarchical ECG Classification Using Neural Networks.

Area of Science:

• Biomedical Engineering
• Computer Science
• Cardiology

Background:

• Cardiac arrhythmia detection is crucial for cardiovascular disease management.
• Wearable devices use neural networks for continuous heart monitoring but face accuracy and energy efficiency challenges.

Purpose of the Study:

• To develop architecture-level solutions for deploying neural networks in wearable healthcare devices for cardiac arrhythmia classification.
• To improve both accuracy and energy efficiency in smart wearable arrhythmia detection systems.

Main Methods:

• Proposed a hierarchical neural network architecture activating only necessary components for energy efficiency.
• Introduced a severity-based classification approach for user and medical professional benefit.
• Utilized computation-in-memory hardware with resistive random-access memory (RRAM) for in-situ processing.

Main Results:

• Achieved high accuracy in cardiac arrhythmia classification using the MIT-BIH dataset.
• Demonstrated an average energy consumption of 0.11 μJ per heartbeat classification.
• Reported an area consumption of 0.11 mm², representing a 25× energy and 12× area improvement over state-of-the-art.

Conclusions:

• The proposed architecture offers a significant advancement in energy efficiency and accuracy for wearable arrhythmia detection.
• This approach enables more effective and sustainable continuous cardiac monitoring solutions.
• Severity-based classification provides actionable insights for users and clinicians.