NeuroCARE: A generic neuromorphic edge computing framework for healthcare applications.

Area of Science:

• Biomedical signal processing
• Neuromorphic computing
• Edge AI

Background:

• Existing neural network methods for biomedical signal classification are computationally expensive and power-hungry, limiting their use on edge devices.
• Spiking neural networks (SNNs) offer high energy efficiency and good performance, making them suitable for edge applications.

Purpose of the Study:

• To propose and evaluate NeuroCARE, a generic neuromorphic framework for edge healthcare and biomedical applications.
• To demonstrate the framework's effectiveness across diverse tasks including epileptic seizure prediction, arrhythmia detection, and hand gesture recognition.

Main Methods:

• NeuroCARE utilizes a sparse spike encoder to convert raw biomedical signals into spike sequences.
• A spike-based computing engine combines Convolutional Neural Network (CNN) and SNN advantages.
• An adaptive weight mapping method efficiently converts CNNs to SNNs without performance loss.

Main Results:

• NeuroCARE achieved high accuracy: 92.7% for seizure prediction, 96.7% for arrhythmia detection, and 85.7% for hand gesture recognition.
• Compared to CNNs, computation complexity was reduced by over 80.7%.
• Energy consumption and area occupation were reduced by over 80% and 64.8%, respectively.

Conclusions:

• The proposed neuromorphic computing approach, NeuroCARE, is energy and area efficient with high precision.
• This framework facilitates the deployment of advanced biomedical signal processing on edge platforms like mobile and wearable devices.