An Ultralow-Power Real-Time Machine Learning Based fNIRS Motion Artifacts Detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Machine learning models demand significant computational resources, limiting their application in power-constrained wearable devices.
• Motion artifacts are a major challenge in functional near-infrared spectroscopy (fNIRS) data acquisition, affecting signal quality and analysis.

Purpose of the Study:

• To develop an ultra-low-power, real-time machine learning-based motion artifact detection module for fNIRS systems.
• To enable the deployment of fNIRS technology in wearable devices with limited processing power and memory.

Main Methods:

• Implementation of a machine learning classifier on a field-programmable gate array (FPGA).
• Optimization for ultra-low power consumption and minimal resource utilization (LUTs, FFs).
• Real-time processing of fNIRS data for motion artifact detection.

Main Results:

• Achieved a high classification accuracy of 97.42% for motion artifact detection.
• Demonstrated ultra-low dynamic power consumption of 0.021 W.
• Reported low FPGA resource utilization (38,354 LUTs, 6,024 FFs), outperforming conventional CPU-based Support Vector Machine (SVM) methods.

Conclusions:

• FPGA-based fNIRS motion artifact classification is feasible within low power and resource constraints.
• The developed module meets crucial embedded hardware requirements for wearable systems.
• High classification accuracy is maintained, paving the way for robust wearable fNIRS applications.