A machine learning approach for automated wide-range frequency tagging analysis in embedded neuromonitoring systems

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Electroencephalography (EEG) is crucial for neuroscience and neural disease diagnostics.
• Frequency-tagging paradigms efficiently assess brain function via periodic stimulation.
• Low-frequency EEG signals (<6Hz) are inherently noisy, complicating automated analysis and device design.

Purpose of the Study:

• To propose an algorithm for artifact removal and automated detection of EEG frequency-tagging responses.
• To enable efficient EEG analysis on resource-constrained systems and smart diagnostic devices.
• To validate the algorithm's performance across a wide range of stimulation frequencies, including low ones.

Main Methods:

• Developed a machine learning-based pattern recognition algorithm for EEG signal processing.
• Implemented artifact removal and automated frequency-tagging response detection.
• Tailored the algorithm for a parallel ultra-low-power (PULP) processing platform.
• Tested the algorithm on a visual stimulation protocol.

Main Results:

• Achieved over 90% accuracy in frequency detection, even for very low stimulation frequencies (<1Hz).
• Demonstrated effective artifact removal and automated response detection.
• Algorithm operates within a low power budget of 56mW.

Conclusions:

• The proposed algorithm significantly improves automated EEG analysis for frequency-tagging paradigms.
• Enables the development of smart, low-power EEG devices for automated diagnostics.
• Overcomes limitations of traditional EEG processing in noisy, low-frequency ranges.