Optimizing Time-Frequency Feature Extraction and Channel Selection through Gradient Backpropagation to Improve Action...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Voluntary motor intention can be predicted using neural oscillating patterns (time-frequency features) from local field potentials (LFPs).
• These LFPs are recorded from the sub-thalamic nucleus (STN) or thalamus in patients with movement disorders undergoing deep brain stimulation (DBS).
• Accurate decoding of motor states is crucial for advanced brain-computer interface (BCI) applications.

Purpose of the Study:

• To optimize signal conditioning for improved time-frequency feature extraction from LFP signals.
• To enhance the real-time decoding performance of voluntary motor states using deep learning.
• To develop and validate a BCI pipeline for classifying discrete pinch grip states.

Main Methods:

• A deep learning-based BCI pipeline was designed using Pytorch for offline analysis of LFP data.
• The pipeline focused on optimizing channel combinations and frequency domain feature extraction.
• LFPs were recorded from 5 patients with bilateral DBS electrode implants.

Main Results:

• Optimized signal conditioning and feature extraction significantly improved classification accuracy for pinch grip detection.
• The pipeline achieved a maximal average accuracy of 79.67±10.02% for detecting all pinches.
• Classification accuracy for identifying pinch laterality (left vs. right hand) reached 67.06±10.14%.

Conclusions:

• Deep learning-based optimization of signal conditioning enhances time-frequency feature extraction from LFPs.
• The developed BCI pipeline demonstrates improved performance in decoding voluntary motor states, specifically pinch grip actions.
• This approach holds promise for more effective BCI systems in patients with movement disorders.