Single-Trial Classification of Bistable Perception by Integrating Empirical Mode Decomposition, Clustering, and...

Area of Science:

• Neuroscience
• Signal Processing
• Computational Biology

Background:

• Decoding brain activity from local field potentials (LFPs) is crucial for understanding perception and developing brain-computer interfaces (BCIs).
• Nonstationary and multivariable time series data, such as LFPs, present challenges for traditional feature extraction methods.
• Bistable perception, like structure-from-motion (SFM), offers a paradigm to study neural correlates of conscious experience.

Purpose of the Study:

• To propose and validate an Empirical Mode Decomposition (EMD)-based feature extraction method for decoding bistable structure-from-motion (SFM) perception from macaque monkey LFP recordings.
• To investigate the effectiveness of different frequency bands and clustering approaches for feature extraction.
• To demonstrate the potential of this method for applications in brain-computer interfaces (BCIs) and evoked potential estimation.

Main Methods:

• Empirical Mode Decomposition (EMD) to decompose nonstationary LFP time series into intrinsic mode functions (IMFs).
• Unsupervised K-means clustering to group IMFs and residues across trials and channels.
• Supervised Common Spatial Patterns (CSP) to design spatial filters for clustered spatiotemporal signals.
• Support Vector Machine (SVM) classifier for single-trial perception decoding.

Main Results:

• The CSP features from the gamma frequency band cluster yielded the highest decoding accuracy for SFM perception.
• The EMD-based feature extraction method demonstrated robust performance in decoding perceptual states.
• The approach showed potential for evoked potential estimation, indicating broader applicability.

Conclusions:

• The proposed EMD-based feature extraction method is effective for decoding complex perceptual states from LFP data.
• This technique offers a powerful tool for analyzing nonstationary multivariable time series in neuroscience and BCI research.
• The gamma band features derived from EMD-clustered LFPs are particularly promising for future brain-computer interface development.