Fast, accurate, and interpretable decoding of electrocorticographic signals using dynamic mode decomposition.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Dynamic Mode Decomposition (DMD) decomposes spatiotemporal signals into fundamental oscillatory components.
• Current kernel-based machine learning methods using DMD (e.g., nonlinear Grassmann kernel) enhance neural decoding accuracy but face limitations in computational time, algorithm compatibility, and interpretability.
• Real-time and interpretable neural decoding remains a significant challenge in neuroscience.

Purpose of the Study:

• To develop a novel mapping function for transforming DMD into spatial DMD (sDM) features.
• To enable the use of these sDM features in any machine learning algorithm, overcoming limitations of kernel-based approaches.
• To enhance the speed, accuracy, and interpretability of neural decoding.

Main Methods:

• Proposed a mapping function to convert DMD into spatial DMD (sDM) features.
• Applied sDM features to electrocorticographic (ECoG) signals from movement and visual perception tasks.
• Evaluated decoding accuracy and computational time against conventional methods.

Main Results:

• sDM features significantly improved neural decoding accuracy and reduced computational time compared to conventional methods.
• sDM features demonstrated higher trial-to-trial reproducibility than high-gamma power, a common neural signal feature.
• The informative components of sDM features exhibited characteristics similar to high-gamma power.

Conclusions:

• The proposed sDM features provide a computationally efficient and accurate method for neural decoding.
• sDM features enhance interpretability by offering insights into signal components relevant for decoding.
• This approach facilitates real-time neural decoding applications across various machine learning algorithms.