Wavelet Shrinkage and Thresholding Based Robust Classification for Brain-Computer Interface

Taposh Banerjee, John Choi, Bijan Pesaran, Demba Ba, Vahid Tarokh

Abstract

A macaque monkey is trained to perform two different kinds of tasks, memory aided and visually aided. In each task, the monkey saccades to eight possible target locations. A classifier is proposed for direction decoding and task decoding based on local field potentials (LFP) collected from the prefrontal cortex. The LFP time-series data is modeled in a nonparametric regression framework, as a function corrupted by Gaussian noise. It is shown that if the function belongs to Besov bodies, then the proposed wavelet shrinkage and thresholding based classifier is robust and consistent. The classifier is then applied to the LFP data to achieve high decoding performance. The proposed classifier is also quite general and can be applied for the classification of other types of time-series data as well, not necessarily brain data.

Introduction

Brain-Computer Interfaces (BCIs) aim to translate neural activity into control signals for external devices. A primary challenge in BCI development is the inherent noise and complexity of Local Field Potentials (LFPs). While traditional methods often rely on power spectral density or Fourier-based analysis, these approaches assume a high degree of signal smoothness (Sobolev classes). However, neural signals frequently exhibit transient, "spiky," or sparse characteristics that are poorly captured by global frequency transforms. This research introduces a robust classification framework using wavelet shrinkage and thresholding, designed to handle a broader class of signals known as Besov bodies.

Methods: Wavelet Shrinkage and Thresholding

The proposed method treats the LFP signal as a function $f$ observed with additive Gaussian noise. To extract robust features, the authors employ a three-step pipeline:

  1. Discrete Wavelet Transform (DWT): The time-series data is decomposed into wavelet coefficients, providing localized information in both time and frequency domains.
  2. Soft-Thresholding: A shrinkage function is applied to the coefficients. By setting coefficients below a specific threshold to zero and reducing others, the signal is effectively denoised. This approach is mathematically proven to be "minimax" optimal for Besov bodies, ensuring the estimator performs well even in worst-case noise scenarios.
  3. Feature Extraction and Classification: Denoised coefficients are processed via Principal Component Analysis (PCA) for dimensionality reduction across multiple electrode channels. A Linear Discriminant Analysis (LDA) classifier is then used for final categorization.
LFP signal approximation using Haar wavelet coefficients after shrinkage.

Experimental Setup

The classifier was validated using LFP data from a macaque monkey performing two distinct saccade tasks across eight possible target directions. As shown in the experimental paradigm, the tasks involve different cognitive demands:

  • Memory-aided task: The subject must saccade to a remembered target location after a delay period.
  • Visually-aided task: The target remains visible throughout the trial.

Data was recorded using a 32-micro-electrode array implanted in the prefrontal cortex, totaling 1,467 trials across nine sessions.

Experimental paradigm for memory-aided and visually-aided saccade tasks.

Results and Discussion

The wavelet-based approach demonstrated exceptional decoding performance:

  • Direction Decoding: The model achieved an average accuracy of 88% in predicting eye movement direction across eight targets. This performance is comparable to previous Fourier-based methods but offers superior theoretical robustness for non-smooth signals.
  • Task Decoding: The classifier distinguished between memory-aided and visually-aided tasks with 98.5% accuracy.

The transition from Sobolev to Besov spaces allows the model to accommodate a wider variety of neural signal morphologies. By utilizing a minimax estimator as a feature, the system maintains consistency across different signal-to-noise ratios, making it a strong candidate for practical BCI applications.

Conclusion and Future Work

This study presents a mathematically grounded, robust framework for LFP classification. While the results are promising, future work will focus on validating the model across multiple subjects and optimizing parameter selection (e.g., wavelet type and threshold levels) to enable real-time, "plug-and-play" BCI systems. The generality of the nonparametric regression framework also suggests potential applications in other time-series domains beyond neuroscience.

Summary updated

Abstract

A macaque monkey is trained to perform two different kinds of tasks, memory aided and visually aided. In each task, the monkey saccades to eight possible target locations. A classifier is proposed for direction decoding and task decoding based on local field potentials (LFP) collected from the prefrontal cortex. The LFP time-series data is modeled in a nonparametric regression framework, as a function corrupted by Gaussian noise. It is shown that if the function belongs to Besov bodies, then the proposed wavelet shrinkage and thresholding based classifier is robust and consistent. The classifier is then applied to the LFP data to achieve high decoding performance. The proposed classifier is also quite general and can be applied for the classification of other types of time-series data as well, not necessarily brain data.
BibTeX
@inproceedings{banerjee2018wavelet,
  author = {Taposh Banerjee and John Choi and Bijan Pesaran and Demba Ba and Vahid Tarokh},
  title = {{Wavelet Shrinkage and Thresholding Based Robust Classification for Brain-Computer Interface}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2018},
  url = {https://openreview.net/forum?id=AHAI1GMx3Q},
}
Figure