Spike-Representation of EEG Signals for Performance Enhancement of Brain-Computer Interfaces

Sai Kalyan Ranga Singanamalla, Chin‐Teng LinFrontiers in Neuroscience 2022 · DOI

TL;DR

This paper proposes a novel method to improve Brain-Computer Interface (BCI) performance by converting noisy EEG signals into a "spike representation" using Spiking Neural Networks (SNNs). This transformation enhances the signal's robustness to noise and artifacts, leading to improved classification accuracy for critical BCI tasks like P300 and Error-Related Negativity (ERN).

What Problem Does This Paper Address?

Brain-Computer Interfaces (BCIs) are a promising technology for enabling communication and control for individuals with motor impairments. Electroencephalography (EEG) is a popular neuroimaging modality for BCIs due to its portability and non-invasiveness. However, EEG signals are inherently noisy and susceptible to artifacts, which significantly degrade the performance of BCI systems, especially for event-related potentials (ERPs) like P300 and Error-Related Negativity (ERN). These ERPs have biomarkers that appear in very short time segments, making them particularly vulnerable to noise. While invasive methods offer cleaner signals, they require complex surgery. This paper addresses the challenge of improving EEG-based BCI performance by leveraging the idea that EEG signals are an aggregate of underlying neuronal spiking activity. The core problem is to find a way to extract more robust and informative features from EEG that are less affected by noise and artifacts.

Methods

The authors propose a novel approach that utilizes a Spiking Neural Network (SNN) to transform EEG signals into a "spike representation." The process involves two main stages:

  1. Forward Process (Template Generation): An SNN, trained using surrogate-gradient descent, is used to generate "template" EEG signals for each class (e.g., target vs. non-target). This involves training the SNN to reproduce representative EEG signals from the dataset.
  2. Inverse Process (Spike Representation Extraction): The trained SNN model is then used in reverse to convert individual EEG samples into their corresponding spike representations. This is achieved by leveraging the learned weights of the SNN.

The SNN architecture itself is recurrent (RSNN) and uses either Leaky Integrate-and-Fire (LIF) or Spiking Neural Unit (SNU) neuron models. A double exponential synaptic filter is applied to the SNN's spike trains to generate a continuous signal, which is then transformed into a multi-channel EEG signal. The training of the SNN weights is performed using surrogate-gradient descent, a technique that allows for gradient-based optimization of SNNs despite the non-differentiable nature of spiking neurons.

Datasets

The study utilized publicly available datasets for two ERP-based BCI modalities:

  • P300 Dataset: Sourced from Van Veen et al. (2019).
  • Error-Related Negativity (ERN) Dataset: Sourced from Kappenman et al. (2021).

For both datasets, 20 subjects' data were used. The datasets were pre-processed with standard pipelines, including bandpass filtering, epoch extraction, and channel selection. For P300, two channels (FCz & Fz) were selected, and for ERN, two channels (Pz & Oz) were chosen.

Key Insights and Ideas

  • EEG as a Proxy for Spiking Activity: The fundamental idea is that EEG, despite its noisy nature, contains information about the underlying neuronal spiking activity. By using SNNs, which are biologically inspired by spiking neurons, we can potentially "reverse-engineer" this spiking information.
  • Spike Representation for Noise Robustness: Transforming EEG into a spike representation appears to make the signal inherently more robust to noise and artifacts. This is because the spike representation might capture the essential temporal dynamics of neural events more effectively than the raw, aggregated EEG signal.
  • SNNs as Feature Extractors: Beyond their role as classifiers, SNNs are demonstrated here as powerful feature extractors. The latent spike representation generated by the SNN captures information that is more discriminative for BCI tasks.
  • Improved Within-Class Correlation: Analysis using PCA and correlation metrics suggests that the spike representation enhances the similarity among samples belonging to the same class, making it easier for classifiers to distinguish between them.

Key Results

The proposed spike-representation method demonstrated significant improvements in classification accuracy for both P300 and ERN datasets:

  • P300 Dataset: Accuracy improved from a baseline of 67.73% (EEG signal) to 69.87% using xGboost with the SNU-based SNN spike representation.
  • ERN Dataset: Accuracy improved from a baseline of 79.22% (EEG signal) to 82.27% using Naive Bayes with the LIF-based SNN spike representation.

These improvements were statistically significant (p < 0.05) across various classifiers (SVM, NVB, xGboost). The study also showed that the SNN's trained weights played a crucial role in this performance enhancement, outperforming random weight transformations and Liquid State Machines.

Limitations and Future Work

  • Template Generation Dependency: The template signal extraction relies on the assumption that a given stimulus produces similar neural activity. However, templates can be noisy or contain irrelevant information, which might affect the spike representation. Future work could explore more robust template generation or representation learning techniques (e.g., supervised contrastive learning) to address this.
  • Node Prioritization Complexity: While node prioritization (selecting only the most informative neurons) showed minor performance gains, it requires estimating variance from multiple datasets, which can be cumbersome. Developing more efficient methods for node prioritization is suggested.
  • Computational Cost: Training SNNs and generating spike representations for each sample can be computationally demanding. Future research could focus on optimizing these processes for real-time applications.
  • Generalizability: While tested on ERP signals, further exploration of the spike-representation approach on other BCI modalities (like motor imagery) is warranted.

The authors suggest that future directions could involve exploring pruning and deep-rewiring techniques for SNNs to further optimize performance and efficiency.

Summary updated

Abstract

Brain-computer interfaces (BCI) relying on electroencephalography (EEG) based neuroimaging mode has shown prospects for real-world usage due to its portability and optional selectivity of fewer channels for compactness. However, noise and artifacts often limit the capacity of BCI systems especially for event-related potentials such as P300 and error-related negativity (ERN), whose biomarkers are present in short time segments at the time-series level. Contrary to EEG, invasive recording is less prone to noise but requires a tedious surgical procedure. But EEG signal is the result of aggregation of neuronal spiking information underneath the scalp surface and transforming the relevant BCI task's EEG signal to spike representation could potentially help improve the BCI performance. In this study, we designed an approach using a spiking neural network (SNN) which is trained using surrogate-gradient descent to generate task-related multi-channel EEG template signals of all classes. The trained model is in turn leveraged to obtain the latent spike representation for each EEG sample. Comparing the classification performance of EEG signal and its spike-representation, the proposed approach enhanced the performance of ERN dataset from 79.22 to 82.27% with naive bayes and for P300 dataset, the accuracy was improved from 67.73 to 69.87% using xGboost. In addition, principal component analysis and correlation metrics were evaluated on both EEG signals and their spike-representation to identify the reason for such improvement.
BibTeX
@inproceedings{singanamalla2022spike,
  author = {Sai Kalyan Ranga Singanamalla and Chin‐Teng Lin},
  title = {{Spike-Representation of EEG Signals for Performance Enhancement of Brain-Computer Interfaces}},
  booktitle = {Frontiers in Neuroscience},
  year = {2022},
  doi = {10.3389/fnins.2022.792318},
}
Figure