[Paper Review] Recognition of Words from the EEG Laplacian

J. Acácio de Barros, C. G. Carvalhaes+2 more|arXiv (Cornell University)|Apr 26, 2012

EEG and Brain-Computer Interfaces12 references3 citations

TL;DR

This study proposes using the EEG Laplacian to enhance word recognition from brain signals by reducing artifacts from distant neural sources. By applying spherical spline interpolation and bandpass filtering to averaged EEG responses, the method improves recognition accuracy for seven auditory words, demonstrating consistent performance across subjects and highlighting spatial isomorphism in neural responses.

ABSTRACT

Recent works on the relationship between the electro-encephalogram (EEG) data and psychological stimuli show that EEG recordings can be used to recognize an auditory stimulus presented to a subject. The recognition rate is, however, strongly affected by technical and physiological artifacts. In this work, subjects were presented seven auditory simuli in the form of English words (first, second, third, left, right, yes, and no), and the time-locked electric field was recorded with a 64 channel Neuroscan EEG system. We used the surface Laplacian operator to eliminate artifacts due to sources located at regions far from the electrode. Our intent with the Laplacian was to improve the recognition rates of auditory stimuli from the electric field. To compute the Laplacian, we used a spline interpolation from spherical harmonics. The EEG Laplacian of the electric field were average over trials for the same auditory stimulus, and with those averages we constructed prototypes and test samples. In addition to the Laplacian, we applied Butterworth bandpass digital filters to the averaged prototypes and test samples, and compared the filtered test samples against the prototypes using a least squares metric in the time domain. We also analyzed the effects of the spline interpolation order and bandpass filter parameters in the recognition rates. Our results show that the use of the Laplacian improves the recognition rates and suggests a spatial isomorphism between both subjects.

Motivation & Objective

• To improve the recognition of auditory words from EEG signals by reducing noise and artifacts from distant brain sources.
• To investigate the effectiveness of the surface Laplacian operator in enhancing EEG signal quality for cognitive decoding.
• To evaluate the impact of spline interpolation order and filter parameters on recognition performance.
• To explore spatial consistency (isomorphism) in neural responses across subjects using Laplacian-transformed EEG data.

Proposed method

• Applied the surface Laplacian operator to EEG data using spherical spline interpolation based on spherical harmonics to suppress volume-conduction effects.
• Averaged EEG responses across trials for each of the seven auditory words (first, second, third, left, right, yes, no) to form prototypes.
• Preprocessed test samples with Butterworth bandpass digital filters to isolate relevant frequency bands.
• Used a least squares metric in the time domain to compare filtered test samples against prototype signals for classification.
• Systematically varied spline interpolation order and filter parameters to assess their impact on recognition accuracy.

Experimental results

Research questions

• RQ1 Can the EEG Laplacian improve recognition rates of auditory words compared to raw EEG signals?
• RQ2 How does the choice of spline interpolation order affect the quality of the Laplacian estimate and subsequent recognition performance?
• RQ3 What is the optimal bandpass filter configuration for enhancing word recognition from Laplacian-transformed EEG data?
• RQ4 Is there a consistent spatial pattern (spatial isomorphism) in neural responses to the same words across different subjects?

Key findings

• The use of the EEG Laplacian significantly improved word recognition rates by reducing artifacts from distant neural sources.
• Higher-order spline interpolation led to better Laplacian estimates and improved recognition accuracy, with optimal performance observed at specific interpolation levels.
• Butterworth bandpass filtering enhanced recognition by isolating relevant neural oscillations, with peak performance at specific frequency ranges.
• Recognition rates were consistent across subjects, suggesting a spatial isomorphism in neural responses to the same auditory stimuli.
• The combination of Laplacian transformation and filtered signal comparison achieved reliable classification of the seven target words.