EEG as a means of communication: preliminary experiments in EEG analysis using neural networks
Charles W. Anderson, Saikumar V. Devulapalli, Erik A. Stolz · 1994 · Proceedings of the First Annual ACM Conference on Assistive Technologies (Assets '94) · doi:10.1145/191028.191071
Summary
This paper explores the use of electroencephalography (EEG) as a communication channel for paralyzed individuals, representing early brain-computer interface (BCI) research. The core premise is that if distinct mental states can be reliably identified through EEG pattern recognition, a paralyzed person could communicate with devices like wheelchairs by composing sequences of these mental states. The authors acknowledge that EEG pattern recognition is inherently difficult and depends critically on finding appropriate signal representations in which patterns become distinguishable. The study compares three different representations of EEG signals: raw signals, a reduced-dimensional representation using the Karhunen-Loeve (K-L) transform, and a frequency-based representation. Classification was performed using a two-layer neural network implemented on a CNAPS server — a 128-processor SIMD (Single Instruction, Multiple Data) parallel architecture by Adaptive Solutions, Inc., reflecting the computational demands of real-time EEG processing even with 1990s neural network approaches.
Key findings
The frequency-based representation of EEG signals yielded the best classification accuracy at 73% on untrained (test) samples, outperforming both the raw signal representation and the K-L transform reduced-dimensional approach. This result indicated that spectral features of EEG — the distribution of signal energy across different frequency bands — contain the most discriminative information for distinguishing between mental states. While 73% accuracy was insufficient for practical communication, it demonstrated the feasibility of using neural networks to classify EEG patterns and established a baseline for future BCI research. The comparison of three representation methods provided valuable methodological guidance, showing that signal preprocessing and feature extraction choices significantly impact classification performance — a finding that would be confirmed repeatedly in subsequent BCI research.
Relevance
This paper is an early entry in what has become one of the most active areas of assistive technology research: brain-computer interfaces. Published in 1994, it predates the major expansion of BCI research that occurred in the late 1990s and 2000s. The work is notable for framing EEG-based communication explicitly as an assistive technology for paralyzed individuals at a time when most EEG research focused on clinical diagnostics or neuroscience. The use of neural networks for EEG classification was prescient — deep learning and neural network approaches now dominate modern BCI signal processing. For practitioners, the paper illustrates both the promise and limitations of brain-computer interfaces: while the technology offers a communication pathway for people with the most severe motor impairments (such as locked-in syndrome), achieving the accuracy and reliability needed for practical use remains an ongoing challenge three decades later.
Tags: brain-computer interface · electroencephalography · neural network · machine learning · paralysis · assistive technology · pattern recognition · alternative input