A Comparative Study of Time Delay Neural Networks and Hidden Markov Models for Electroencephalographic Signal Classification

In this paper, we analyze the performance of Time Delay Neural Networks (TDNN) and Hidden Markov Models (HMM) for Electroencephalogram (EEG) signal classification. The specific focus of this study is Brain-Computer Interfacing (BCI), where near-real time detection of mental commands during a multi-channel EEG recording is desired. We argue that HMM and TDNN should be preferred over the rigid, one-size-fits-all methods of the more traditional EEG signal classifiers. To analyze the utility of modern classification methods for BCI, we compare and discuss the performance of our suggested TDNN and HMM EEG classifiers with the reported best results on BCI 2003 EEG benchmark dataset Ia. INTRODUCTION The goal of Brain Computer Interfaces (BCI), also known as Thought Translation Devices (TTD), is to provide individuals with a new communication channel that conveys commands directly from the brain (Wolpaw 2002). Other areas that may benefit from this area of research include diagnosis of memory problems, cognitive development of the children (Taylor and Baldeweg 2002), diagnosis and treatment of attention-deficit disorder, and many other medical conditions that need tools for classification of brain’s electrophysiological activities. Classification of Electroencephalogram (EEG) is an important part of current BCI research. Using EEG-based BCI is arguably superior to other related modalities such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET), since the brain signals of interest happen at a rate that is only within the temporal resolution of EEG. Moreover, compared to PET and fMRI, EEG is convenient, portable, and affordable. In this paper we examine the application of two nonlinear intelligent signal analysis tools to non-invasively collected EEG signals in order to classify the spatiotemporal signatures of imagined commands; as while the researchers in neurobiology have mostly been utilizing linear discriminant for BCIs, one can argue that such approach is mathematically justifiable only if the dimension of the feature space is high enough. Many attempts have been made to build an EEG-based BCI system (Wolpaw 2002). Generally speaking, the most important steps of these BCI systems are feature extraction and classification. For feature extraction, adaptive auto regressive models, Hjorth parameters, power spectrum, and principle component features have been used (Obermaier 2001, Keirn and Aunon 1990). Various classifiers have also been used,