Detection of Multiple Sclerosis using Neural Network Ensembles
Abstract
Event related potentials, or ERPs, are averaged waveforms of EEG cerebral activity measured using non-invasive electrodes attached to the scalp. Certain diseases resulting in neuropsychological impairment seem to result in changes in the waveform and latency of the P300 ERP. One example of this type of disease is multiple sclerosis. Experiments using artificial neural networks to analyze ERP data sets and discriminate between multiple sclerosis patients and a control population have been performed in previous research. These experiments yielded classification accuracies of 81% for single channel (electrode) training and 90% based on a vote of 2 out of 3 single channel outputs. This research describes a series of experiments on the same data sets undertaken with the goal of seeking higher accuracies. Single channel training using inputs that were pruned to isolate the P300 resulted in classification accuracies approaching 86%. by constructing a multi-channel ensemble network that used raw outputs of the single channel networks as inputs, classification accuracies of 96% were achieved.
Recommended Citation
G. Eastman and D. C. St. Clair, "Detection of Multiple Sclerosis using Neural Network Ensembles," Intelligent Engineering Systems Through Artificial Neural Networks, vol. 13, pp. 821 - 826, American Society of Mechanical Engineers, Dec 2003.
Department(s)
Mathematics and Statistics
Document Type
Article - Conference proceedings
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2024 American Society of Mechanical Engineers, All rights reserved.
Publication Date
01 Dec 2003
