Robust long term neural signal decoding by estimating unobserved features
Vijay Aditya Tadipatri, Ahmed H. Tewfik, James Ashe
Abstract
Chronic effects of electrode implantation in the brain tissue alter the neural channel signal-to-noise ratio (SNR) over time. Variability of signal quality over time poses a difficult challenge in long-term decoding of neural signals for Brain Computer Interface (BCI). Specifically, all channels observed during a neural recording session may not be observed during the next recording session. This paper describes a novel approach that effectively overcomes these challenges by identifying reliable channels and features in any given trial, estimating unobservable or unreliable features and adapting the neural signal classifier with no user input in real time. The proposed decoder predicts one of eight arm directions with an accuracy, unmatched in the literature, of above 90% in two monkeys over 4-6 weeks, achieving robustness against time and also varying environmental conditions. Application of these decoders reduces neural prosthetic training time and user frustration thus improving the usability of BCI.
BibTeX
@inproceedings{icassp2015_robustlongtermne,
title = {Robust long term neural signal decoding by estimating unobserved features},
author = {Vijay Aditya Tadipatri and Ahmed H. Tewfik and James Ashe},
booktitle = {ICASSP 2015},
year = {2015}
}