Panel 4: Signal Analysis

1. Robust techniques are important and need to be looked into further (3 of the 6 panelists are using robust techniques)
2. Feedback is a major concern and needs to be evaluated. Biasing a subject's feedback during training could be of help.
3. Qualities which affect on-line signal processing results such as subject expectations, motivation, distraction, and frusteration need to be accounted for and controlled as much as possible.
4. Artifact remains a major consideration and better routines are always welcome.

Audience comments: Bipolar recording can eliminate 60 Hz (the montage matters).

1. Motivate subjects.
2. Fix a noisy/distracting environment.
3. Fix a poor user interface
4. Work perfectly!

Audience comments: It would be helpful to have a measure of subject frusteration (or lack thereof).

• AR Model Parameter Estimation
  1. LSQ estimation is optimal for Gaussian processes
  2. LSQ estimation can break significantly with additive outlier contamination
  3. Robust Estimation Method: Generalized Robust Maximum Likelihood Estimate (GM-estimate) see R.D. Martin and D.J. Thomson, "Robust-resistant spectrum estimation", Proc. IEEE, no. 9, vol. 70, pp. 1097-1115, 1982.
• AR Spectrum Estimation
  1. Direct Estimate using an AR model from a robust estimate of the parameters
  2. Using a prewhitening method again utilizing robustly estimated AR parameters
• Robust Signal Estimation
  1. Robust Signal Estimator is based upon a modified Kalman Filter (see Martin and Thomson)
  2. Produces an estimate of the signal without the influence of additive outliers
• The nature of additive outliers
  1. to qualify as a time-series outlier, it only has to be "different" on the innovations (residual) scale, not the process scale
  2. Innovation scale is typically 10-10,000 times smaller than the process scale
  3. Outliers will often be impossible to detect in a visual plot of the raw data
• Effects of additive outliers on LSQ-based estimation
  1. GM methods perform almost as well as LSQ methods on uncontaminated Gaussian data
  2. with a 10 per cent level of additive outlier contamination, LSQ parameter estimates can have mean squared errors four to five times higher than GM methods (see example)
• Estimation of Event Related Potentials
• Examples of work that have utilized AR modeling in the signal processing of EEG signals: Birch, Pfurtscheller, Cerutti, Jansen, Smith, Lager, Spreckelsen

Audience Comments:

• (???) When the data is assumed nonstationary, one should use short data segments, but how long a time window is necessary?
• (Donchin) Should measure the effectiveness of feedback and possibly use the error-related potential for error detection. a delay in feedback impedes performance.
• (???) feedback should be evaluated: perhaps using a frusteration index
• (???) Might want to bias feedback in the beginning for a subject
• (Vaughan) Feedback early in the training in our system failed.
• (McFarland) What subjects expect is important!
• (Penny) A feedback bias may be adaptive on-line
• (Kostov) A subject can be manipulated to think that they're doing better or worse than in reality.
• (???) Might want to standardize potential ways for telling if a subject is trying to manipulate the system with something other than EEG signals
• (Wolpaw) EMG is an obvious artifact and subjects can control the mu rhythm with muscle. We can't prove subjects aren't doing this, but we can show that EMG is decorrelated with performance. Note that if we have people tense their muscles it is not the same as what people might naturally be doing during the experiment. If there is a sharp, focussed band peak in the spectral band, then the signal is most likely not EMG.
• (Makeig) Muscles stay in one place and are stationary, thus (independent component analysis) ICA can be used to remove muscle activity from contaminated signals.
• (Wolpaw) That depends on the muscle signal activity.
• (Makeig) ICA can be a method for separating different individual muscle signals from each other.
• (Gevins) That statement is too strong. (Note taker's note: a discussion followed, but no conclusion was reached.)

Major factors in choosing a model

Discussed by Christof Guger

Example discussed: AAR model vs. CSP (see slides)

Audience Comments:

• (Gevins) Why use 27 electrodes? Answer: the algorithm was trained off-line on pilot data and the number of needed coefficients were decided by the discrimination of the task
• (Kostov) How did you optimize the order of the AAR? Answer: empirically
• (Penny) What happens if the AAR gives feedback? Answer: it's almost the same
• (Donchin) Are the differences between the two techniques really different? CSP was only 5\% better and is a more difficult algorithm.
• (Gevins) What is the error? Answer: Screen position of the cursor (binary hit or miss).
• (Donchin) There is a trade-off between an increase in errors and the amount of time it takes for the chosen algorithm to run. What is the cost of errors? It's application specific.
• (Hinterberger) 160 trials of data per subject.
• (Wolpaw) Everybody can use bit rate as a measure.
• (Mason) We need to evaluate techniques independent of the application.
• (Gevins) Can consider using a laplacian.

The Thought Transform Device

Discussed by Thilo Hinterberger

See Powerpoint slides for details.

Audience Comments: (Gevins) Do you place electrodes on the face rather than the head? Answer: yes. (Gevins) What about skin potentials? Answer: can't control for that.

Artifact Rejection

Discussed by Will Penny

It's important to take into account uncertainty and reject outliers. See this page or this document for more details.

Audience Comments:

• (Kostov) Artifacts can overlap with your class categories, so it might not be possible to reject them. coefficients were decided by the discrimination of the task
• (Gevins) Rejecting outliers is like having a can't decide category which Jacques Vidale did 30 years ago.
• (McFarland) What if artifact is in your training set? Need to have a clean training set for this method.
• (Robinson) Can the priors be defined?