Information Technology Reference
In-Depth Information
broadly categorized as probability density estimation, nearest neighbor search, and
neural networks [88, 66, 2]. In particular, probability density estimation or Bayes es-
timation categorizes the measurement in order to minimize the probability of error,
nearest neighbor search finds the class that is associated with the nearest neighbors
of the measurement, while neural networks consist of simple interconnected compu-
tational elements that have the end result of dividing the feature space into specific
regions [59, 60, 58].
Among these classifiers, neural networks seem to be the most widely used meth-
ods in biomedical applications. However, choosing the best classifier as well as a
feature set for a particular case is often an empirical task. Thus, a set or “ensemble”
of different classifiers is often used for a single classification task [116].
13.6.3 Feature Selection
Selecting the features that minimize a cost function, such as the probability of
misclassification, can be done exhaustively by examining each subset. However,
this process is of complexity
N
n
and may become intractable for large feature
sets. Alternatively, there are a number of methods that reduce the complexity of the
task, including “branch and bound,” “sequential forward and backward selection,”
“Plus-l-take-away-r algorithm,” and “max-min feature selection” [122, 19, 118].
13.7 Closed-Loop DBS
Following the discovery of the effects of electrical brain stimulation on the symp-
toms of Parkinson's disease [13] in 1987, investigations were initiated to explain
how the stimulus achieved the desired result [101, 54]. Also, methods for admin-
istrating the newfound treatment as an implantable “brain pacemaker” were being
explored [106, 146, 134, 54, 127, 78, 39]. In particular, the first disclosure of such
an apparatus was the original patent on DBS filed by Rise and King [127] of the
Medtronic corporation in 1996, where a system consisting of an electrode sensor,
a microprocessor, stimulus generator, and additional peripheral circuitry was pro-
posed for the purpose of measuring tremor-related symptoms in the arm and adjust-
ing stimulus parameters based on the measurements. Subsequently, another patent
was filed by John [78] in 2000, elaborating on the original proposal by including
provisions for multiple sensors such as electrodes implanted in the brain and/or
surface electrodes on the scalp and limbs. In addition, John proposed particular sig-
nal processing methods for assessing the measured data including the computation
of signal variance, correlation, discrete Fourier transform, peak detection, and Ma-
halanobis distance or
Z
-scores. Also, provisions for wireless data telemetry to an
external PC or handheld processor were included in that patent.
