Brain–Computer Interfaces - Index of

324 Y. Li et al.
(ROC) analysis approach, we first introduce several concepts. Let us consider a
two-class prediction problem (binary classification), in which the outcomes are
labeled as either a positive (p) oranegative(n) class. There are four possible
outcomes from a binary classifier. If the outcome from a prediction is p
and the actual value is also p, then it is called a true positive (TP); however,
if the actual value is n then it is said to be a false positive (FP). Similarly, a
true negative (TN) implies that both the prediction outcome and the actual value
are n, while the false negative (FN) implies the prediction outcome is n however
the actual value is p. A ROC curve is a graphical plot of sensitivity versus
(1-specificity), or equivalently, a plot of the true positive rate versus the false positive
rate for a binary classifier system as its discrimination threshold is varied
(http://en.wikipedia.org/wiki/Receiver_operating_characteristic).
ROC analysis [53] has become a standard approach to evaluate the sensitivity
and specificity of detection procedures. For instance, if a threshold θ is assumed to
be a detection criterion, then a smaller θ indicates a higher sensitivity to events but a
lower specificity to the events of interest. This is because those events that are of no
interest are also easily detected. ROC analysis estimates a curve that describes the
inherent tradeoff between sensitivity and specificity of a detection test. Each point
on the ROC curve is associated with a specific detection criterion. The area under
the ROC curve has become a particularly important metric for evaluating detection
procedures because it is the average sensitivity over all possible specificities.
An example is presented here to illustrate the use of ROC curve analysis in evaluating
asynchronous BCIs [52]. The two axes of the ROC curves are the true positive
rate (TPR) and the false positive rate (FPR). These quantities are captured by the
following equations:
TPR(θ) =
TP(θ)
FP(θ)
, FPR(θ) =
, (24)
TP(θ) + FN(θ) TN(θ) + FP(θ)
Fig. 8 Sample-by-sample evaluation of true and false positives (TPs and FPs, respectively) and
true and false negatives (TNs and FNs, respectively). Events are defined as the period of time from
2.25 to 6.0 s after the rising edge of the trigger, which occurs 2.0 s after the start of each trial [52]