of finite memory like the boxcar, and give more weight to the latest prediction error

values like the recursive gamma average).

 Also, a quantitative analysis of the segmentation history could be performed by

computing a Receiver Operating Characteristic [29], differentiating false alarm from

detection by a threshold on the number of concurrent samples without change after a

switch, and assessing the performance of the system in terms of accurately detected

switches with respect to spurious switches.

 Lastly, it is important to note that for the particular case of ventilatory data where

the segmentation is based on differences in waveforms, a more straightforward approach

might be to use some feature detection or template matching system, that focus on the

shape of the signal without depending on the length of a breath or its amplitude like the

competitive experts do.