creates the "build up" problem when the instantaneous error is too high. Another way to

compute an estimate of the MSE is to use a boxcar average of the instantaneous error:

this approach would give as much weight to all past instantaneous error values used to

compute it (so the change detection might not be as fast), but within one window length

all previous error values are forgotten so the buildup problem would be restricted to that

length. Unfortunately, looking at the memory depth used for the models (one hundred

samples on the test set, fifty samples on the real data), it appears that using a boxcar

average of the instantaneous error in this case would solve the problem: most of the

difference in waveform between the two regimes is represented by the samples gathered

right after the end of the inspiratory phase, and lasts for less than forty samples, then the

expiratory phase follows the same model in the two regimes. So in a two-experts system,

changing the error estimate would not help, but with a three-experts system modeling the

two different inspiratory phases and the expiratory phase, the memory depth could maybe

be decreased, so would a window size for the boxcar estimate that could then become

more interesting.

 5.3 Simulations with Another Set

5.3.1 Description of the data

 The data set previously used was a particular case of respiratory signal under

assisted ventilation, but many different ones exist: indeed, the set waveform of the

delivered flow can be different (square or sine for example), individual breathing patterns

can be different, and also the patient might fight the ventilator.

 Figure 5-7 shows another data set at 50 Hz with two different kinds of breaths, and

we can see that the waveforms are not similar to the ones of the former data set.