each breath was initiated. The goal of this thesis is then to "classify" each breath, or

segment the times series between the two different modes: mandatory or spontaneous

breath.

3.1.2 Concerns

 Considering the data and the task we propose to study, several problems appear.

 To perform the segmentation, we need to use one of the times series and try to

model it. The pressure exhibits a near square waveform, with very high and very low

frequencies, and this would be hard for any adaptive system to model, so the flow

waveform is chosen (and since it is centered around zero and has an amplitude of two

L.sec-1 whereas the pressure spikes at forty cmH2 0., is does not need any preprocessing

before being used in a non-linear model). But the segmentation is visually more easily

done on the pressure graph, so it is kept as a visual reference.

 Another problem is the obvious non-stationarity of the real data sample, due to

several factors: as said previously, the lung parameters are not constant in reality, the

patient might become agitated and breathe too fast for the ventilator settings, there is

noise in the recording that can be non-stationary, and other parameters could change. In

order to compare the performance of the several models and segmentation methods, we

need to have some control over the data. For that purpose, we create a synthetic test set

using the real samples, so that some control is gained over the variability of the original

data while its specificity is preserved

 3.2 Synthetic Data Set Generation

 As just stated, a synthetic test set is important because it allows us to test the

systems in a controlled environment, to know that there are only two regimes, meanwhile

keeping some of the nonstationarity and versatility of the original data, which still gives