the process' parameters change abruptly after a certain period of time. Although a single
adaptive system theoretically could model any function, such sudden change renders this
task very hard. An interesting approach to counter this problem is then to divide the time
series in several stationary or quasi-stationary regions, and identify the different models
corresponding to those regions [2]. This approach, called segmentation, is more
comprehensive than tracking, since it does not necessarily require forgetting past
properties to learn new ones, and a past model can be re-used on a new segment
belonging to the same regime. Many algorithms attempting to perform such tasks have
been developed in recent years, under the framework gated experts models [3], but have
not been extensively tested.
The goal of this thesis is to assess the performance and practicality of this
framework approach on modeling and performing online segmentation of a particular
type of real world non-stationary time series: data recorded from patients under assisted
ventilation. Medical time series constitute indeed a great challenge as far as data analysis
goes, because they are inherently non-stationary, often presenting impulse-type or
complicated shapes, and they depend on so many parameters that it is often difficult to
model them: in our case, a lung's behavior can be modeled in analogy with an RC
electrical circuit, R being the lung resistance and C the lung compliance (equivalent to
the capacity), and those are only two of the basic parameters that can change over time
creating non-stationarity. There are many directions of research in this field, but a lot of
the latest ones share a common point: instead of being fitted with linear models,
biomedical signals are analyzed through segmentation [4, 5] with non-linear or chaotic