Providing Case-Based Retrieval as a Decision Support Strategy in Time Dependent Medical Domains - Computational Intelligence in Healthcare: Advanced Methodologies - page 202

can be represented without the need of making it explicit in a more gener-
alized form, thus mitigating the well known knowledge acquisition bottle-
neck that affects other methodologies (such as Rule-based or Model-based
Reasoning). Proper maintenance policies [25] can then be implemented in
order to control the case base growth while preserving its problem solving
competence;
- cases allow to integrate different knowledge types . When generalized domain
knowledge is available, extracted from textbooks or physicians committees
expertise, and formalized by means of rules, ontologies, or computerized
guidelines, its integration with experiential and unformalized knowledge may
represent a significant advantage. Cases can be quite naturally adopted to
complement formalized knowledge, and to make it operational in a real set-
ting, as testified e.g. by the wide number of multi-modal reasoning systems
proposed in the literature [29].

Various Case-based Reasoning/retrieval works dealing with cases with time
series features have been recently published, in the medical domain (see sections 2
and 3), as well as in different ones (e.g. robot control [42], process forecast [34,43],
process supervision [13], pest management [9], prediction of faulty situations
[19]). Moreover, general (e.g. logic-based) frameworks for case representation in
time dependent domains have been proposed [32,18,27,8].
However, adopting case-based retrieval can be non trivial in time dependent
applications, since the need for describing the process dynamics impacts both
on case representation and on retrieval itself, as analysed in [32]. As a matter of
fact, in classical CBR, a case consists of a problem description able to summarize
the problem at hand, and of a case solution , describing the solution adopted for
solving the corresponding problem. The problem description is typically repre-
sented as a collection of pairs. However, for intrinsically data

intensive applications, where data come in the form of time series, data them-
selves cannot be simply stored as feature values “as they are”. Pre-processing
techniques are required in these situations, in order to simplify feature mining
and knowledge representation, and to optimize the retrieval activity.
Most of the approaches proposed in the literature to this end are founded
on the common premise of dimensionality reduction , which allows to reduce
memory occupancy and to simplify time series representation, still capturing the
most important characteristics of the time series itself.
Dimensionality is typically reduced by means of a mathematical trans-
form , able to preserve the distance between two time series (or to underestimate
it). Widely used transforms are the Discrete Fourier Transform (DFT) [2], and
the Discrete Wavelet Transform (DWT) [10]. Another well known mathemati-
cal methodology is Piecewise Constant Approximation (PCA) [22,23]. Retrieval
then works in the transformed time series space, and, in many cases, operates
in a black-box fashion with respect to the end users: they just input the query,
and collect the retrieved cases, but usually do not see (and might not understand
the meaning of) the transformed time series themselves. Details of mathematical
methods for dimensionality reduction and time series retrieval will be presented

feature,value