Intelligent Adaptive Monitoring for Cardiac Surveillance - Computational Intelligence in Healthcare: Advanced Methodologies - page 317

detailed. Learning chronicles and decision rules are presented in Section 8. We provide
some evaluation results in Section 9 before concluding.

2 Related Work

Stacey and McGregor's survey [35] lists many intelligent monitoring systems that share
several objectives and features with ours. To cite a few, R ESUME [34], VIE - VENT [22]
and its successors like ASGAARD [33] have introduced general knowledge representa-
tion paradigms for temporal abstraction with a deep integration of domain knowledge.
VIE - VENT could process data streams whereas R ESUME was limited to databases. For
efficiency reasons, instead of being general, our work focusses on cardiac knowledge
in order to extract rich information from ECG or pressure signals online. From the tem-
poral reasoning and diagnosis point of view, N EO - GANESH [8] shares the concept of
temporal patterns, called scenarios, with our approach. Recent developments have led
to knowledge-based temporal abstraction (KBTA) where machine learning techniques
are used to extract the most discriminating patterns which can be identified in normal
and several pathological states. There is a wide literature about temporal abstraction
in medical domains [20,35]. The Calicot system [3], dedicated to cardiac arrhythmias
detection, and the one proposed by Guimaraes [13] which is focused to sleep-related
respiratory disorders are two examples. Collaborative knowledge discovery approaches
exploiting the properties of multi-agent systems (open character, autonomy of its com-
ponents) have been proposed recently[16] for the exploration of mechanical ventilation
asynchronies.
Self-adaptation techniques have been advocated to cope with changes in the envi-
ronment of the computation. This has been an issue for many years in the monitoring
field, especially in the medical domain. With the advent of pervasive and ubiquitous
computing, self-adaptive software is becoming a major concern in the field of software
engineering (see [21] for an overview).
Two main operational paradigms have been used for self adaptive software: dynamic
planning and control theory-like architecture [29]. Dynamic planning has been used
for adaptive monitoring or computation in medical systems such as G UARDIAN [19],
A SGAARD [32] or in image processing [2]: giving some goals (describing a therapy or
a final execution state), a plan is computed by planning or instantiating a plan model.
The plan operations are followed until observations contradict plan expectations. Past
and actual plan operations can be used to evaluate the execution of the plan. Future
plan operations can be used to focus data abstraction to specific processing and to adapt
the involved algorithms. As in our approach, these systems provide upwards and down-
wards control. However, except for G UARDIAN where a knowledge-based approach is
advocated, it is not clear how adaptation is implemented in those systems. In our case,
a declarative approach based on event action rules is used.
An alternative method consists in viewing adaptation as in control theory: a super-
vision module analyzes continuously inputs and outputs in order to reconfigurate the
system by selecting and adapting the processing components. This approach has been
adopted for self-adaptive signal processing [17] or self-adapting numerical software ap-
plication [9], for instance. Like many other approaches in software engineering [21], the