Information Technology Reference
In-Depth Information
What the operator really needs is a decision support system that could help him decide
whether an alarm needs some action or can be skipped safely. In the 1980's the concept
of kwnowledge based system emerged with the aim to associate deep knowledge to
diagnosis. An intelligent monitoring system integrates such a knowledge-based system
into a monitoring system.
The first step of intelligent monitoring is temporal abstraction. This means trans-
forming numerical time series into symbolic event sequences. There is a huge literature
in this domain (for surveys see [35,20]), e.g. in the cardiac domain. The second step
is devoted to the reasoning task. Among proposals, model-based diagnosis [28,5] has
the main avantage of using an explicit model that can be used to diagnose the series of
events observed during monitoring as well as giving comprehensible explanations to the
operator. As diseases have an important temporal dimension, we have proposed to rep-
resent them by sets of events linked by temporal constraints on their occurrences. Such
sets of events are called chronicles [11]. They can model (a faulty model, in this case)
the evolution of a disease during time or local typical temporal phenomena, e.g. typical
wave sequences of an electrocardiogram (ECG) that represent cardiac beats character-
istic of some rhythm problem. Their recognition on an input stream is based on efficient
processing of temporal constraint networks [6]. This makes chronicles good candidates
for monitoring.
One of the main challenge of temporal abstraction in intelligent monitoring systems
is to closely couple signal processing tasks and higher level tasks involved in diagnosis.
One source of difficulty is that, generally, recorded data are highly dynamic and sub-
ject to changes. For example, the patient may move letting some sensor transmit very
noisy data. Also, the patient state may evolve quickly due to the effect of some drug or
disease evolution. We propose to introduce a central module called a pilot that analyzes
continuously the monitoring context, i.e. the nature and quality of signals as well as the
hypotheses devised by the diagnosis module. The aim of the pilot is to select the best
signal processing algorithms and the right abstraction level for data abstraction. The pi-
lot makes use of decision rules in order to bring high flexibility for taking into account
new monitoring conditions or new monitoring domains.
The major bottleneck of knowledge based approaches is knowledge acquisition and
maintenance. Machine learning has been advocated for this task. Since monitored dis-
eases have a temporal relational dimension, first order models are good candidates for
knowledge representation. This is why we have used Inductive Logic Programming [25]
for learning chronicles. Devising decision rules for the pilot could also be tedious and
time consuming [26]. We have proposed decision tree learning for inducing decision
rules from the performance of algorithms in a representative set of contexts.
This article summarizes the work done during several years as an active collabora-
tion with experts in biological signal processing and the department of cardiology of the
local hospital and which led to the implementation of an experimental platform called
Calicot. Section 2 we survey some work on intelligent monitoring and self-adaptation.
Section 3 describes the medical applicative context. In Section 3, an overall overview of
the system architecture is given. In Sections 5 and 6, we present the temporal abstrac-
tion and diagnosis methods. In Section 7, a solution to adaptation to context change is
Search WWH ::
Custom Search