Information Technology Reference
In-Depth Information
Analysis) to determine the most informative context attributes [26] from performance
data obtained by executing a set of signal processing algorithms in many situations with
different kinds and level of noise and different diseases. Then, decision production rules
conditioned by the selected attributes were built manually by an expert. The second,
more systematic, method makes use of decision tree learning [27]. From a similar set of
performance data, decision trees are learnt. The nodes of such a tree represent a partition
of the values of some attribute and the leaves represent a class, here an algorithm. Every
path from the root node to some leave can be translated into a decision rule having as
conditions the tests in the path nodes and as conclusion the algorithm given in the leave.
It is worth-noting that the performance of learnt rules is very close to the performance
of expert rules.
9
Evaluation
The Calicot prototype is implemented in Java
1
. Its graphical interface displays moni-
tored signals annotated with complex events related to recognized chronicles (see Fig.
7 ). Many experiments were conducted in order to assess its monitoring quality and
evaluate its performance. Calicot has been evaluated on real clinical data recorded in
ICU but, until now, has not been used in clinical routine. This section gives some re-
sults concerning the performance of the prototype on QRS detection, with and without
piloting.
The context analyser that was implemented is based on a wavelet decomposition-
recomposition in three subbands in which the root mean square is computed to obtain
the triplet
ls, ms, hs
(low, medium, high subbands). This triplet together with the an-
notated context attributes
r, n, SN R
(rhythm, noise type, Signal-to-Noise Ratio) forms
the context descriptor used by the pilot to decide which algorithm to use. Piloting rules
were extracted by decision tree learning from the performance results of 7 QRS de-
tectors taken from the litterature [27]. Three decision trees were induced: D1 (using
all attributes:
r
×
n
×
SNR
×
ls
×
ms
×
hs
), D2 (using a subset of the attributes:
r
hs
).
Ten ECGs, lasting around 30 minutes each (containing about 18.000 QRSs in to-
tal) and including ten various ventricular and supra-ventricular arrhythmic contexts,
were extracted from the MIT-BIH Arrhythmia database [23]. Real clinical noise,
from 5
dB
to -15
dB
, was introduced randomly in each ECG with probabilities
P
×
ls
×
ms
×
hs
) and D3 (using the subbands attributes only:
ls
×
ms
×
(
no noise
)=
P
(
bw
)=
P
(
ma
)=
P
(
em
)=1
/
4
and
P
(5
dB
)=1
/
2
,P
(
−
5
dB
)=
1
to reproduce difficult clinical ECG situations and to assess
the system performance in specific contexts as well as when the context changes. The
performance was evaluated from the standard measures:
TP
(True Positive - correct re-
sult),
FN
(False Negative - missed result) and
FP
(False Positive - false result) were
used to compute the sensitivity
Se
/
3
,P
(
−
15
dB
)=1
/
6
TP
TP
+
FN
TP
TP
+
FP
=
, the positive predictivity
PP
=
)=
(1+
β
2
)
∗
PP
∗
Se
β
2
∗
PP
+
Se
and the F-measure
FM
(
β
,where
β
=1
(same weight for
Se
and
PP
).
To estimate the upper bound performance reachable by Calicot with the pilot, the
best detector performance (i.e. achieved by the detector with maximal F-measure) for
1
http://www.irisa.fr/dream/Calicot/
Search WWH ::
Custom Search