Intelligent Adaptive Monitoring for Cardiac Surveillance - Computational Intelligence in Healthcare: Advanced Methodologies - page 329

Ta b l e 1 . Recognition results (on 15525 cardiac beats)

detector sens(%) P+(%) FM(1) (%) switches

af2 92.59 81.47 86.67 -

95.91 76.72 85.25 -

benitez

80.64 86.61 83.52 -

df2

87.29 75.33 80.87 -

gritzali

94.76 82.74 88.35 -

kadambe

95.29 68.36 79.61 -

mobd

79.66 85.70 82.57 -

pan

bestChoice 92.39 91.04 91.71 1486

95.09 82.01 88.06 741

pilot D1

94.74 82.92 88.43 294

pilot D2

93.30 81.53 87.02 1478

pilot D3

91.32 80.74 85.70 443

pilot D1 *

92.33 81.38 86.51 343

pilot D2 *

94.69 81.20 87.42 1642

pilot D3 *

This is consistent with the fat that this decision tree was learnt from examples described
with less attributes, so the decision rules cannot discriminate situations ass prcisely as
pilot D1 or pilot D3 * . Without using kadambe , pilot D3 * switched 1642
times showing that it uses the available algorithms much more.
Compared to kadambe , the advantage of using a pilot is that it uses explicit declar-
ative rules which can be easily updated. This demonstrates the value of using a smart
adaptation of QRS detection algorithms according to both signal, patient and diagnosis
context.

10 Conclusion and Perspectives

We have presented an approach to intelligent monitoring with self-adaptive capabilities
in the cardiac domain. Our proposition associates temporal abstraction, online diagno-
sis by chronicle recognition, self-adaptation to the monitoring context and automatic
knowledge acquisition to learn chronicles and adaptation decision rules. A prototype
named Calicot has been implemented.
Efficiency has been a constant concern during the conception and implementation of
Calicot, as it was intended to run online. Thus, a temporal abstraction method taking
advantage of the domain and data specificities has been proposed. Though they repre-
sent complex event, chronicles can be efficiently recognized on multiple data streams,
with one or two orders of magnitude less than real time in our case. To enhance the
performance we have also proposed an architecture for self-adaptation, featuring a pilot
which can reconfigure the processing chain or tune the module parameters when the
monitoring context changes. Finally, symbolic machine learning is used, offline, to get
discriminating patterns, on the one hand, and adaptation decision rules, on the other