Intelligent Adaptive Monitoring for Cardiac Surveillance - Computational Intelligence in Healthcare: Advanced Methodologies - page 328

Fig. 7. The graphical interface of Calicot showing the recognition of bigeminy and trigeminy
episodes.
each ECG chunck was also kept. These results were used to define the gold standard
( bestChoice ).
Table 1 synthesizes the results of arrhythmia recognition without the pilot and every
QRS detector algorithms, bestChoice (used as gold standard) and the different pilot-
ing rule sets. According to the F-measure, pilot D2 outperforms all the other meth-
ods. The best non piloted monitoring performance ( FM =88.35%) is obtained when
using algorithm kadambe for temporal abstraction, followed by af2 ( FM =86.67%),
and benitez ( FM =85.25%). These three algorithms outperform the others with an F-
measure greater by 1.73%. The best piloted monitoring performance is obtained by the
pilot D2 rule set followed by pilot D1 and pilot D3 . pilot D1 performs
better than non piloted af2 but worst than non piloted kadambe . The upper bound
that can be reached is given by bestChoice with FM =91.71%. This shows that the
piloting strategy could be notably improved with more accurate rules (at most by 3.36%
on the F-measure).
kadambe associates wavelet analysis with heuristics for self-adaptating to the sig-
nal, thus, it can be considered as being “piloted”. To asses the pilot more fairly, new
piloting rules, pilot D1 * , pilot D2 * and pilot D3 * , were learned excluding
kadambe . The best performance of non piloted algorithms was obtained for af2 fol-
lowed by benitez . The piloted rule sets exibited the best performance: pilot D3 *
with FM = 87.42% outperformed non piloted af2 with FM = 86.67% improving FM
by 0.75% which is considered a good improvement in the QRS detection field. The
number of switches (column switches ) shows that the best possible performance
makes 1486 switches. pilot D2 reached good scores with far less switches (294).