C. Feature Extraction
For subsequent beat classification, four heart beat features
as defined by Krasteva & Jekova [1] were used. The features
with respect to the templates were: (i) difference in absolute
area (ArDiff, using normalized waveform area) and (ii) maximal
cross-correlation coefficient (MaxCorr). Again, 400 ms
windows centered on every detected R-peak were used for
this computation. Furthermore, the width of detected QRS
complexes QRSwidth was computed using the Pan-Tompkins
integrator output [8]. The last computed feature was the RR-
interval R-R. The computed features are shown as rounded
shapes in Fig. 2.
D. Beat Classification
Beats were classified using two different characteristics:
the WAVEFORM characteristic, which distinguished normal
and abnormal beats. The abnormal beats class
had the subclasses fpremature ventricular contraction
(PVC), PVC/aberrant, bundle branch block, escape beat
(generic), atrial premature contraction (APC), aberrantg,
the PACE/RHYTHM characteristic, which distinguished
normal and abnormal pace. The abnormal pace class
had the subclasses ffusion of two beats, AV-block,
tachycardia, bradycardiag.
Discrimination of the different subclasses was performed by
a decision tree (Fig. 2) as proposed in [1]. The different beat
classes are shown in rectangular boxes in Fig. 2.