Fig 1.
Visualisation of shift between R peak and middle of heartbeat for different classes of arrhythmia.
Fig 2.
Negentropy as a function of phase shift α with 95% confidence interval.
Fig 3.
Concept diagram of proposed method.
Table 1.
Labels used in the MIT-BIH database with number of learning examples assigned to each class.
Fig 4.
Example ECG signal after applying filtration and alignment of QRS complex.
Following type of beats were presented: a) NOR—normal beat, b) PVC—arrhythmias including premature ventricular contraction, c) PAB—paced beat, d) RBB—right bundle branch block beat, e) LBB—left bundle branch block beat, f) APC—atrial premature complexes, g) VFW—ventricular flutter wave, VEB—ventricular escape beat.
Fig 5.
Examples of Daubechies 1-10 wavelets used in experiments 1-3.
Fig 6.
Classification performance when the wavelet transform or ICA are applied separately.
Fig 7.
Classification performance when utilizing different wavelet functions for modification of mixing matrix in ICA.
Table 2.
P-values obtained with Wilcoxon test comparing BAC-score from expierment 1 and 2.
Fig 8.
Classification metrics for various functions for the mixing matrix creation in ICA.
Table 3.
Best MLP networks used for heartbeat classification task.
Table 4.
Comparison of our methods with other results reported in literature.
Table 5.
Best MLP network for heartbeat classification compared to other results from literature.
Fig 9.
Comparison of Recall obtained for our method and CNNs from [5] for selected type of beats.
Number of heartbeats for each type is equal to: PVC—7130, PAB—7028. RBB—7259, LBB—8075, APC—2546, VFW—472, VEB—106. For APC, VFW and VEB beats with lower amount of learning examples available our method provided better results than convolutional neural networks.