Convolutional Neural Network for Individual Identification Using Phase Space Reconstruction of Electrocardiogram

Figures

• 

  Figure 1. (a) A single-cycle ECG with P, QRS complex, and T waves. (b) The transformed phase portrait using a time delay of 20 ms.

• 

  Figure 2. Phase portraits transformed from single-cycle ECG, using various time delays over various-density grid partitions. QRS complex, T, and P waves are transformed to a major outer loop, a secondary inner loop, and a minor loop, respectively, in each portrait. The distinctness of the phase portrait decreases as its grid partition is reduced from 64 × 64 to 16 × 16 partitions. For a visual comparison, all phase portraits are displayed with the same picture size.

• 

  Figure 3. Accuracy of individual identification over the first 30 epochs on the underlying phase portraits (a) embedded using five time delays (τ = 2, 8, 16, 24, and 36 ms) and (b) generated over five grid partitions (16 × 16, 32 × 32, 64 × 64, 128 × 128, and 256 × 256).

• 

  Figure 4. Map of average accuracy of individual identification on the underlying phase portraits using the combinations of five grid partitions (16 × 16, 32 × 32, 64 × 64, 128 × 128, and 256 × 256) and ten time delays (2, 4, 8, 12, 16, 20, 24, 28, 32, and 36 ms) during four training stages (epochs 16~30, 31~45, 46~60, and 61~75).

• 

  Figure 5. Plot of average accuracies of individual identification on the underlying phase portraits using the combinations of five grid partitions (16 × 16, 32 × 32, 64 × 64, 128 × 128, and 256 × 256) and ten time delays (2, 4, 8, 12, 16, 20, 24, 28, 32, and 36 ms) during four training stages: (a) epochs 16~30, (b) epochs 31~45, (c) epochs 46~60, and (d) epochs 61~75.

• 

  Figure 6. (a) Waveforms of 100 consecutive single-beat ECG in a subject. (b) The transformed phase space trajectories using a time delay of 20 ms.

• 

  Figure 7. The identification accuracy using the scaled-down network and original network which were designed for portrait images with 32 × 32 and 256 × 256 pixels, respectively, but the portrait images were derived on the basis of phase-space reconstruction over 32 × 32 grid partition.