Table 4.
Mean absolute errors (± the standard deviation) in ms between predicted and ground truth activation times at different subsampling ratios of the septal ground truth.
| Subsampling ratio (%) | GCN-active | GCN-random | DEP-random | NN-random |
|---|---|---|---|---|
| 1 | 8.5 ± 7.9 | 8.7 ± 8.1 | 9.5 ± 8.9 | 14.9 ± 11.4 |
| 2 | 8.3 ± 8.0 | 8.4 ± 8.2 | 9.1 ± 8.6 | 13.7 ± 11.5 |
| 3 | 8.0 ± 7.9 | 8.3 ± 8.3 | 9.3 ± 8.4 | 9.7 ± 10.1 |
| 4 | 7.8 ± 7.9 | 8.0 ± 8.0 | 9.0 ± 8.3 | 9.3 ± 9.5 |
| 5 | 7.5 ± 7.7 | 7.9 ± 8.0 | 8.7 ± 8.1 | 9.3 ± 9.6 |
| 6 | 7.4 ± 7.6 | 7.7 ± 7.7 | 8.6 ± 8.1 | 8.7 ± 9.3 |
| 7 | 7.2 ± 7.5 | 7.6 ± 7.7 | 8.5 ± 8.1 | 8.1 ± 9.4 |
| 8 | 7.1 ± 7.4 | 7.7 ± 7.7 | 8.4 ± 8.0 | 7.4 ± 9.4 |
| 9 | 7.0 ± 7.4 | 7.5 ± 7.6 | 8.3 ± 7.9 | 7.2 ± 9.4 |
| 10 | 7.0 ± 7.4 | 7.4 ± 7.5 | 8.2 ± 7.8 | 7.1 ± 9.4 |
Active sampling based on an ensemble of four GCNs (GCN-active) is compared against three methods with random sampling: the graph convolutional network (GCN-random), the personalized computational model (DEP-random), and the nearest neighbor projection (NN-random).