Table 2.
Comparative performance analysis of AFIB detection methods on the PTB-XL-Improved test set and the original PTB-XL test set. All methods were trained and evaluated as 12-lead ECG classifiers using the same patient-level train–test split. For the encoder–decoder, repeated training across 13 random seeds yielded empirical 95% confidence intervals within ±0.1% for the reported metrics.
| Method | Dataset | Set | AUC | Sensitivity | Specificity | F1-Score | PPV |
|---|---|---|---|---|---|---|---|
| Encoder–decoder (ours) | PTB-XL-Improved | Test | 99.15* | 97.81 | 95.63* | 96.75* | 95.72* |
| CNN-BiLSTM [40,41] | PTB-XL-Improved | Test | 97.15 | 95.62 | 90.62 | 93.29 | 91.06 |
| ResNet [42,51] | PTB-XL-Improved | Test | 97.27 | 98.75* | 87.50 | 93.49 | 88.76 |
| CNN [31,35–37] | PTB-XL-Improved | Test | 79.18 | 82.19 | 60.31 | 74.08 | 67.44 |
|
| |||||||
| Encoder–decoder (ours) | Original PTB-XL | Test | 99.78* | 99.01* | 98.02 | 98.52* | 98.04* |
| CNN-BiLSTM [40,41] | Original PTB-XL | Test | 98.84 | 97.69 | 92.85 | 95.38 | 93.17 |
| ResNet [42,51] | Original PTB-XL | Test | 99.69 | 97.03 | 98.84* | 95.15 | 93.33 |
| CNN [31,35–37] | Original PTB-XL | Test | 94.29 | 86.14 | 90.10 | 87.88 | 89.69 |