Table 4.
Deep learning models for atrial fibrillation detection.
| Authors, Year | Number of Subjects | Leads | Classes | Database | Method | Performance (%) | ||
|---|---|---|---|---|---|---|---|---|
| Spec. | Sen. | Acc. | ||||||
| Acharya et al., 2017 [30] | 21,709 2 s ECG segments 8683 5 s ECG segments |
Lead II | SR, AF, AFL and VF | MIT-BIH DB, MIT-BIH AFDB, CU VTDB |
11-layer CNN | 93.13 81.44 |
98.09 99.13 |
92.50 94.90 |
| Xia et al., 2018 [40] | 162,536 5 s ECG segments | 2 Lead | AF and non-AF | MIT-BIH AFDB | STFT (RGB) + CNN STFT (grayscale) + CNN SWT + CNN |
98.24 97.17 97.87 |
98.34 98.60 98.79 |
98.29 97.74 98.63 |
| Faust et al., 2018 [41] | CV: 20 subjects BV: 3 subjects |
- | SR and AF | MIT-BIH AFDB | HRV + bidirectional LSTM | 98.67 99.61 |
98.32 99.87 |
98.51 99.77 |
| Fan et al., 2018 [45] | 5154 SR recordings 7713 AF recordings |
Single Lead | SR and AF, AF and O |
PhysioNet/CinC 2017 | MS-CNN | 98.77 98.84 |
93.77 80.26 |
98.13 97.19 |
| Andersen et al., 2019 [19] | 23 long-term recordings 48 short-term recordings 18 long-term recordings |
Single Lead | SR and AF | MIT-BIH AFDB, MIT-BIH DB, MIT-BIH SRDB |
CNN + LSTM | 96.95 86.04 95.01 |
98.98 98.96 - |
97.80 87.40 - |
| Fujita et al., 2019 [56] | 25,287 2 s ECG segments | Single Lead | SR, AF, AFL and VF | MIT-BIH DB, MIT-BIH AFDB, MIT-BIH VFDB |
8-layer CNN | 96.07 | 99.43 | 98.61 |
| Attia et al., 2019 [47] | 649,931 10 s ECG recordings | 12 Lead | SR and AF (includes AFL) | Mayo Clinic ECG Laboratory | CNN | 83.4 | 82.3 | 83.3 |
| Baalman et al., 2020 [48] | 1499 10 s ECG recordings | Lead II, 8 Lead |
SR and AF | AFACT | R-centered SC-ECG + RNN R-to-R-wave SC-ECG + RNN |
- | - | 94.00 96.00 |
| Cai et al., 2020 [43] | 16,557 10 s ECG recordings | 12 Lead | SR and AF AF and non-AF SR, AF and O |
Chinese PLA General Hospital Wearable 12-Lead, The China Physiological Signal 2018 |
DDNN | 99.19 97.04 95.85 |
99.44 98.63 98.38 |
99.35 98.21 97.74 |
| Lai et al., 2020 [46] | 510,472 10 s ECG recordings | Multi Lead | AF and non-AF | Hexin Patch Lead II, MIT-BIH DB |
8-layer CNN | 93.4 | 93.1 | 93.1 |
| Jin et al., 2020 [59] | 150,060 5 s ECG recordings | - | AF and non-AF | MIT-BIH AFDB | Multi-domain feature + TAC-LSTM | 98.76 | 98.14 | 98.51 |
| Wang et al., 2020 [60] | 22,174 ECG segments 1265 ECG segments |
Single Lead | SR, AF and AFL | MIT-BIH AFDB, MIT-BIH DB |
CNN + MLP CNN + ENN CNN + IENN |
99.3 99.6 |
97.1 99.3 |
98.3 99.4 |
| Nurmaini et al., 2020 [61] | 6114 samples (9 s) | Single Lead | SR and AF SR, AF and non-AF |
PhysioNet AFDB, MIT-BIH AFDB, MIT-BIH Malignant Ventricular Entropy, An Indonesian Hospital |
13-layer one-dimensional CNN | 99.91 99.17 |
99.91 98.90 |
99.98 99.17 |
| Mousavi et al., 2020 [42] | 167,422 5 s ECG recordings 8528 ECG recordings |
Single Lead | AF and non-AF SR and AF |
MIT-BIH AFDB, PhysioNet/CinC 2017 |
BiRNN (HAN-ECG) | 98.54 | 99.08 | 98.81 |
| Chen et al., 2021 [54] | - | 2 Lead 12 Lead |
SR and AF | MIT-BIH DB, AHA DB, QT DB, CSE DB |
Multiple feature extraction + CNN | - | - | 98.92 |
| Petmezas et al., 2021 [52] | 970,009 beats | 2 Lead | SR, AF, AFL and J | MIT-BIH AFDB | CNN + LSTM + FL | 99.29 | 97.87 | - |
| Jo et al., 2021 [49] | - | 12 lead, 6 Lead, Single Lead |
AF and non-AF | Sejong ECG DB, PTB-XL ECG DB, Charman et al. ECG DB, PhysioNet DB | CNN | 99.5 | 99.9 | 99.6 |
| Zhang et al., 2021 [55] | 80,000 ECG segments 83,464 ECG segments 19,220 ECG segments |
Lead I | AF and non-AF | Wearable Lead I-II, MIT-BIH AFDB, PhysioNet/CinC 2017 |
LSTM + CNN | 95.19 94.49 96.66 |
97.73 96.46 92.09 |
95.44 95.28 96.23 |
Acc, accuracy; AF, atrial fibrillation; AFL, atrial flutter; BiRNN, bidirectional RNN; BV, blindfold validation; CNN, convolutional neural network; CV, cross-validation; DDNN, deep densely connected neural network; ENN, Elman neural network; FL, focal loss; FRM-CNN, CNN-based AF screening framework; HAN, hierarchical attention network; HRV, heart rate variability; IENN, improved Elman neural network; J, junctional rhythm; LSTM, long short-term memory; MLP, multilayer perceptron; MS-CNN, multi-scaled fusion of deep CNN; N, noisy; O, others; RNN, recurrent neural network; SC, single-cycle; Sen, Sensitivity; Spec, Specificity; SR, sinus rhythm; STFT, short-time Fourier transform; SWT, stationary wavelet transform; TAC-LSTM, twin-attentional convolutional LSTM; VF, ventricular fibrillation.