Table 1.
Deep learning-based methods for heart sounds classification.
| S. No | Reference | Method | Input Features | Segment | Optimizer | Categories | Performance on Test Dataset MAcc, Se, Sp, Acc |
|---|---|---|---|---|---|---|---|
| CNN-Based Methods | |||||||
| 1 | Maknickas et al., 2017 [25] |
2D-CNN | MFSC | No | RMSprop | N, A | 84.15, 80.63, 87.66, * |
| 2 | Tarik Alafif et al., 2020 [26] | 2D-CNN + transfer learning | MFCC | NO | SGD | N, A | *, *, *, 89.5% |
| 3 | Deng et al., 2020 [24] | CNN + RNN | Improved MFCC | No | Adam | N, A | 0.9834, 0.9866, 0.9801, * |
| 4 | Abduh et al., 2019 [27] | 2D-DNN | MFSC | No | * | N, A | 93.15, 89.30, 97.00, 95.50 |
| 5 | Chen et al., 2018 [28] | 2D-CNN | Wavelet transform + Hilbert–Huang features | No | * | N, M, EXT | 93.25, 98, 88.5, 93 |
| 6 | Rubin et al., 2016 [29] | 2D-CNN | MFCC | Yes | Adam | N, A | 83.99, 72.78, 95.21, * |
| 7 | Nilanon et al., 2016 [30] | 2D-CNN | Spectrograms | No | SGD | N, A | 81.11, 76.96, 85.27, * |
| 8 | Dominguez et al., 2018 [31] | 2D-CNN | Spectrograms | No | * | N, A | 94.16, 93.20, 95.12, 97.05 |
| 9 | Bozkurt et al., 2018 [32] | 2D-CNN | MFCC + MFSC | Yes | * | N, A | 81.5, 84.5, 78.5, 81.5 |
| 10 | Chen et al., 2019 [33] | 2D-CNN | MFSC | No | Adam | N, A | 94.81, 92.73, 96.90, * |
| 11 | Cheng et al., 2019 [34] | 2D-CNN | Spectrograms | No | * | N, A | 89.50, 91.00, 88.00, * |
| 12 | Fatih et al., 2019 [35] | 2D-CNN | Spectrograms | No | * | N, M, EXT | 0.80 (Accuracy on dataset A) 0.79 (Accuracy on dataset B) |
| 13 | Ryu et al., 2016 [36] | 1D-CNN | 1D time-series signals | No | SGD | N, A | 78.69, 66.63, 87.75, * |
| 14 | Xu et al., 2018 [37] | 1D-CNN | 1D time-series signals | No | SGD | N, A | 90.69, 86.21, 95.16, 93.28 |
| 15 | Xiao et al., 2020 [38] | 1D-CNN | 1D time-series signals | No | SGD | N, A | 90.51, 85.29, 95.73, 93.56 |
| 16 | Humayun et al., 2020 [39] |
tConv-CNN (1D-CNN) | 1D time-series signals | Yes | Adam | N, A | 81.49, 86.95, 76.02, * |
| 17 | Humayun et al., 2018 [40] |
1D-CNN | 1D time-series signals | Yes | SGD | N, A | 87.10, 90.91, 83.29, * |
| 18 | Li et al., 2019 [41] | 1D-CNN | Spectrograms | No | * | N, A | *, *, *, 96.48 |
| 19 | Li et al., 2020 [42] | 1D-CNN | 497 features from time, amplitude, high-order statistics, cepstrum, frequency cyclostationary and entropy domains | Yes | Adam | N, A | *, 0.87, 0.721, 0.868 |
| 20 | Xiao et al., 2020 [43] | 1D-CNN | 1D time-series signals | No | N, A | *, 0.86, 0.95, 0.93 | |
| 21 | Shu Lih Oh et al., 2020 [44] |
1D-CNN WaveNet | 1D time-series signals | NO | Adam | N, AS, MS, MR, MVP | 0.953, 0.925, 0.981, 0.97 |
| 22 | Baghel et al., 2020 [45] | 1D-CNN | 1D time-series signals | No | SGD | N, AS, MS, MR, MVP | *, *, *, 0.9860 |
| RNN-Based Methods | |||||||
| 23 | Latif et al., 2018 [46] | RNN (LSTM, BLSTM, GRU, BiGRU) | MFCC | Yes | * | N, A | 98.33, 99.95, 96.71, 97.06 (LSTM) 98.61, 98.86, 98.36, 97.63 (BLSTM) 97.31, 96.69, 97.93, 95.42 (GRU) 97.87, 98.46, 97.28, 97.21 (BiGRU) |
| 24 | Khan et al., 2020 [47] | LSTM | MFCC | No | * | N, A | *, *, *, 91.39 |
| 25 | Yang et al., 2016 [48] | RNN | 1D time-series signals | No | * | N, A | 80.18, 77.49, 82.87, * |
| 26 | Raza et al., 2018 [49] | LSTM | 1D time-series signals | No | Adam | N, M, EXT | *, *, *, 80.80 |
| 27 | Westhuizen et al., 2017 [50] | Bayesian LSTM LSTM | 1D time-series signals | No | N, A | 0.798, 0.707, 0.889, 0.798 0.7775, 0.675, 0.880, 0.778 |
|
| Hybrid Methods | |||||||
| 28 | Wu et al., 2019 [51] | Ensemble CNN | pectrograms + MFSC + MFCC | No | * | N, A | 89.81, 91.73, 87.91, * |
| 29 | Noman et al., 2019 [52] | Ensemble CNN | (1D time-series signals + MFCC) | Yes | Scikit | N, A | 88.15, 89.94, 86.35, 89.22 |
| 30 | Tschannen et al., 2016 [53] | 2D-CNN+SVM | Deep features | Yes | * | N, A | 81.22, 84.82, 77.62, * |
| 31 | Potes et al., 2016 [23] | AdaBoost + 1D-CNN | Time and frequency features, MFCC | Yes | * | N, A | 86.02, 94.24, 77.81, * |
| 32 | Gharehbaghi et al., 2019 [54] | STGNN + MTGNN | Time-series signal | No | * | N, A | *, 82.8, *, 84.2 |
| 33 | Deperlioglu et al., 2020 [55] | AEN | 1D time-series signals | No | * | N, M, EXT | 0.9603 (Accuracy for normal), 0.9191 (Accuracy for extrasystole), 0.9011 (Accuracy for murmur) |
* Abbreviations—N: normal heart sounds, M: murmur heart sounds, EXT: extrasystole heart sounds, AS: aortic stenosis, MS: mitral stenosis, MR: mitral regurgitation, MVP: mitral valve prolapse, MS: mitral stenosis, Acc: accuracy, MAcc: mean of specificity, Sp: specificity, Se: sensitivity.