Table 4.
Literature for heart sound classification using deep learning.
| Year | Author | Segmentation method | Dataset | Performance | ||||
|---|---|---|---|---|---|---|---|---|
| 2019 | Wu et al. [56] | CNN | PhysioNet (2575 normal heart sounds and 665 abnormal heart sounds) | Hold out testing | ||||
| Sensitivity | Specificity | Accuracy | ||||||
| 86.46% | 85.63% | 86.0% | ||||||
| Ten-fold cross validation | ||||||||
| Sensitivity | Specificity | Accuracy | ||||||
| 91.73% | 87.91% | 89.81% | ||||||
| 2019 | Abduh et al. [57] | DNN | PhysioNet | Sensitivity | Specificity | Accuracy | ||
| 89.30% | 97% | 95.50% | ||||||
| 2018 | Gharehbaghi and Lindén [58] | DTGNN | 130 recordings of the heart sound signal | Sensitivity | Specificity | CR | ||
| 83.9% | 86% | 85.5% | ||||||
| 2018 | Chen et al. [59] | DNN | PASCAL | Sensitivity | Specificity | Accuracy | Precision | |
| 98% | 88.5% | 93% | 89.1% | |||||
| 2018 | Yaseen et al. [60] | DNN | 5 categories of heart sound signal, 200 per class (N, AS, MR, MS, MVP) | Sensitivity | Specificity | |||
| 94.5% | 98.2% | |||||||
| 2018 | Han et al. [61] | CNN | 2575 normal recordings and 665 abnormal recordings | MAcc | Sensitivity | Specificity | ||
| 91.50% | 98.33% | 84.67% | ||||||
| 2018 | Ren et al. [62] | CNN | PhysioNet | 19.8% higher than the baseline accuracy obtained using traditional audio processing functions and support vector machines. | ||||
| 2018 | Morales et al. [63] | CNN | PhysioNet | Accuracy | Sensitivity | Specificity | ||
| 97% | 93.20% | 95.12% | ||||||
| 2018 | Baris et al. [64] | CNN | UoC-murmur database (innocent murmur versus pathological Murmur) and PhysioNet-2016 database (normal versus pathological) | MAcc | Specificity | Sensitivity | ||
| 81.5% | 78.5% | 84.5% | ||||||
| 2018 | Messner et al. [65] | DNN | PhysioNet | F1 ≈ 96% | ||||
| 2017 | Ghaemmaghami et al. [66] | DNN | 128 recordings from male and female subjects with healthy hearts | Accuracy | Sensitivity | Specificity | ||
| 95.8% | 83.2% | 99.2% | ||||||
| 2017 | Sujadevi et al. [67] | RNN & LSTM&GRU | Dataset A from PhysioNet | Accuracy | Precision | |||
| RNN 4 layer | 53.8% | 55.8% | ||||||
| LSTM 4 layer | 76.9% | 83.3% | ||||||
| GRU 4 layer | 75.3% | 78.2% | ||||||
| Dataset B from PhysioNet | Accuracy | Precision | ||||||
| RNN 4 layer | 65.2% | 68.1% | ||||||
| LSTM 4 layer | 74.7% | 94.5% | ||||||
| GRU 4 layer | 74.4% | 69.7% | ||||||
| 2017 | Chen et al. [68] | DNN | 311 S1 and 313 S2 from 16 people (11 males and 5 females) | Accuracy: 91.12% | ||||
| 2017 | Yang and Hsieh [69] | RNN | PhysioNet | MAcc: 84% | ||||
| 2017 | Zhang and Han [70] | CNN | Dataset A from PASCAL | Normalized precision: 0.77 | ||||
| Dataset B from PASCAL | Normalized precision: 0.71 | |||||||
| 2017 | Faturrahman et al. [71] | DBN | MITHSDB [72] | Accuracy: 84.89% | ||||
| AADHSDB [73] | Accuracy: 86.15% | |||||||
| 2017 | Maknickas and Maknickas [74] | CNN | PhysioNet | Train accuracy: 99.7% | ||||
| Validation accuracy: 95.2% | ||||||||
| 2016 | Thomae et al. [75] | DNN | PhysioNet | Sensitivity | Specificity | Score | ||
| 96% | 83% | 0.89 | ||||||
| 2016 | Tschannen and Dominik [76] | CNN | PhysioNet | Sensitivity | Specificity | Score | ||
| 84.8% | 77.6% | 0.812 | ||||||
| 2016 | Potes et al. [77] | AdaBoost & CNN | PhysioNet | Sensitivity | Specificity | MAcc | ||
| 94.24% | 77.81% | 86.02% | ||||||