|
| Studies | Dataset | Technique | Outcomes | Limits |
|
| Yeh et al. [1] | Private and PTB DB | ResNet, AlexNet, and SqueezeNet | Accuracy and kappa statistics of ResNet, AlexNet, and SqueezeNet in ECG waveform classification (0.97, 0.96), (0.96, 0.95), and (0.75, 0.67). | (i) No data augmentation,
(ii) less accurate, and
(iii) worked on waveform classification |
|
| Wasimuddin et al. [2] | ECG-ID | CAD and machine learning approach | CAD and machine learning approach working on 2D image based on classification and worked on the R peak of the ECG and showed an accuracy of 98.5%. | (i) Handcrafted features,
(ii) small dataset, and
(iii) accuracy is remarkable but slow because of handcrafted features |
|
| Hsu et al. [7] | MIT-BIH | AlexNet and ResNet 18 | ECG into the fingerprint by using the transfer learning methods and proved the predicted accuracy of 94.4%. | (i) No data augmentation and
(ii) handcrafted features |
|
| Elgendi and Menon [8] | Private | Machine learning approach | Supervised ML algorithms confirmed that ECG is an optimal wearable biosignal for assessing driving stress, with an overall accuracy of 75.02%. | (i) Low accuracy,
(ii) augmentation not performed, and
(iii) handcrafted features |
|
| Gaddam and Sreehari [12] | MIT-BIH | AlexNet | Transferred deep learning convolution neural net with 1D and 2D structure with 95.6% accuracy. | (i) Augmentation not performed and
(ii) low accuracy |
|
| Simjanoska et al. [14] | 4 private datasets used | Machine learning | The proposed method achieved 8.64 mmHg of the mean absolute error in the case of SBP. | (i) Handcrafted and
(ii) low accuracy |
|
| Acharya et al. [15] | PTB DB | CNN layers | CNN for automated detection of myocardial interaction using ECG signals, and inferred the data with noise (93.5%) and without noise (95.22%). | (i) Low accuracy and
(ii) less number of classes |
|
| Tomer Golany [21] | Private | GAN-based generative models such as GAN, DCNN, SIMCGAN, and SIMDCGAN. | Simulator-based network for ECG to improve deep ECG classification was used and compared all GAN-based models to find the accurate result of ECG and got SIMDCGAN as a refined and result-oriented model. | (i) Low accuracy,
(ii) augmentation not performed, and
(iii) handcrafted features |
|
| Sehirli et al. [28] | PTB-XL | RNN (LSTM and GRU) | Compared the performance of the RNN with the long short-term memory (LSTM) and gated recurrent unit (GRU) and then observed that the LSTM technique is the latent method for the sequential data and time series with the accuracy of 97.7%. | (i) Less accurate,
(ii) small dataset, and
(iii) augmentation not performed |
|
| Strodthoff et al. [33] | PTB-XL | ResNet and inception | Deep learning of ECG analysis by using datasets showed an 89.8% result. | (i) No augmentation and
(ii) low accuracy |
|
| Rahman et al. [45] | MIT-BIH | CAA-TL model (deep learning) | Different transfer learning approaches analyzed with data augmentation achieved 98.38% accuracy. | (i) No data fusion,
(ii) low accuracy, and
(iii) no hardware implementation |
|
