Table 5 Summary of intelligent DL-based SAA
Refs., Year | Database | Signal for FE, Segment length | Method description | ICF | FS | FV | Limitation | Key Findings |
|---|---|---|---|---|---|---|---|---|
[7], 2018 | CUDB, VFDB, MITDB | Stand-alone ECG, 2 s | The 11-layers CNN for classification of SH/NSH 2 s-ECG segment. | NA | NA | 10-folds CV | Only one CNN structure. | Simple full CNN. |
Validation of full CNN with 10-fold CV. | Time-consuming for validation of the full CNN. | Less complexity due to no FS and FV. | ||||||
Relative short of segment length. | ||||||||
[8], 2018 | CUDB, VFDB | ECG segment, NSH signal, SH signal, using MVMD, 8 s | NSH, SH signals generated by MVMD. | NA | NA | 5-folds CV | Time-consuming for selection of CNNE. | Improvement of LF quality due to multiple channels. |
Use of ECG segment, NSH and SH signals as input channels of CNN | Improvement of final performance due to secondary training of ML classifier. | |||||||
Grid search with nested 5-fold CV to select best structure and parameters of CNNE using ML classifiers. | No need of FE and FS. | |||||||
Validation of feature vector extracted by CNNE with different ML classifiers. | ||||||||
[46], 2020 | CUDB, VFDB | Modes using FFREWT, 8 s | FFREWT for ECG segment decomposition into 6 modes. | NA | NA | 10-folds CV | pre-selected CNN, structure. | High performance. Effective FFREWT for generation of different modes containing VF, VT, normal components of ECG. |
First 5 modes used as the input of CNN for detection of SH and NSH ECG segment, VF and VT rhythms | ||||||||
[47], 2019 | CUDB, VFDB, AHADB | Stand-alone ECG, 4 s | 1D parallel CNN, LSTM and ANN for classification of 4 s-ECG segment. | NA | NA | NA | pre-selected CNN, LSTM, ANN structure. | High performance. |
No validation. | Multiple DLs for deep feature extraction. | |||||||
Relative short of segment length. | ||||||||
[48], 2020 | EMS | Stand-alone patient’s ECG, 4 s | Fully CNN architecture and ResNet CNN model | NA | NA | 10-folds CV | pre-selected CNN structure. | High performance for 4 s-ECG segment. |
Classification of SH/NSH for different ECG segment length. | Relative short of segment length. | |||||||
Validation with 10-fold CV. | ||||||||
[49], 2020 | CUDB, VFDB, MITDB, OHCA1,OHCA2 | Stand-alone ECG, 5 s | Random search based method for hyper-parameters of optimal deep CNN models using number of sequential CNN blocks, number of filters, kernel sizes. | NA | NA | NA | No validation. | Productive method for selection of a deep CNN structure with optimal hyper-parameters. |
Median values to rank the optimal deep CNN models trained with various learning rates and ECG segment lengths to select the best deep CNN model. | Time-consuming for hyperparameter optimization | Model ability related to SH/NSH classification is depended largely on hyper-parameters. | ||||||
[50], 2021 | CUDB, VFDB | Stand-alone ECG, 5 s | DNN using a feature set extracted from ECG segments pre-processed by DWT, EMD, VMD. | 24 | NA | NA | No validation. | Relative short of segment length |
Comparison to various ML classifiers | pre-selected DNN | Effective decomposition techniques for data processing. | ||||||
[51], 2021 | CUDB, VFDB, MITDB, AHADB | Time-frequency maps, 3 s,5 s,8 s,10 s | Conversion of ECG segments into time-frequency maps using CWT. | NA | NA | 10-folds CV | Time-consuming for selection of an optimal 2D CNN structure | Relative short of segment length |
Investigation of eight 2D CNN structures. | Productive transformation method using CWT. | |||||||
Consideration of different ECG segment lengths. |