|
| Author | Approach | Database | ECG beats | Advantages | Disadvantages |
|
| Balasundaram [3] | Wavelet analysis | MITDB | Rhythmic ventricular tachycardia WA (VT)
Organized ventricular fibrillation (OVF)
Disorganized ventricular fibrillation (DVF) | This method performs well in the overlap zone between VT and VF. | The limitation is that it cannot be used as a risk-stratifier, because this method cannot determine the probability of future VF episodes. |
|
| Sayantan [4] | Gaussian-Bernoulli deep belief network and active learning | SVDB MITDB | AAMI | By using the expert interaction, this method is robust and it can overcome the variance in data distribution in interpatient scenarios. | This method handles intraclass variations poorly. |
|
| Javadi [5] | ME and NCL | MITDB | Normal (N)
Premature ventricular contraction (PVC)
Others | Combined with NCL and ME, it can enable the training algorithm of ME to establish a balance in bias-variance-covariance trade-offs, and it improves the accuracy and generalization of the model. | This work does not provide further insights of the classification boundaries. |
|
| Rajpurkar [6] | CNN | MITDB | Atrial fibrillation (AFIB)
Atrial flutter (AFL)
Complete heart block (CHB)
Ectopic atrial rhythm (EAR)
(14 rhythm types), etc. | This method uses a very deep CNN; the model can achieve a high accuracy rate under a big dataset.
It can be used for a single-lead wearable monitor. | If two ECG signals are similar, the model sometimes makes mistakes, such as Wenckebach and AVB_Type2, Supraventricular Tachycardia (SVT), and atrial flutter (AFL). |
|
| Li [7] | CNN | MITDB | AAMI | This method uses the SMOTE algorithm to balance the classes in dataset. | It only extracted the time-domain features. |
|
| Alif [8] | 2D CNN | MITDB | AAMI EC57 | This method utilizes CNN to extract features automatically. | It only extracted the time-domain features. |
|
| Marinho [9] | Feature extraction: Fourier, goertzel, higher order statistics (HOS), and SCM.
Classifier: support vector machine, multilayer perceptron, bayesian, and optimum-path forest | MITDB | ANSI/AAMI
AAMI2 | This is the first time that SCM has been applied to feature extraction.
This paper combines different feature extraction methods, the accuracy is improved, and the classification rate is relatively high. | It is dependent on QRS’s window lengths. |
|
| Faziludeen [10] | Wavelets and SVM | MITDB | Normal (N)
Premature ventricular contraction (PVC)
Left bundle branch block (LBB) | This work uses one against one (OAO) SVM to classify ECG signals. | This method requires designing features manually, and the classification of ECG signals is less. |
|
| Radovan [11] | SVM | PhysioNet/CinC | Normal (N)
Atrial fibrillation (A)
Other rhythm (O)
Noisy records (P) | Combined with SVM and simple threshold-based rules, it can improve performance. | This method requires designing complex features manually. |
|
| Mondéjar-Guerra [12] | Multiple SVMs | MITDB | Normal (N)
Supraventricular ectopic beat (SVEB)
Ventricular ectopic beat (VEB)
Fusion (F) | This work trains and integrates specific SVM models for each type of feature; it offers a satisfactory performance. | Data fusion is relatively simple. |
|
