Key result
ECG-trained RNN detects AF with ~96% accuracy but fails to identify sinus tachycardia.
Why the study?
Existing deep learning architectures for detecting tachycardia diseases lack interpretation and understanding of the results obtained, prompting an exploration of the features learned by these models.
Can deep learning models trained on atrial fibrillation accurately detect other types of tachycardia in ECG signals?
Can deep learning models trained on atrial fibrillation accurately detect other types of tachycardia in ECG signals?
Deep learning models trained on atrial fibrillation can detect ventricular fibrillation due to the shared absence of P-waves, but fail to detect sinus tachycardia which features upright P-waves.
Caution against broad tachycardia detection with AF-trained models; leaves open need for diverse arrhythmia-specific ECG datasets.
Electrocardiogram (ECG) signal is one of the most reliable methods to analyse the cardiovascular system. In the literature, there are different deep learning architectures proposed to detect various types of tachycardia diseases, such as atrial fibrillation, ventricular fibrillation, and sinus tachycardia. Even though all types of tachycardia diseases have fast beat rhythm as the common characteristic feature, existing deep learning architectures are trained with the corresponding disease-specific features. Most of the proposed works lack the interpretation and understanding of the results obtained. Hence, the objective of this letter is to explore the features learned by the deep learning models. For the detection of the different types of tachycardia diseases, the authors used a transfer learning approach. In this method, the model is trained with one of the tachycardia diseases called atrial fibrillation and tested with other tachycardia diseases, such as ventricular fibrillation and sinus tachycardia. The analysis was done using different deep learning models, such as RNN, LSTM, GRU, CNN, and RSCNN. RNN achieved an accuracy of 96.47% for atrial fibrillation data set, 90.88% accuracy for CU-ventricular tachycardia data set, and also achieved an accuracy of 94.71, and 94.18% for MIT-BIH malignant ventricular ectopy database for ECG lead I and lead II, respectively. The RNN model could only achieve an accuracy of 23.73% for the sinus tachycardia data set. A similar trend is shown by other models. From the analysis, it was evident that even though tachycardia diseases have fast beat rhythm as their common feature, the model was not able to detect different types of tachycardia diseases. The deep learning model could only detect atrial fibrillation and ventricular fibrillation and failed in the case of sinus tachycardia. From the analysis, they were able to interpret that, along with the fast beat rhythm, the model has learned the absence of P-wave which is a common feature for ventricular fibrillation and atrial fibrillation but sinus tachycardia disease has an upright positive P-wave. The time-based analysis is conducted to find the time complexity of the models. The analysis conveyed that RNN and RSCNN models could achieve better performance with lesser time complexity.
No takes yet. Share an insight, caveat, or question.
Sanjana et al. (2020) studied Tachycardia (Atrial Fibrillation, Ventricular Fibrillation, Sinus Tachycardia) (n=8,642). Deep learning models (RNN, LSTM, GRU, CNN, RSCNN) trained on Atrial Fibrillation data vs. Tested on unseen tachycardia datasets (Ventricular Fibrillation, Sinus Tachycardia) was evaluated on Accuracy of RNN for atrial fibrillation dataset. A recurrent neural network trained on atrial fibrillation ECG data achieved 96.47% accuracy for atrial fibrillation and 90.88% for ventricular fibrillation, but failed to detect sinus tachycardia (23.73% accuracy).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: