Options
September 8, 2023
Journal Article
Title
Detection of Ventricular Tachycardia Using Artificial Intelligence
Title Supplement
Abstract
Abstract
Introduction: Cardiovascular diseases are the leading cause of death worldwide. The associated ventricular tachycardias are currently diagnosed via long-term electrocardiogram (ECG) recordings. This work explores the potential of using Artificial Intelligence for real-time classification of ventricular tachycardias with a few selected features directly on such a device.
Methods: In this study, Einthoven leads I and III are used to extract 49 heart rate variability (HRV) features from the time, frequency, and time-frequency domains using the Pan-Tompkins algorithm. Additionally, 26 features based on signal morphology and 26 statistical features are extracted. Sequential forward selection, feature importance from Gemini index, and ANOVA tests are used for selecting the best features and Artificial Neural Networks (ANN), Random Forrest (RF), and Support Vector Machines (SVM) for classification. The applied Charité data set consists of 447,774 signals recorded with an event recorder. To account for the unbalanced data set, oversampling of the data with ventricular tachycardia is performed.
Results: On an independent test dataset, the ANN demonstrates the best classification performance, achieving an area under the receiver operating characteristic curve (ROCAUC) of 0.71, RF 0.7, respectively, and SVM 0.68. The most relevant features selected for classification are based on HRV, including the percentage of consecutive RR intervals below 50 ms. When atrial fibrillation patients are excluded, the ANN achieves a best ROCAUC of 0.74. Compared to results reported in the literature for other datasets, the ROCAUC is relatively low. This can be partly attributed to the poorer signal quality of the event recorders.
Conclusion: The classification accuracies presented in this study are not yet sufficient for practical use. The algorithms need to be adapted to account for the poorer signal quality of event recorders. The study demonstrates that HRV features have a significant impact on classification, and strong overlaps with atrial fibrillation patients are present.
Methods: In this study, Einthoven leads I and III are used to extract 49 heart rate variability (HRV) features from the time, frequency, and time-frequency domains using the Pan-Tompkins algorithm. Additionally, 26 features based on signal morphology and 26 statistical features are extracted. Sequential forward selection, feature importance from Gemini index, and ANOVA tests are used for selecting the best features and Artificial Neural Networks (ANN), Random Forrest (RF), and Support Vector Machines (SVM) for classification. The applied Charité data set consists of 447,774 signals recorded with an event recorder. To account for the unbalanced data set, oversampling of the data with ventricular tachycardia is performed.
Results: On an independent test dataset, the ANN demonstrates the best classification performance, achieving an area under the receiver operating characteristic curve (ROCAUC) of 0.71, RF 0.7, respectively, and SVM 0.68. The most relevant features selected for classification are based on HRV, including the percentage of consecutive RR intervals below 50 ms. When atrial fibrillation patients are excluded, the ANN achieves a best ROCAUC of 0.74. Compared to results reported in the literature for other datasets, the ROCAUC is relatively low. This can be partly attributed to the poorer signal quality of the event recorders.
Conclusion: The classification accuracies presented in this study are not yet sufficient for practical use. The algorithms need to be adapted to account for the poorer signal quality of event recorders. The study demonstrates that HRV features have a significant impact on classification, and strong overlaps with atrial fibrillation patients are present.
Author(s)
Open Access
File(s)
Rights
Use according to copyright law
Additional link
Language
English