Noise-Aware Deep ECG-based VT/VF Detection Method for Trustworthy Life Threatening Arrhythmia Monitoring Devices

Abstract

Timely detection of cardiovascular arrhythmias such as ventricular tachycardia (VT) and ventricular fibrillation (VF) is pivotal for preventing sudden cardiac arrest and can be achieved through continuous cardiac health monitoring. Portable wearable health monitoring devices have limitations related to misdiagnosis, low memory, low battery power, and low processing time/latency. The paper presents a noise-aware convolutional neural network (CNN) based single-stage VT/VF detection method which optimizes hyperparameters (activation function, kernel size, layer, and kernel count) for improved performance and reduced computational complexity. The 1-dimensional CNN (1D-CNN) is trained with 3 classes, constituting of VT/VF, non-VT/VF, and noisy segments of 5s duration. The optimized VT/VF detection method with 5-layers (size: 8.10MB) is implemented on Raspberry Pi-4 platform and validated on untrained datasets to obtain an overall recall/sensitivity, specificity, accuracy, and F1-score of 100%, 99.98%, 99.99%, and 100%, respectively with latency 1.93ms for an electrocardiogram (ECG) segment of 5s duration. Compared to 2-class VT/VF detection methods, the 3-class noise-aware method demonstrates its accuracy and robustness by reducing false alarms (0 to 100%) and improving accuracy (-0.05 to 91.04%). The method offers 8.00-88.04% energy savings compared to other methods that include or omit signal quality analysis. Thus, the single-stage VT/VF detector improves energy efficiency and reliability of battery-operated wearable cardiac health monitoring devices and edge health analytics systems. © 1963-2012 IEEE.