Deep learning and signal processing-based methods for detection of life-threatening cardiac arrhythmias

Abstract

Atrial fibrillation (AF), ventricular tachycardia (VT), and ventricular fibrillation (VF) are three of the most prevalent life-threatening arrhythmias and are the major contributors to cardiovascular mortality. Due to the intermittent and paroxysmal nature of the arrhythmias it is highly difficult to detect AF, VT, or VF. Although the clinically sustained AF, VT, and VF requires urgent attention, early detection of these arrhythmias prevent it from reaching this stage. Early detection of the life-threatening arrhythmias are possible through continuous monitoring of electrocardiogram (ECG). As continuous monitoring creates a huge amount of data, it is not possible to manually check for arrhythmias, so there is a need for automated life-threatening arrhythmia detection methods which are reliable and can be implemented on wearables and portable cardiac health monitors. The realtime ECG signals are also highly corrupted by noises which lead to misclassifications. Since low-power devices have low memory capacity and battery constraints, the life threatening detectors must be noise-aware and lightweight with fast processing time, and low energy consumption. In the thesis, we propose lightweight, noise-aware methods for AF, VT/VF detection using signal processing, machine learning, and deep learning approaches. For quality aware AF recognition, we developed four frameworks utilizing the ECG waveform and the characteristics of AF. A novel R-peak detection method is developed to compute RRI interval. An optimized convolutional neural network (CNN) was developed using ECG to detect AF. The second method used R peak detection algorithm to calculate the RR intervals and find a discriminative feature to detect AF using machine learning classifiers. The RR interval, absence of P-wave, and presence of fibrillatory wave characteristics were used with an optimized CNN to develop the third framework. The fourth framework presents a noise-aware, single-stage AF detection method developed using AF, non-AF, and noisy ECG signals as its classes. The quality-aware VT/VF recognition methods include, a simple discriminative feature to detect VT/VF, a signal quality aware VT/VF detection method, and a single-stage noise-aware VT/VF detection method. The thesis also introduces a noise aware AF/VT/VF detector designed to prevent misclassification of AF as VT/VF and avoid unnecessary shocks during cardiac arrest.