Event-Driven Arrhythmia Classification using Level-crossing ADCs

Arrhythmia, an abnormal heart rhythm, is a significant medical condition. Continuous monitoring of electrocardiogram (ECG) signals plays a crucial role in early detection of life-threatening arrhythmias and remote processing of ECG signals using small wearable devices enables faster and low-cost diagnosis. This thesis explores the use of a novel class of Analog-to-Digital Converters (ADCs) known as Level-Crossing ADCs (LC-ADCs) for acquiring non-uniformly sampled ECG signals for the application of event-driven arrhythmia classification. The research investigates the impact of LC-ADC parameters on signal quality using different LC-ADC architectures at varying resolutions and quantization steps. Utilizing an artificial neural network (ANN) and the MIT-BIH Arrhythmia Database, the study demonstrates that event-driven ECG signals, when compared to uniformly-sampled data, achieve comparable classification accuracy at significantly lower complexity. The best event-driven model achieves over 97% accuracy with a 79% reduction in ANN complexity and a signal-to-distortion ratio (SDR) exceeding 21dB. Additionally, a new methodology for optimizing design parameters of LC-ADCs is presented, enhancing sampling accuracy and reducing data stream rates. The study introduces a one-dimensional convolutional neural network (1D-CNN) based classifier for evaluating event-driven data from various LC-ADC models. Comparative analysis with uniformly sampled data reveals the ability of a 7-bit LC-ADC, with 2385 Hz clock frequency and 6-bit clock resolution, to offer 3x data compression while maintaining an SDR of 21.19 dB and significantly reducing computational requirements for cardiac arrhythmia classification. The 1D-CNN achieved 99.2%, 89.98% and 91.64% overall accuracy, sensitivity and specificity, respectively. An open-source event-driven arrhythmia database is also presented. Furthermore, this research addresses the challenges posed by variable-length two-dimensional data vectors resulting from event-driven ECG signals. A mapping approach is proposed, converting these vectors into fixed-length feature vectors using functional approximation and Chebyshev polynomials. The proposed three-layered ANN ECG classifier with Chebyshev coefficients demonstrates an average accuracy of 98.80% and average sensitivity of 91.5%, outperforming state-of-the-art models with minimal parameters (20k). Additionally, efficient signal reconstruction methods from event-driven signals are also proposed, with the iterative approach showcasing accurate reconstruction with negligible error. Finally, an event-driven ECG classification system is proposed which only requires 0.1 MIPS (mega instructions per second) per ECG beat. In summary, this thesis contributes novel insights into the application of LC-ADCs for ECG signal acquisition, event-driven classification, and efficient signal reconstruction. The proposed methodologies demonstrate promising results for enhancing the accuracy and efficiency of arrhythmia classification in wearable devices, paving the way for advancements in real-time cardiovascular monitoring.