Enhancing the diagnostic quality of ECGs in mobile environments

Abstract

As the leading cause of deaths worldwide, Cardiovascular Disease (CVD) has imposed a serious burden onto society. Being reactive in approach, the current healthcare infrastructure struggles to address the problem properly. In contrast, a proactive approach can offer better disease management by predicting impending episodes. The key to the proactive approach is remote continuous monitoring. Traditional long-term monitoring faces serious challenges in transmission as data must be sent out continuously for further processing, leading to short battery life of the sensor node and defying the very notion of continuous monitoring. However, with intelligent signal processing algorithms, signal analysis may directly take place at the sensor node itself. Hence one does not need to transmit data until abnormality is detected. This in turn may save the energy at the sensing node and therefore preserve the notion of continuous monitoring.

In this thesis, we first investigate the automated feature detection of Electrocardiogram (ECG) fiducial points by proposing two different algorithms based on time-domain morphology and gradient, as well as time-frequency-domain with Discrete Wavelet Transform (DWT) respectively. Secondly, to tolerate the possible misdetection errors from ECG fiducial points detection algorithms, we investigate spectral energy as a feature for normal and abnormal ECG classification in feature calculation, as well as statistical analysis of its variation and classification performance under worst-case misdetection. Our exploration shows that spectral energy mostly manages to tackle misdetection error and shows better classification performance than wave duration-based classification. Thirdly, we explore the possibility of adding more time and frequency domain features for enhancing the classification accuracy. Different levels of improvements in classification performance can be observed with respect to the classification models and number of ECG leads involved. Finally, hardware architecture is proposed to integrate spectral energy calculation and Linear Discriminant Analysis (LDA) classifier for an on-body ECG classifier. Verification and power estimation of the system is carried out and shown to be efficient for on-body ECG normal and abnormal classification.

More information

Identifiers

Catalogue record

Export record