Automated computerised detection of atrial fibrillation from the electrocardiogram

Abstract

Atrial Fibrillation (AF) is the most common cardiac arrhythmia affecting approximately 2% of the general population with prevalence expected to increase dramatically between now and 2050. The electrocardiogram (ECG) can be used to analyse the electrical activity of the heart. Human interpretation of the ECG is the current standard for diagnosis of AF with automated analysis programs generating a significant amount of false alarms. The main sources of these false alarms are generally; low signal-to-noise ratio (SNR) of key ECG features during ambulation and the presence of other arrhythmia such as sustained ectopic beats. This project focuses on improving automated AF detection from the ECG using three main strategies (1) Implementation of machine learning techniques, (2) design and characterisation of advanced ECG electrode materials and (3) optimisation of ECG electrode placement. Initially, an assessment of multivariate classification models for automated detection of AF using R-R intervals alone has been conducted. A Random Forests based model improved overall sensitivity, specificity and positive predictive value over the irregularity measurements alone. As ECG signal magnitude and stability is largely dependent on the impedance properties of the monitoring electrode a range of electrode inks and electrolytic hydrogels were sourced and used to manufacture Ag/AgCl ECG electrodes. The electrodes were characterised using electrical impedance spectroscopy and the results modelled using an equivalent circuit model. The electrode ink with an 80:20 Ag/AgCl ratio performed best demonstrating the lowest value of electrode-electrolyte charge transfer resistance. The hydrogel with the lowest bulk resistivity exhibited a reduced electrode-skin impedance across all investigated frequencies. In an effort to improve SNR of the ECG a new bipolar lead selection algorithm has been developed and assessed on body surface potential map data to determine optimal Holter electrode placement for maximum R-wave and P-wave amplitude, the key features to automated AF detection. The new bipolar ECG leads (R-lead and P-lead) showed significant improvement in peak-to-peak signal amplitude and signal root mean square over all other previously described ECG lead systems. With the future of long-term arrhythmia monitoring being directed toward disposable patch based ECG systems, a further lead selection method was developed which accounted for inter-electrode distance to determine the optimal placement of ECG patches to maximise R-wave and P-wave signal magnitude at short inter-electrode distance.

Date of Award	Aug 2017
Original language	English
Supervisor	Jim McLaughlin (Supervisor)