Factors Influencing Automated Detection of Atrial Fibrillation

Peter Macfarlane1, Shahid Latif2, Brian Devine2
1Past president, 2University of Glasgow


Introduction: There has been recent increasing interest relating to automatically analysing a single channel ECG to detect cardiac arrhythmias. However, little interest has been expressed in determining how accuracy might be affected by using lead I as opposed to using the 12 lead ECG or 6 limb leads.

Aim: The aim of this pilot study was to determine the effect of reducing the number of leads from 12 to 6 to 1 and separately, to look at the effect of analysing a single 30s recording as one continuous recording versus five contiguous recordings constituting a 30s recording.

Methods: One hundred 10s 12 lead ECGs correctly reported as atrial fibrillation (AF) were extracted from a Glasgow database. All chest leads were removed from the data and the remaining 6 limb leads then used for analysis of rhythm. Similarly, lead I alone was used. Separately, 100 consecutive single lead I ECGs classified as AF in the PhysioNet 2017 database were analysed, both as single 30s recordings and as five contiguous 10s recordings commencing 0,5,10,15 and 20s from the start of the recording. An algorithm made the final diagnosis from 5 reports. All analyses were made with the Glasgow program.

Results: For the 10s 12 lead recordings, 95% were reported as AF using 6 limb leads and 91% using lead I. Using the 30s recordings, 92% were reported as AF using a single 30s analysis and 91% as AF using the five recordings. In the latter study, there was local disagreement with the PhysioNet classification in 2/8 and 4/9 false negative AF reports effectively increasing sensitivity to 94% and 95% respectively.

Conclusion: Single lead and 6 lead recordings are not as sensitive as 12 lead recordings in detecting AF and five 10s reports combined are no more accurate than a single 30s report.