Co-Author(s):
Philip Currie, MBBS - Cardiovascular Services
David Playford, MBBS FRACP PhD FCSANZ FESC FACCC - Professor, The University of Notre Dame Australia
Luke Bollam, BEng - Alerte Digital Health
Razali Mohamad, MEng, BCM - Alerte Digital Health
Kushwin Rajamani, MBBS
Rukshen Weerasooriya, MBBS
universite de Bordeaux
Place de la Victoire, 33000 Bordeaux, France
Introduction | Objectives: Implanted loop recorders (ILR) have an important role to play in the detection of atrial fibrillation (AF).  Continuous home monitoring of data from ILR may result in false positive AF detections due to a number of factors including noisy signal and frequent ectopic beats.  A pilot study was designed to evaluate the use of AI to improve AF detection from Biotronik Biomonitor 3 (BM3) recordings using a source agnostic AI strategy.
Methods: A total of 1015 individual BM3 recordings were used in this study.  The BM3 data was uploaded to an ECG AI web application and was then manually labelled by cardiologists.  An AI model was created based upon labelled data from a clinical database of 18,000 beat and rhythm labelled 12-lead ECGs representing a wide range of arrhythmias and beat disorders. For the purpose of this pilot study, the following minimal arrhythmia labelling criteria were pre-defined: High relevance (eg atrial fibrillation), Low relevance (eg sinus rhythm plus ectopics), Not appraisable (eg noise). The following experiments (Exp) were conducted: Exp 1.  An AI model was created according to the minimal labelling criteria only using 12-lead ECG data; Exp 2. The existing AI model was assessed on BM3 data with no training using BM3 data; Exp 3. The existing AI model was fine tuned using BM3 data under the following 2 scenarios: Exp 3a. 29% of BM3 data (295 recordings) used for fine tuning and evaluated using 71% of BM3 data (717 recordings). Exp 3b 68% of BM3 data (688 recordings) used for fine tuning and evaluated using 32% of BM3 data (327 recordings).
Results: Exp 1 AUC 0.975004