B-AB24-04 ARTIFICIAL INTELLIGENCE ENABLES DRAMATIC REDUCTION OF FALSE ATRIAL FIBRILLATION ALERTS FROM INSERTABLE CARDIAC MONITORS

Abstract

Insertable Cardiac Monitors (ICMs) detect Atrial Fibrillation/Flutter (AF) episodes with high sensitivity at the cost of lower specificity. The resulting false AF detections hinder longitudinal monitoring by increasing clinician review burden and over-reporting of AF episode and duration metrics. Cloud-based solutions may offer an extra layer of technology between the ICM and clinician that would reduce the presentation of false AF information. This work was conducted to develop and validate a cloud-based artificial intelligence (AI) algorithm to suppress false AF detections from ICM devices. A deep convolutional network was progressively trained across several large ECG datasets to detect and suppress non-AF episodes. Episodes ≥ 1 hour or correlated to patient symptoms bypass the AI to ensure clinician review. This algorithm was evaluated on prospectively collected data from 434 LINQ II patients consecutively implanted from July 25 - August 31, 2020 and followed for 3 months post-implant. Reasons for monitoring included: known AF (147); stroke (108); syncope (88); suspected AF (42); palpitations (29); and other (20). Each episode’s 2-min ECG was independently adjudicated by up to 3 reviewers. Relative performance of the AI was assessed with respect to episodes, alerts (all episodes from a given day), and duration (time in AF). The dataset included 5,025 ICM-detected AF episodes from 147 patients, comprising 2,186 total AF alerts. True AF was adjudicated within 3,090 of the ICM-detected episodes, resulting in 1,275 true AF alerts from 89 patients. The sensitivity of the AI relative to true ICM detections was 98.9% (3056/3090) and 99.3% (1266/1275) at the episode and alert level, respectively. The relative specificity of the AI was 71.4% (1381/1935) and 74.1% (675/911) at the episode and alert level, respectively. The AI retained 100% (99.96%) of the total true AF duration detected by the ICM with a positive predictive value of 98.6%. The AI can eliminate three-fourths of false AF alerts from the ICM, while preserving nearly all true AF alerts. This could dramatically reduce clinician review burden and facilitate longitudinal monitoring via more accurate AF episode and duration metrics.