Background: The identification of atrial fibrillation (AF) using artificial intelligence (AI), either with medically prescribed ECG monitors or non-prescription devices such as watches, is not to be taken lightly. Little focus has been placed on the cost, anxiety and potential therapeutic consequences of a false positive diagnosis. The potential value of AI in AF diagnostics is not debated here, but if it is to be used, the ECG used to determine the truth set should be capable of sorting AF from its various mimickers. Many current methods are limited by the duration of the ECG and the fidelity of the P-wave as well as by the validation process details. These limitations can lead to misdiagnosis of other arrhythmias with RR interval variability as AF by the AI engine.
Objective: Our objective was to build an AI to detect AF with better than 90% sensitivity and 90% specificity capable of identifying the onset and offset of episodic AF events, without conflating other atrial (or ventricular) arrhythmias that mimic AF with RR interval variability.
Methods: For this work, we built a convolutional neural network (CNN) that would analyze Carnation Ambulatory Monitor (CAM™) (Bardy Diagnostics, Inc., Seattle, WA) ECG and its associated RR interval data to produce an AF yes/no output for every half-second of ECG data. Training of the CNN was done with CAM ECG recordings from 1,227 patients, 474 with paroxysmal or persistent AF, and 753 without AF. Included in the 753 patients were 148 patients with dense atrial or ventricular ectopy, and both atrial flutter (AFL) and sustained atrial tachycardia (AT) with variable conduction. This distinction is pertinent to the diagnostic and therapeutic medical and procedural management, and stroke risk should these disorders be diagnosed as AF. Our validation was comprised of two-hour excerpts of CAM ECG data chosen at random from 50 AF-positive patients and 50 AF-negative patients (200 hours total). AF was diagnosed if variable p-wave morphology was present for at least 30 seconds. AF presence and duration were confirmed by a team of experienced cardiac electrophysiology clinicians. Disagreements between the three validating electrophysiology clinicians were adjudicated at weekly review meetings.
Results: The AI differentiates AF not only from normal sinus rhythm, but also from other conditions such as atrial ectopy, ventricular ectopy, atrial flutter and atrial tachycardia with variable conduction. Our results were 96.82% sensitive and 99.86% specific with a positive predictivity of 99.79% for detecting 30 seconds of AF or longer.
Conclusions: Our P-wave centric continuous ECG monitoring technology allows our neural network, or AI, to differentiate between AF and a host of rhythms that mimic AF. AI systems that do not make these distinctions may mislead both patients and clinicians.