Artificial intelligence-enabled electrocardiogram: can we identify patients with unrecognized atrial fibrillation?

Abstract

Atrial fibrillation (AF) is known to affect at least 30 million people worldwide [1,2], although this may be an underestimation. AF can be asymptomatic and fleeting and often goes undetected. In fact, it has been estimated that approximately one million Americans live with unrecognized AF [3]. The proportion of patients with paroxysmal AF versus persistent AF varies with age (with paroxysmal AF more common in patients <50 years), and it is estimated that about 25% of patients with AF have a paroxysmal pattern [4]. Identifying patients with undiagnosed AF is important as they have a fivefold increased risk of stroke [1,2] and the first manifestation of AF may be a disabling stroke. Furthermore, AF-related strokes carry a particularly poor prognosis [3,5]. When AF is recognized, interventions including oral anticoagulation or left atrial appendage closure can lower stroke risk and mortality [5,6]. Due to its frequently paroxysmal nature, AF is often under detected. Currently, prolonged electrocardiographic monitoring is implemented to detect patients with suspected AF – a process that is expensive, resource intensive, and at times poorly tolerated. In nearly 5,000 patients referred for continuous 24-hour monitoring, the prevalence of paroxysmal AF was 2.5% [7]. It has been estimated that even among a highrisk cohort of patients with ischemic strokes, 20% remain cryptogenic despite thorough diagnostic evaluation [5]. Apart from the low yield, long-term cardiac monitoring is resource intensive, expensive, and impractical for broad-scale application. A frequent clinical dilemma is whether or not to anticoagulate patients without documented AF based on incomplete information; studies of empiric anticoagulation following embolic stroke of uncertain source have found no benefit and harm (i.e. bleeding) [8,9]. Therefore, it is essential to detect paroxysmal AF to guide therapy to prevent stroke. Recently, we developed an artificial intelligence-enabled electrocardiogram (AI-ECG) algorithm using over 500,000 normal sinus rhythm standard 10-second 12-lead ECGs from over 180,000 patients using machine learning to identify those with a high likelihood of undocumented AF [10]. This work demonstrated that the application of a convolutional neural network (CNN) to a single ECG recorded during sinus rhythm could effectively identify paroxysmal AF, with an area under the receiver operator curve (AUC) of 0.87 (95% confidence interval [CI], 0.86–0.88), sensitivity of 79.0% (95% CI, 77.5–80.4%), specificity of 79.5% (95% CI, 79.0–79.9%), F1 score of 39.2% (95% CI, 38.1–40.3%), and overall accuracy of 79.4% (95% CI, 79.0–79.9%). The diagnostic yield improved when applied to patients with multiple ECGs (AUC 0.90). With the impressive performance of the AI-ECG algorithm, the question becomes: what is the AI seeing that the human eye is missing? Due to the nature of CNNs, identification of the signal features selected by the AI is currently not possible. We presume that underlying structural changes (e.g. myocyte hypertrophy, fibrosis, chamber enlargement) precede the onset of AF and that these substrate changes result in subtle yet detectable ECG changes. It has been reported that normal sinus rhythm on a surface ECG may not accurately reflect atrial function. One report found approximately one-third of patients with AF undergoing cardioversion to lack sinus contraction of the left atrial appendage despite a surface ECG demonstrating sinus rhythm [11]. Another report showed that about one-fourth of patients had a surface ECG revealing sinus rhythm despite fibrillation of the left atrial appendage documented via transesophageal echocardiogram [12]. These studies suggest that there may be unrecognized patterns on the ECG associated with AF that are detectable during sinus rhythm by means of deep neural network. The exposure of CNN to over half a million ECGs enables it to extract and process subtle features not routinely noticed by the human eye. In a recently published clinical case, we reported a patient with a cryptogenic stroke deferred anticoagulation therapy, based on the lack of documented AF on long-term cardiac monitoring, who developed another stroke a few years later [13]. Retrospective AI-ECG analysis of this patient’s available ECGs in sinus rhythm demonstrated the patient had a high likelihood of undiagnosed AF years before the incident stroke events. Based on these findings, one may have considered it reasonable to initiate anticoagulation therapy earlier in the course and possibly preventing harm to the patient. This exemplifies a potential clinical role for the AI-ECG algorithm in management decisions and poses the question: could