Comprehensive Summary
The goal of this study was to determine the efficacy of an AI algorithm in interpreting electrocardiogram (ECG) results to improve atrial fibrillation (AF) diagnosis and treatment with prescription non-vitamin K antagonist oral anticoagulants (NOACs). The trial was conducted at an academic center and a community hospital in Taiwan, where non-cardiologists were randomly assigned either the AI intervention group (n=120) or the control group (n=113), and screened 8,857 and 8960 eligible patients, respectively. The study focused on ECGs of patients who had not been diagnosed with AF, had no previous NOAC or warfarin prescriptions, and had no history of stroke. The AI system analyzed all ECGs, but only physicians in the intervention group were given a notification of the AI’s label before it was compared to a formal ECG report made by a cardiologist. 275 patients in the AI intervention group and 245 patients in the control group were identified by the AI as having a high risk of AF. There were 48 and 35 cases of AI error in the groups where the algorithm identified AF, but a cardiologist did not. Diagnosis rates of AF increased by 40% and prescription rates of NOACs increased by 85% in the intervention group. There was no statistically significant difference between groups in clinical outcomes such as cardiology outpatient visits, ischemic stroke, hemorrhagic stroke, new-onset heart failure, or death. Even in the intervention group, NOAC prescription rates were still lower than what is seen by cardiologists. This study is the first to make an AI ECG-based clinical decision support system for AF identification, but highlights the need for larger-scale trials to ensure efficacy and determine if there is any difference in clinical outcomes due to better AF diagnostic ability.
Outcomes and Implications
Atrial fibrillation is a leading cause of ischemic stroke, and often underdiagnosed by non-cardiologists. This leads to lower prescription of non–vitamin K antagonist oral anticoagulants (NOACs), which can reduce the likelihood of stroke and death. Most clinical decision support systems are for patients already diagnosed with AF, not for identifying it from ECGs. ECG examinations are commonly and easily performed, so non-cardiologists need to be able to use that data to accurately diagnose and treat patients, and can do so more successfully with the use of AI. While diagnosis rates of AF and prescription of NOACs both increased, clinical outcomes between groups with and without the AI algorithm’s assistance were the same. Further and larger studies must be done to determine if there are any positive patient outcomes from non-cardiologists' use of AI to diagnose AF.